ADAPTIVE UPDATE METHOD FOR EMBEDDED REAL-TIME

JET ENGINE MODELS

by

Frederick John Pineo

B.S., Clarkson College of Technology (1980)

SUBMITTED IN PARTIAL FULFILLMENT OF

THE REQUIREMENTS OF THE

DEGREE OF

MASTER OF SCIENCE

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

February 1985

Q Frederick J. Pineo 1985

The author hereby grants to M.I.T. permission to reproduce and to distribute

copies of this thesis document in whole or in part.

Signature of Author
Signature redacted
Depar ent of Mechanical Engineering
A0 January 19, 1985
Signature redacted
Certified by
J. Karl Hedrick
Thesis Supervisor

Signature redacted
Accepted by
Ain A. Sonin
Chairman, Departmental Committee on Graduate Students
MASSACHUSETTS iNSTITUTE
OF TECHNOLOGY

MAR 2 2 1985
LIBRARIES ARCHIVES
 77 Massachusetts Avenue
Cambridge, MA 02139
MITLibrares http://Ilibraries.mit.edu/ask

DISCLAIMER NOTICE
Due to the condition of the original material, there are unavoidable
flaws in this reproduction. We have made every effort possible to
provide you with the best copy available.

Thank you.

Some pages in the original document contain text

that is illegible.
 ADAPTIVE UPDATE METHOD FOR EMBEDDED REAL-TIME

JET ENGINE MODELS

by

Frederick John Pineo

Submitted to the Department of Mechanical Engineering

on January 18, 1985 in partial fulfillment of the
requirements for the Degree of Master of Science
in Mechanical Engineering

ABSTRACT

Real-time jet engine models that are compact enough to be programmed

along with the control laws in the digital engine control (embedded) have
recently come into use. As with all models, there are the inevitable
modelling errors that limit the capability to discern failures and/or to
monitor the health of the engine components. A technique to adapt the
model to reflect the particular engine that the control is mated to is a
useful way to improve the fidelity of the model. This would allow the
model to change as the engine quality changes. Several techniques were
explored for the adaptation, and that selected was a modification to the
extended Kalman filter. The filter gains are pre-computed at a repre-
sentative condition and applied to the non-linear engine model. The
results show that the adapted model tracks an analytical truth model of the
engine in steady state and dynamic modes, and that it can be used for
engine condition monitoring, analytical redundancy and possibly synthetic
control.

Thesis Supervisor: Dr. J. Karl Hedrick

Title: Professor of Mechanical Engineering

2
 TABLE OF CONTENTS

ABSTRACT ...................... ....................................... 2

ACKNOWLEDGEMENTS ..................................................... 4

LIST OF FIGURES ...................................................... 5

LIST OF TABLES ....................................................... 6

LIST OF SYMBOLS .. .................................................... 7

CHAPTER

1. INTRODUCTION ............................... 9
Evolution of Real Time Models ........... 10
Benefits of an Adaptive Algorithm ....... 11
Outline of This Study ................... 12

2. ENGINE MODEL ............................... 14

Jet Engine Description .................. 14
Model Requirements ...................... 17
Detailed Model Description .............. 18

3. ADAPTIVE ALGORITHM DEVELOPMENT ............. 39

Parameter Set Selection ................. 40
Gain Matrix Methodology ................. 48
Inverse Jacobian Gain Matrix ......... 49
Recursive Least Squares Derived Gain Matrix 53
Constant Gain Extended Kalman Filter .. 0 . . 57
.

Extended Kalman Filter ............... 63

Robustness Issues ....................... 65
Singular Values ...................... 66
Robustness Analysis .................. 68

4. NUMERICAL RESULTS .......................... .. 0 . 72

.

..... 0
Static Performance ...................... .. 0 . 72
Dynamic Tracking ........................ .. 0 . 79

5. PROCESSOR REQUIREMENTS ..................... 107

Required Processor Time-Current Computer 107
Practical Issues of the Embedded Model .. 108

6. SUMMARY AND CONCLUSIONS .....................

REFERENCES ........................................................... 113

APPENDIX: DATA FOR PARAMETER SET SELECTION ........

3
 ACKNOWLEDGEMENTS

I gratefully acknowledge the guidance of my advisor, J. Karl Hedrick,

whose advice and direction during the past six months was instrumental in
keeping my work on track and in focus. I also thank the General Electric
Company for financial support to perform the work, and Mike French and Mike
Idelchik at GE for technical support. Finally, I thank my wife, Jean, for
moral support during the many hours spent.

4
 LIST OF FIGURES

Page

2.1-1 Engine Schematic and Station Descriptions ........ 15

2.3-1 Compressor Characteristics ......................... 20

2.3-2 Flow Chart for Real Time Model ..................... 33

2.3-3 Heat Soak Representation ........................... 37

2.3-4 Thermocouple Compensation .......................... 38

3.1-1 Flight Envelope Check-Out Points ................... 44

3.2.1-1 Pole Plot-Scaled Inverse Jacobian Feed-Back

Matrix .............................................. 54

3.3.2-1 Singular Values for Kalman Filter .................. 70

4.2-1 thru 4 Baseline Accel-Dynamic Simulation ................ 80-83

4.2-5 thru 9 Scaled Inverse Jacobian Gain Update of

Parameters ......................................... 85-89

4.2-10 thru 14 Lease Squares Derived Gain Update of

Parameters ......................................... 91-95

4.2-15 thru 19 Constant Gain Extended Kalman Filter Update

of Parameters ..................................... 96-100

4.2-20 thru 24 Extended Kalman Filter Update of Parameters ...... 101-105

4.2-25 Summary of Update Methods' Errors .................. 106

A-1 Flow Path for Variable Set Selection ............ 136

5
 LIST OF TABLES

Page

3.1-1 Engine Variables ............................. 41

3.1-2 Sets of Parameters ........................... 45

3.1-3 Sensor Errors ................................ 47

3.1-4 Engine-to-Engine Variations .................. 47

3.2.4-1 Summary of Continuous Extended Kalman Filter 63

.
3.2.2-1 Eigenvalues of FINO Engine Model ............. 69

4.1-1 Baseline Performance of the Model ............ 73

4.1-2 Modelling Error Statistics ................... 74

4.1-3 Improvements by Model Adaptation ............. 74

4.1-4 Effects of Engine to Engine Variations ....... 76

4.1-5 Effects of Sensor Errors ..................... 77

5.1-1 Processor Time ............................... 108

A-1 thru A-12 Results for Sets 1-12 ........................ 117-128

A- 13 Envelope Survey of Parameter Sets ............ 129

A-14 thru A-17 Effects of Sensor Errors ..................... 132-135

A-18 RSS of Sensor Error Sensitivities ............ 136

A-19 thru A-21 Effects of Engine to Engine Variations ....... 137-139

A-22 Final Summary ................................ 140

6
 LIST OF SYMBOLS

A, B, C, D; F, G, H State space representation matrices

A8 Exhaust nozzle area
A Area
C Specific heat
p
e Vector of output errors
FAR Fuel to air ratio
FF41 Turbine nozzle flow function
FG Gross thrust
FN Net thrust
h Film coefficient
HPC Compressor horsepower
HPF Fan horsepower
HPHPT High pressure turbine horsepower
HPLPT Low pressure turbine horsepower
J Jacobian matrix
K Kalman gain matrix
m Mass
N1 Fan rotor speed
N Compressor rotor speed
2
P Total pressure
Ps Static pressure
Q, R Kalman filter weighting matrices
r Vector of parameters
T Total temperatures
u Vector of inputs
UBHP Unbalanced torque in high pressure spool
UBLP Unbalanced torque in low pressure spool
W Flow
WF Fuel flow
W--R Corrected flow
x Vector of states

7
XJHP HP rotor inertia
XJLP LP rotor inertia

Y- Vector of outputs

6 Pressure correction P/14.7 psi

n Efficiency
6 Temperature correction T/518.7 *R
T Time constant
Singular value

Subscripts

c Compressor

f Fan

t Turbine

2t LP turbine

Superscripts

A Estimate

8
 1. INTRODUCTION

The advent of advanced microelectronic circuitry has led to the

application of digital engine control to recent turbofan engines.

Increased demands on control law complexity to wring out the optimum

performance have made hydromechanical controls too heavy, complex, and

costly, and also less reliable due to their complexity. Microcircuitry can

replace the hardware in the hydromechanical control, and increase perfor-

mance and reliability with a decrease in weight. Several programs have

demonstrated the usefulness of digital controls. The Navy-sponsored Full

Authority Digital Electronic Control (FADEC) program (Ref. 1) demonstrated

control systems for advanced variable cycle engines. The NASA U.S. Air

Force Digital Electronic Engine Control (DEEC) program (Ref. 2) has demon-

strated performance improvements on the Pratt and Whitney F100 engine. The

NASA Cooperative Control Program (Ref. 3) on the YF-12C triple sonic

aircraft was successful in demonstrating integrated inlet/autopilot/auto

throttle systems.

The control laws for advanced applications will require many highly

reliable sensors. There are two principle solutions to reliability. The

first is hardware redundancy, where all or some sensors are duplicated.

There is the obvious weight penalty here. The second is analytical

redundancy, which uses a reference model of the engine to provide esti-

mates. It is dependent upon the accuracy of the model. This thesis

concerns the use and adaptive update of a current model of an advanced

afterburning turbofan engine produced at the General Electric Company

9
Aircraft Engine Business Group in Lynn, Massachusetts. The model is

embedded in the digital control and processes in real time during engine

operation.

1.1 Evolution of Real Time Models

Models of jet engines can be classified into two types: those that

operate in real time, and those that do not. The first category include

all models whose equation set processing time is less than or equal to the

update time of the model. The second are typically high accuracy models

that try to capture all possible effects occurring in the engine cycle.

The distinction between the two depend on model complexity and the particu-

lar processor. As processors have matured in recent years, the amount of

complexity has grown proportionately. A recent survey of jet engine models

provides insight into the development (Ref. 4). The highlights are pre-

sented here.

The reference classifies the modelling approaches into three cat-

egories. These are Parameter Synthesis (PS), Pseudolinear (PL), and

Simplified Component (SC). The first method utilizes tabulations of sensed

variables as a function of each other. There are no explicit dynamics.

This method was used to to demonstrate hard failure detection and replace-

ment on a J-85 single spool turbojet engine (Ref. 5). It was also used on

a TF-30 two spool afterburning turbofan engine around the flight envelope

(Ref. 6). The second method, pseudolinear, has been used on several

occasions. It consists of a dynamic linear state space structure, with

individual coefficients in the equations varying as a function of the

state. One application was on a Pratt and Whitney F100 two spool

10
afterburning turbofan engine (Ref. 7). This provided an analytical judge

for the fan turbine inlet temperature.. The final type, the simplified

component, is the subject of this investigation.

The simplified components model is based upon highly detailed non-

linear engine simulations. Detail is selectively removed in order to

achieve real time capability while maintaining accuracy. The model con-

tains information about each component of the engine, and thus is modular

in nature. The physics of each process are represented in order to achieve

accuracy, along with appropriate dynamics. Simplified component approaches

have been used in at least three works. Spang and Corley (Ref. 8) used it

for the Quiet Clean, Short Haul, Experimental Engine (QCSEE), French

employed the technique for a variable cycle engine (Ref. 9), and Pisano et

al. modeled a T700 turboshaft engine using it in Ref. 10. The history of

real time models shows that complexity was added as microprocessors

developed, and the latest models are highly non-linear representations of

the engine cycle.

1.2 Benefits of an Adaptive Algorithm

The potential benefits of analytical redundancy are straightforward.

There is the weight savings due to the elimination of physically redundant

sensors, and the cost savings. Additionally, analytical redundancy can

potentially distinguish between soft (drift and small in range) failures

and hard failures (large out-of-range). However, since all models have

errors, the magnitude of failure detection is limited to the modelling

accuracy. Thus, an algorithm that adapts the model to the particular

11
engine can reduce the window for detection, plus provide additional bene-

fits.

A good adaptive algorithm, in essence, becomes a self teaching control,

and can provide the following benefits. First, it can provide engine

condition monitoring (ECM) capability. The detection of component charac-

teristic drifts with engine operating time as the engine deteriorates or

gets dirty has a significant impact on maintenance procedures. For example,

high pressure turbine efficiency typically decreases due to rubs or other

effects, and detection of this fact for a poor engine would easily identify

the cause and replacement action. Second, the adaptation capability would

be extremely useful to minimize modelling errors in consideration of engine

to engine variations. The model could adapt its internal component charac-

terizations to match those of the particular engine that the controller is

on. Third, the adaptation would increase accuracy of the estimates of

sensed variables, due to the capabilities mentioned above. Finally, with a

highly accurate model, the estimates of non-measurable variables could be

used for synthetic control. For example, high pressure turbine inlet

temperature is important for performance and hot parts life, but currently

its measurement is impossible. If a good synthesis of it were possible,

the engine could be controlled to it providing increased performance and

better hot parts life management. It is for the above reasons that this

work was undertaken.

1.3 Outline of This Study

The first task is to describe the component level real time model that

is used in the study. Chapter 2 gives a fairly detailed description of the

12
model and its operation. The model is typical in that each component in

the engine system is described. Chapter 3 deals with the adaptive algorithm

development. Part 3.1 describes the selection of parameters used to adapt

the model. Part 3.2 describes the dynamic methodology used to employ the

parameters. Part 3.3 discusses robustness issues of the update method.

Numerical results are shown in Chapter 4. A discussion of microprocessor

technology is included in Chapter 5. Finally, conclusions and recommenda-

tions are presented in Chapter 6.

13
 2. ENGINE MODEL

The engine model used in this study was developed at the General

Electric Company in Lynn, Massachusetts, and is described in Reference 9.

Many of the details of that description are presented here, in order to

give a clear picture of the work performed. Some changes to the model that

occurred after its initial release are also described in the pages that

follow. The model is contained in General Electric Transient Program No.

82086D per Reference 11.

2.1 Jet Engine Description

Before addressing the model details, a brief description of the actual

engine that it represents is necessary. The subject engine is a modern,

low-bypass turbofan engine with an afterburner for thrust augmentation.

The components, from inlet to exhaust are shown in Figure 2.1-1. The fan,

or low pressure compressor provides initial supercharge to the high pres-

sure compressor (HPC). Part of the compressed air exiting the fan is

bypassed around the high pressure spool in the bypass duct, and is mixed

with the core stream at the mixing plane. The highly compressed air from

the HPC is mixed with fuel and burned in the combustor, providing suffi-

cient energy to drive the fan and compressor. The hot gases from the

combustor are expanded in the high pressure turbine, which is on the same

shaft as the HPC, producing the horsepower required to drive the HPC plus

all parasitic horsepower such as in the accessory gearbox and for airframe

power extraction. The hot gases are further expanded in the low pressure

14
 0 0 S9 40

B pass Duct F.Tv

WF-

2 21 13 16 6

25 14 3 41 45 56 ) 89

I .1
U,1

Mixer
High Pressure Exhaust
Fan High Pressure Turbine
Nozzle
Compressor

Combustor Low Pressure

CTurbine
Afterbu rner

1-4
ENGINE SCHEMATIC AND STATION DESCRIPTIONS

9
turbine (LPT) which drives the fan on their common shaft. After the LPT,

the gases are mixed with the bypass air, and entered into the augmentor

section. Here, additional fuel may be injected to provide additional

thrust. Finally, the hot gases are expanded in the variable exhaust

nozzle, and the expansion provides thrust to propel the jet aircraft.

The jet engine cycle described is a highly non-linear process. It is

further complicated by the following factors: it operates over a wide

range of flight conditions (altitude and mach number combinations); it

operates over a wide range of power settings; and has a high throttle

response capability. As evidence of the last statement, the engine can

accelerate from flight idle to maximum power in five seconds, which is a

fifty-fold increase in thrust.

It is common to describe the points in the process by numbered sta-

tions. The station designations for each component are shown in Figure

2.1-1, and amplified here. Inlet to the fan is station 2, and exit is

station 21. Because of the hub to tip work split, the inlet to the bypass

duct is a different condition than the inlet to the high pressure compressor.

Thus, station 13 designates the inlet to the bypass duct, and station 25

designates the entrance to the HPC. Exit from the HPC is station 3, and

exit from the combustor is station 41. The high pressure turbine and low

pressure turbine are considered close-coupled, so the common designation

between them is station 45. Exit from the LP turbine is station 56, which

is just before the mixed station 6. The bypass duct exit is station 16,

which is axially the same as 56, but in the bypass stream before mixing.

Afterburner fuel flow is introduced as station 61, and fully mixed and

16
burned by station 7, the inlet to the exhaust nozzle. The nozzle throat is

station 8, and nozzle exit station is 9.

2.2 Model Requirements

There are many levels of modelling sophistication for the turbofan

engine cycle, and there exists the classic trade of complexity versus

accuracy. The requirement imposed by having the model contained in the

engine's digital control imposes a heavy penalty on complexity, but if the

model is to be used for a number of control purposes, there may be an

equally heavy penalty on inaccuracy.

The modelling requirements are composed of those dictated by engine

operation and those dictated by the digital engine control. Regardless of

model structure the model must be operational from ground idle (lowest

power setting) to maximum afterburner; for altitudes from sea-level to

60,000 ft, and for flight mach numbers to 1.6. In this case, operational

means stable model characteristics with modelling errors no greater than

five percent on engine measured variables. For this application, the

engine measured variables are fan rotor speed (N1 ), compressor rotor speed

(N2 ), compressor discharge static pressure (PS3C), and low pressure turbine

total temperature (T5). The type of processor in the digital engine

control will dictate the size and nature of the equation set used to

describe the engine. The processor used in this case imposed the following

requirements. First, all equations describing the process characteristics

are expressed as polynomial equations no greater than third order. Second,

there are no tables or table look-up involved. Third, in order to operate

in real time, all equations are evaluated in an overall update time of .01

17
seconds. Fourth, no iteration is permitted between time steps. Finally,

only integer exponents are allowed, plus square roots.

It may be noted that the choice of processor has a strong influence on

model structure. As stated previously, the processor for this application

required polynomials. Other machine languages may be far more efficient

with table look-up routines. Also, as the state of the art advances, model

complexity can advance beyond the present model. Alternatively, the

advances may permit the extended Kalman filter to become practical as

necessary. All previous attempts at the full Kalman filter equations have

faced the computational burden of updating the covariance equations in real

time, but have concluded that the burden is too much (Refs. 8, 10).

Further analysis of current and future processor technology is reserved for

Chapter 5.

2.3 Detailed Model Description

The description of the model is borrowed heavily from Reference 9. In

many places, the description is further amplified. The overall perspective

for this model is a component by component description of the jet engine

aero-thermodynamic cycle with sufficient complexity to provide accuracy and

sufficient simplicity to meet the real time processing requirements. The

first phase of the description involves the original model, and the second

adds detail of the latest incorporations to increase model accuracy. It

should be noted that many of the later additions violate some of the rules

stated in Section 2.2. The main reason for these violations is that

advances in microprocessors since 1981 have made the violations a minor

18
computational issue. The model is described by component from inlet to

exhaust.

The first component in the engine is the fan, or low pressure compres-

sor. It takes the shaft energy provided by the low pressure turbine, and

through the blade aerodynamics, converts it to pressure energy. The fan

operates over a wide range of pressure ratio and flow in order to control

thrust. The operational characteristics of the fan are generically repre-

sented by a four-dimensional fan map as in Figure 2.3-1A. The figure shows

the highly non-linear relationships between airflow, speed, pressure ratio,

and efficiency. The x-axis is corrected fan airflow, where the correction

parametrically collapses performance over a variety of inlet conditions.

The definition of corrected airflow is

Wi V e_
W1R = W1 x /T2/518.67 _ 2 2.3-(1)
P2/14.696 62

The y axis is fan pressure ratio. Also shown are the corrected speed lines

and lines of constant efficiency, called efficiency islands. Thus at a

given corrected speed and corrected flow, there is a defined pressure ratio

at a given efficiency. Finally, the stall or surge line is shown.

The map characteristics are compressed and linearized by transforming

it as a function of exit corrected flow in place in inlet corrected flow.

The new map is shown in Figure 2.3-1B. Thus define exit corrected flow

W 21 V e_
25 2.3-(2)
W21R = W21 /-T25/518.67 _
P21/14.696 625

Then P21/P2 = f(W21R, NI/ v/e) 2.3-(3)

19
 COMPRESSOR CHARACTERISTICS

A.

0
4IJ

(z 1 Fan
Efficiency
Contours

Corrected Speed
Lines

CORRECTED AIRFLOW Wl v'U/ 6

C14
Corrected
Speed
0 Lines

U2
N

P.4

Exit corrected flow W21 V12T/621

FIGURE 2.3-1
20
Now, the temperature ratio must be calculated in order to calculate fan

work. For adiabatic compression, temperature ratio is a function of

pressure ratio and efficiency.

T21 ( - ) /T 2.3-(4)
T2 P2 f)~

If the fan characteristic is simplified, fan efficiency is most strongly a

function of corrected speed. Thus, the final functional relationship used

in the model is

T21/T2 = f(P21/P2, N//vj) 2.3-(5)

Then fan inlet airflow is a function of exit corrected airflow pressure

ratio and temperature ratio. Since W21 = W1, then

W1= W21R x (P21/14.696) 2.3-(6)

v(T21/518.67

And fan horsepower is calculated

HPF = f(W1, T21, T2) (2.3-(7)

The calculation of airflow may seem backward here, but whether it is

based on inlet or exit flow, there is the necessity to iterate the flow to

match continuity requirements. The iteration is done in time rather than

by multiple passes, as is described in the iteration and integration logic

21
section. The above functional relationships consist of polynomials and

segmented linear functions with a total of 54 arithmetic and logical

operations to represent the fan from choke to stall for speeds from 30% to

100%.

The intercompressor duct downstream of the fan splits the flow between

the bypass duct and the core compressor. The inlet states to the high

pressure compressor are treated as fixed fractions of the states at the

exist from the fan.

P25 = C x P21 2.3-(8)

T25 = C x T21 2.3-(9)

The high pressure compressor is modelled similarly to the fan. Exit

corrected airflow was used to compress and linearize the flow-speed-

pressure ratio characteristic. Thus

W3 x v/T3/518.67 2.3-(10)
W3R
P3/14.696

P3/P25 = f (W3R, N 2/ 25 ) 2.3-(11)

23()

T3/T25 = f(P3/P25, N 2 /v'2) 2.3-(12)

W3 = f(W3R, T3, P3) 2.3-(13)

22
Compressor horsepower is calculated from flow and temperature rise.

However, the parasitic bleeds that are used to cool the high pressure and

low pressure turbine components must be accounted for. These parasitics

are modelled as a constant fraction of inlet airflow.

W25 = C x W3 2.3-(14)

HPC = f(W3, T3, T25) 2.3-(15)

One final parameter is calculated because it is a sensed variable: com-

pressor discharge static pressure. Here it is represented as a constant

fraction of the total pressure P3. The implication is that the mach number

at the exit of the compressor is relatively constant. The above modelling

required 55 operations to represent the compressor for speeds from 75% to

100% (the normal operational range) for loading from full acceleration to

full deceleration.

The air from the HPC is diffused, entered into the main combustor,

mixed with fuel and the fuel air mixture is burned. There is a pressure

drop caused by aerodynamic losses, and a temperature rise due to the

burning of the fuel air mixture. The pressure drop is an aerodynamic

function, and is represented as a simple function of corrected airflow.

P41 = P3 x f(W3R) 2.3-(16)

The temperature rise is a function of the fuel-air ratio and the combustor

inlet temperature, T3.

23
 T41 = T3 + f(FAR41, T3) 2.3-(17)

It is necessary to calculate both airflow and mass flow for the turbine

nozzle flow function check. Again, the parasitics are a fraction of total

core airflow:

WA41 = C x W25 2.3-(18)

W41 = WA41 + WFM/3600 2.3-(19)

Note that WFM, the main burner fuel flow, is conventionally expressed in

pounds per hour. The fuel-air ratio is then

FAR41 = (WFM/3600)/WA41 2.3-(20)

These functional relationships required only 14 steps to describe the

process.

The next step in the model is representing the bypass duct to insure

proper back pressure on the fan and the LP turbine. Note here that due to

the required static pressure balance between the core stream and the bypass

stream at the mixing plane, the correct representation in the model will

insure accuracy. The purpose of bypassing a fraction of the airflow is to

increase overall cycle efficiency, while maintaining a high level of

thrust. In a pure turbojet cycle there is high thrust, but also high

specific fuel consumption. For a turbofan cycle, with a high bypass ratio,

the thermodynamic cycle is more efficient but produces less thrust with the

24
same airflow. Thus, the subject engine has the high thrust properties of

the turbojet cycle, with the better SFC properties of the turbofan.

The duct inlet temperature and pressure must be calculated consistently

with the amount of energy exiting from the fan. For the pressure energy,

the following must hold:

P13 x W13 + P25 x W25 = P21 x W21 2.3-(21)

However, the properties at station 14, which is downstream of the splitter

and includes a loss from 13 to 14, is of interest. Thus we calculate

(W14 = W13)

P14 = C x P3 = C(P21 x W14 P25 x W25 2.3-(22)

The thermal energy balance requires that the following equation holds true:

T14 x W14 = T21 x W21 - T25 x W25 2.3(23)

or

T14 = (T21 x W21 - T25 x W25)/W14 2.3-(24)

There is an input of heat into the bypass stream from the core engine

casing in the hot section. This thermal energy is modeled as a simple

convective heat transfer.

25
 T16 = T14 + f(T3, T25, W14) 2.3-(25)

There is also a pressure loss from inlet to exit of the bypass duct due to

normal aerodynamics and irregularities in the core casings. The loss is

calculated based on "pseudo" mach numbers, which are closely related to the

square of the actual mach numbers.

P16 = P14 x (1 - f(W16, T16, P14)) 2.3-(26)

It is necessary to calculate the static pressure at station 16 in

order to preserve the static pressure balance at the mixing plane. It is a

straightforward compressible flow relation, that is simplified for the

computational speed required.

PS16 = P16 x f(W16, P16, T16, AE16) 2.3-(27)

The high pressure turbine consists of a choked nozzle to direct the

hot gases onto the turbine blades, and the turbine wheel, which is on a

common shaft with the high pressure compressor. The nozzle is choked

throughout the normal operating range in order to control the HPC back

pressure, and to prevent aerodynamic fluctuations downstream of the turbine

from effecting the compressor. The nozzle model is simple:

W41 x vT41/P4 = C 2.3-(28)

or the error in nozzle continuity is

26
 EFF41 = C - W41 x vT4/P4 2.3-(29)

This error will be used in the iteration scheme described later.

Turbine work and horsepower are directly related to pressure ratio.

The design of the two turbines and their close coupling produces the

following convenient relationship:

P45/P4 = f(PS16/P4) 2.3-(30)

Turbine horsepower is then

HPHPT = C x W41 x T41 f(P45/P4) 2.3-(31)

Since the HPC horsepower was previously computed, the unbalanced power in

the shaft can be obtained.

UBHP = (HPHPT/(HPC + f(WFM, WFR))) - 1 2.3-(32)

The function in the relation is an approximation of the parasitic

horsepower in the HP spool required to drive the main fuel pump, the

afterburner fuel pump, the exhaust nozzle actuators, and accounts for

bearing losses.

The low pressure turbine further expands the hot gases to extract

energy to drive the fan. Also, some cooling flows extracted from the HP

compressor are re-introduced into the core stream. Downstream of the

27
turbine wheel are swirl, strut and diffusion losses that are included in

this calculation.

The flow through the turbine is that of the high pressure turbine plus

parasitic flows that are a fraction of W25.

W45 = W41 + C x W25 2.3-(33)

The low pressure turbine nozzle is not hard-choked, so the flow function is

a function of pressure ratio:

W45 x VT.45/P45 = f(PS16/P45) 2.3-(34)

Expressed more conveniently, since T45 is desired:

T45 = (P45 x f(PS16/P45)/W45) 2 2.3-(35)

Then turbine work may be calculated

HPLPT = C x W45 x f(T45) x f(PS16/P45, Ni, T45) 2.3-(36)

Since the fan horsepower was previously calculated, the unbalance in the

low pressure spool is found.

UBHP2T = (HPLPT/(HPF + C)) - 1 2.3-(37)

The constant accounts for bearing losses and windage.

28
 The temperature control sensor is mounted downstream of the rotor, and

the sensed temperature is defined by turbine pressure ratio.

T5 = T56 = T45 x f(PS16/P45) 2.3-(38)

The discharge pressure includes the previously mentioned swirl losses,

diffusion losses, and strut losses. These are functions of mass flow,

which is calculated from W45 and cooling flows.

W56 = W45 + C x W25 2.3-(39)

P56 = PS16 x f(W56vT36/PS16, N /T45) 2.3-(40)

After exiting from the LP turbine, the hot core gases are combined

with the bypass stream air, and after mixing flow into the afterburner.

Flow continuity demands that the total flow is the sum of the two streams

W6 = W56 + W16 2.3-(41)

The energy balance results in

T6 x W6 = T16 x W16 + T56 x W56 2.3-(42)

A bias is included in the final form

T6 = C x (T16 x W16 + T56 x W56)/W6 2.3-(43)

29
The pressure at station 6 includes any pumping gain resulting from the

mixing of two streams at different pressures. There are losses downstream

of the mixing plane. When non-afterburning, the loss is simply the

friction and diffusion losses in the tail pipe. When afterburning, there

is additional loss due to heat addition. The pumping and tail pipe losses

are combined:

P61 = P56 x (1 - f(W6 VT6/P56, N1 / V5)) 2.3-(43)

For non-afterburning, the losses to the nozzle entrance are accounted for,

thus

P7 = P61 2.3-(44)

T7 = T6 2.3-(45)

The introduction and burning of fuel in the afterburner provides

substantial temperature rise for thrust augmentation. The reason can be

seen by examining the gross thrust function

FG/(W8 V'T8) = f(P8/Pamb) 2.3-(46)

Thus

FG = C x W8 x VdT x f(P8) 2.3-(47)

30
As T8 increases, the gross thrust will. Then temperature rise is a

function of fuel air ratio:

T8 = T6 + f(WFM, WFR, W6) 2.3-(48)

The pressure loss is a function of airflow and temperature rise (or ratio).

P8W = P61 x (1 - f(W6 V-T?/P56, T7/T6)) 2.3-(49)

The gases are accelerated to sonic velocity in the exhaust nozzle, and

expanded to ambient conditions in the divergent portion of the nozzle. The

nozzle operates choked by design, and there is some leakage. The nozzle

throat flow coefficient is calculated to provide a good calculation of back

pressure on the engine. The flow coefficient is represented as

CF8 = f(P8/Pamb, A8, P8 - Pamb) 2.3-(50)

The throat area is corrected for inward nozzle deflection at low pressure

ratios. Note here that the nozzle position is part of a feedback loop to

the control, and thus the control senses area.

Nozzle flow is computed from the following equation, where C accounts

for leakage, and WFR is afterburner fuel flow.

W8 = W6 x (1-C) + WFR 2.3-(51)

31
Since the nozzle is choked, and all the variables are calculated, there is

a second continuity check. From the choked nozzle

W8
W v'i = const. 2.3-(52)
P8 x A8 x CF8

or

P8C = C x W8/T8 2.3-(53)

A8 x CF8

Thus, a second error is formed by comparing the choked P8, P8C, to that

calculated from known losses, P8W, EQN 2.3-(49)

EP8 = (P8W - P8C)/P8C 2.3-(54)

A summary of the flow of information from inlet to exhaust is shown in

Figure 2.3-(2). It is seen that the critical parameters for each component

are calculated and passed on to the next component. Now the integration

and iteration routine is discussed.

On the very first pass through the equation set, the initial values of

the rotor speeds N and N2 are guessed, as are the two exit flow functions

W21R and W3R. On subsequent passes, the unbalanced torques and the

continuity errors force the changes in these terms. Thus, for the fan

rotor, the following equation gives the new rotor speed.

NI(t + .01) = NI(t) + C x UBHP2T/XJLP/N1 2.3-(55)

32
49 9 I 9j vp w

IWFR
W21
W16
Bypass Duct T16 jI

W5 6 W8. C0
~i1P25j 0 3 W41 W45
H
z tri
0 x

T214Mjj
T2 0 T41 45 T56 N 0j
rt 0x
H"*
T25 S T3 P45 P56 0 ;j P8
rt

ULJ PS3C
HPC
HPF
N N
ii]
W3R N 2NWF EFF41 UBHPT

W21R
Interation and Integration UBHP2T
~Tj
H Logic EP8
0
I r
U.)
This equation is a simple time step interaction of the speed rate of

change.

dN C x UBHPZT/XJLP/N
1 2.3-(56)
dt

Similarly for the high pressure spool

N 2 (t + .01) = N 2 (t) + C x UBHP/XJHP/N 2 2.3-(57)

The two continuity errors are forced to zero by the time step

iteration of the errors. Thus

W21R(t + .01) = W21R(t) + C x EFF41 + C x EP8 2.3-(58)

W3R(t + .01) = W3R(t) + C x EFF41 + C x EP8 2.3-(59)

The gains on these two loops are high enough to decouple them from the

rotor integrations. Therefore, in each time step, there may be a small

error in the continuity errors, but they are quickly forced to be small by

the iteration logic. It is also noted that within each time step there is

no iteration. This completes the description from the reference, and now

some of the detail added since the reference was published are added.

Three additional major effects were added to the above model. These

are off-schedule variable geometry effects, the effects on non-adiabatic

compression and expansion (also known as heat soak), and additional

34
modelling to represent the transient response of the turbine discharge

rakes. The first will effect the amount of flow and a given compressor

speed. The second effects the transient responses of the components, and

the third is to give a fair comparison between the model prediction of the

sensor output and the actual output.

Both the fan and high pressure compressor have variable inlet guide

vanes and stator vanes to better mach airflow and speed at off design

conditions. The main effects are that when the variable geometry is closed

relative to a scheduled position, the flow and efficiency will be lower.

This effect is especially true during transients, where the inertia of the

VG system causes it to lag the scheduled value. Thus, for better transient

matching between the model and engine, the off-angle effects are included.

If the difference between actual and scheduled is termed DVGL for the fan

and DVGH for the compressor, then the following relationships apply. For

the fan, the effect on corrected flow is a function of corrected speed and

DVGL, as is the effect on efficiency. In the model, efficiency is not a

direct parameter, so the off-angle effect changes temperature ratio at

pressure ratio. For the high pressure compressor the effects are a

function of corrected core speed and DVGH.

The heat soak modelling is added to better match the transient

response of the model to the actual engine. Heat soak in the engine is a

reflection of the non-adiabatic processes occurring in the engine. This is

especially true in the turbine, where the temperature gradient from the hot

gas to metal is as much as 1000*F. The approach used is a lumped parameter

estimate of the heat storage in each component. The alternative to this

type of modelling is an elaborate bookkeeping of all parts exposed to the

35
gas path. The lumped parameter model was generated by analyzing data and

comparing it to adiabatic models. The general form is derived from simple

heat transfer. The following equations briefly describe the modelling.

For the metal absorbing heat from the gas stream

MC dTm/dt = hA(Tg - Tm) = Wg C (Tgi = Tg ) 2.3-(60)

p pg i u

Figure 2.3-(3) expresses these equations in block diagram form. The input

to the system is the adiabatic temperature as calculated in the real time

model. The output includes the non-adiabatic effects of heat soak.

The final addition is to realistically model the temperature sensor

dynamics. This temperature sensor consists of eight two sensor probes

located as equally spaced circumferential locations. The sensors have to

be rugged enough to withstand the long operational life of the engines; and

are therefore slow in response time due to the mass. One representation is

a lag with a variable time constant. From convective heat transfer theory,

the time constant will be proportional to Reynolds number raised to some

exponent. The representation is shown in the figure 2.3-(4). The current

main fuel control uses a lead/lag compensation scheme with a variable lead

and a fixed lag. The variable lead is an attempt to model the sensor

dynamics, and the lag is for stability. The overall modeling used for this

study is shown in figure 2.3-(3). This concludes the description of the

real time model.

36
 fe UUW IF
1W

0 a

Tg IN - AT h cp dT M/dt Tm go
hA -]

q 1 ATg
WgCpg
- Tg out H

0
_.j
ni4
n-

Simplify to the following z

0

(ADIABATIC)

Tg(INNND AI
__ _1G =hA (NON-ADIABATIC)
TS+1

- - g out

k)
(~A4
 W56 Time
Constant
W
Calc. T

T56 T56H po (Temperature Sensor Output)

000

e soI

Cj

T561TLA l T56LGL (Best estimate of gas

TS+1 W . S+1 temperature using lead/lag
compensation)

WFM LedT Lead

T/C
 3. ADAPTIVE ALGORITHM DEVELOPMENT

It was previously stated that the goal of the study is to define a

sensible methodology to force the model to "track" the engine outputs. The

first fact that lead to the selected method was the obvious realization that

all models have errors. One traditional approach to correct for modelling

errors has been to use the Kalman filter equations to update the model.

The second fact considered was that even if the model was exactly correct

for one engine in the population of those produced, it would be in error

for many other engines. Thirdly, for those engines that the model matched

exactly when new, as the engine running time accumulates, the engine

deteriorates, and the modelling errors would grow. Finally, the method

selected should be able to correct small modeling errors without corrupting

the model. Model corruption means that it would take a large and unrealis-

tic change in a model variable to force the model to track the engine. The

method should be able to adapt the model for engine to engine variations.

And further, it should be able to adapt the model as the particular engine

deteriorates. The last item has significant implications. If, for example,

the method involved changing a component efficiency ap engine run time

accumulated, one could reliably predict that that particular component was

deteriorating and should be replaced. This capability will enhance engine

condition monitoring (ECM) and provide advanced diagnostics for overhaul

shops. In summary, the method should be capable of changing component

performance in a logical manner in order to track the engine, and a sensible

set of component characteristics must be determined.

39
3.1 Parameter Set Selection

The engine sensors produce four outputs to "track," and thus the size

of the variable set may be up to four in number. The criteria for selec-

ting the set of parameters are the following. First, each parameter should

be dominant in matching one sensed variable. This will result in a quasi-

diagonal manipulation, and should prevent two variables from "fighting"

each other to match a sensor. Second, the change in the parameters to

match the engine outputs around the flight envelope should be relatively

small. This requirement is related to corruption of the model. Third, the

change in parameters should be small when considering engine to engine

variations in component performance. Similarly the parameters selected must

be insensitive to errors in the measured variables. For example, the main

fuel control uses the metering valve position to indicate fuel flow. When

considering the variations from metering valve to metering valve, broad

spec fuels, fuel density variations and the like, the relationship between

metering value position and fuel mass flow rate may vary by as much as 5%.

Similarly, the exhaust nozzle area feedback signal may be in error by as

much as 4%. Finally, the synthesis of important parameters such as turbine

inlet temperature should be insensitive to all of the above problems.

The questions to be answered are the number of parameters to adjust

and the type of parameters to adjust. The first idea is that by varying

four parameters, there is more flexibility, but perhaps if only three are

matched, the errors in the fourth are small. The second concerns itself

with whether only engine performance parameters alone are selected, or

should biases by imposed on the control inputs to account for errors in

40
measuring fuel flow. The specific questions to be addressed are: should 3

or 4 engine sensors be tracked; what combination of engine parameters and

sensor biases be used; and are the parameter changes required to match

engine performance acceptably small. These questions are answered based

solely on static performance of the model. The dynamic considerations are

attacked in section 3.2.

The first step is to identify potential performance parameters that

can be changed to correct for modelling errors. A logical starting place

is engine to engine variations that exist in the production population.

These variables are listed in the table below.

TABLE 3.1-1 ENGINE VARIABLES

if Fan Efficiency
WIR Fan Corrected Flow @ Corrected Speed

Tc High Pressure Compressor Efficiency

W25R HPC Corrected Flow @ Corrected Speed

DB Combustor Efficiency
Tit HP Turbine Efficiency
FF41 Turbine Nozzle Area Flow Function

12t LP Turbine Efficiency

Wc Percentage Cooling Flow

In addition to these variables, two control variables may require biases to

account for inaccuracy. These are the fuel flow, Wf, and exhaust nozzle

area A8. The two could be in error by as much as 5% and 4% respectively.

Static derivatives are taken with respect to each variable to determine

the effect of each variable on engine sensed parameters. Since the engine

is a two degree of freedom system, two inputs are held constant while

41
taking derivatives. For the nine engine variables, the fuel flow and

exhaust nozzle area were held constant. For example, a one point change in

turbine efficiency (one point equals 1% change) at constant fuel flow and

nozzle area resulted in -14 RPM on the low pressure shaft, +65 RPM on the

high pressure shaft, +0.5 PSI on compressor discharge static pressure, and

-2* in turbine exit temperature. Several parameters were eliminated upon

examination of the derivatives. The effects of a change in high pressure

compressor efficiency are indistinguishable from those of high pressure

turbine efficiency for the variables measured. The two can be isolated

with a measurement of HPC discharge temperature, which is unavailable for

the subject engine. Therefore, eliminate compressor efficiency. Burner

efficiency only effects specific fuel consumption strongly, therefore

eliminate it. The percentage cooling flow is similarly weak, so it too is

deleted from the list. Now six engine variables remain, and further work

is required to assess the selection.

Sets of variables and control biases are made, and the weeding out

process begins by determining the magnitude of changes required to force

the model to track the engine. This exercise requires the establishment of

a "truth model" to base the model against. The truth model could be actual

engine data from one particular engine, or it could be data generated by a

highly detailed description contained in a mammoth computer simulation.

The latter is used in this case. It is an internal General Electric model

of the subject engine, which is a very large scale dynamic simulation

including very complex component performance representations including

vitiation and dissociation effects. It is incapable of operating in real

time. For example, the description of fan performance requires over 50

42
equations and hundreds of arithmetic and logical operations. This model

describes the so-called nominal engine behavior. Thus the comparison is

established between the real time model and the nominal engine at important

flight conditions around the flight envelope. At a given flight condition

there will exist differences between the real time model and the truth

model in fan rotor speed, NI high pressure rotor speed, N 2 , compressor

discharge static pressure, PS3C, and turbine exit temperature T56. In

order to eliminate the errors, four variables in the model are changed to

match them. For example, employ changes in fan efficiency, turbine effi-

ciency, fan speed-flow (WIR) and compressor speed flow (W25R) to account

for differences in N1 , N2 , PS3C, and T56. If it takes +1 point in fan

efficiency, -.5 PTS in turbine efficiency and +0.3% in flows to eliminate

the differences, then one may say that the modelling errors are equivalent

to component mis-representations of this magnitude. Alternatively stated,

the real time model has a nominal implied delta nF of zero points, but an

actual delta of one point is necessary. Thus, through the following

discussion, the parameters involved are deltas from a nominal value of

zero. It may be noted the component efficiencies are not directly calculated

in the real time model. There is no equation which states that nF is 84%,

for example. They are implied in the equation of temperature ratio at a

pressure ratio.

The combinations are selected to produce diagonal dominance in the

derivative matrix. Twelve possible sets are explored. Four sets are for

matching all four measured engine outputs. Eight sets are for matching

only three outputs. The sets are summarized in Table 3.1-2. The table

shows the combinations of engine variables and sensor biases, along with

43
 0 Model Check-out Points

60- N-I/

4-I

40-...
4-4

01-
0 0 0

20

0
I 2
0 .4 .8 1.2 1.6 2.0

Flight Mach Number

FIGURE 3.1-1

44
the sensor outputs to be tracked. The first four sets track all four

outputs, the second four sets track all except PS3C, and the last four

track all outputs except

N2

.
TABLE 3.1-2 SETS OF PARAMETERS

All Parameters are Deltas from a Nominal Value of Zero

Set Number Engine Variables Sensor Biases Matched Sensors

1 WIR, t FF41, nF -- N1 , N 2 , PS3C, T56

2 WIR, nt9 ' F WF N1 , N 2 , PS3C, T56

3 t FF41, nF A8 N1 , N 2 , PS3C, T56

4t F A8, WF N 1, N 2 , PS3C, T56

5 WIR, nt' F -- N1 , N 2, T56
6 WIR, nt WF N 1 , N2 , T56

7 t F A8 N 1 , N2 , T56
8 nt A8, WF N1 , N 2 , T56

9 WIR, FF41, n-- N1, PS3C, T56

10 WIR, , t WF N1, PS3C, T56

11 FF41, t A8 Ni, PS3C, T56

12 , t A8, WF Ni, PS3C, T56

These sets were tested for tracking around the flight envelope. The

points in the flight envelope are depicted in Figure 3.1-1. As can be

seen, the model was checked out at all extremes of flight c6nditions. The

results were judged upon the size of parameter changes, and the ability to

synthesize turbine inlet temperature T41. The findings are shown in

Appendix 1, which contains the detailed results. Two thrust levels were

45
explored. The first is intermediate rated power (IRP) which is the highest

thrust without afterburner, and the second is maximum afterburner, which is

the greatest possible thrust setting.

The following conclusions are made based upon an examination of the

results of this segment of the study. Those sets that bias A8 provide

marginal improvement in T41 synthesis when compared to those that do not.

Also, in many cases the bias became greater than 10% which is unacceptable.

Note here that A8 is "known" from the large scale model, and should not

require a bias. Finally, at low power settings, as changes have a low

influence, especially at 400 square inches. The conclusion is to eliminate

A8 as a potential variable. This step eliminates two of the four variable

sets and four of the three variable sets. For the remaining four variable

sets, the T41 synthesis is virtually identical, but the variations on fuel

flow are small compared to those on turbine nozzle area. Thus this set,

number 1, is eliminated. The remaining sets of 3 variables matching N 2 are

comparable and are not eliminated at this point. Set number 9 is elimi-

nated for the same reason as number 1 above. In summary, sets numbers 2,

5, 6, and 10 remain.

These four remaining sets are further judged by their performance in

consideration of possible sensor errors. The known sensor errors for the

following sensors are tabulated below. These errors are judged to be

random and uncorrelated, and therefore, to assess the effects, the

root-sum-square (RSS) technique is employed (see Appendix 1 for the

detailed results of this analysis). The effects of these sensor errors

were studied at sea level static conditions and at 60000 Ft/mach no = 1.

The conclusions eliminate two sets. Set number 5 is eliminated because of

46
large sensitivities of fan efficiency and fan airflow. Also, the synthesis

of PS3C is poor. Set number 6 is eliminated due to a large sensitivity on

fan airflow. This leaves sets 2 and 10.

TABLE 3.1-3 SENSOR ERRORS

Fan Speed N 0.5% Full Scale (14000 RPM)

Core Speed N2 0.5% Full Scale (17000 RPM)
Static Pressure PS3C 0.25% Full Scale(450 PSIA)
Temperature T5 16c
Fuel Flow WF 5% of Point
Nozzle Area A8 4% of Point

The final selection criteria was to subject the two remaining sets to

random engine to engine variations, and check the effect on the synthesis

of T41. The engine to engine variations arise out of manufacturing toler-

ances and buildup procedures. The General Electric Design Practice suggests

the following as good assumptions on these variations.

TABLE 3.1-4 ENGINE TO ENGINE VARIATIONS

Fan Efficiency flf 0.6 PTS

Fan Speed-Flow WiR 1.0%
Compressor Efficiency Tc 0.7 PTS
Compressor Speed-Flow W25R 1.0%
Turbine Efficiency t 0.8 PTS
Turbine Flow Function FF41 1.0%
LP Turbine Efficiency U2t 1.0 PTS
LP Turbine Flow Function FF45 1.0%
Cooling Flow 0.4%

47
The final conclusion is to use the 4 variable set due to smaller sensitivi-

ties to engine to engine variations. See Appendix 1. It is also more

flexible due to the additional parameter. Now the work can proceed to the

dynamics of the update.

3.2 Gain Matrix Methodology

The task at hand is to develop the means to update the model in a

dynamic sense that insures stability of the model while updating. The

dynamics of update should be relatively fast for two reasons. First, model

errors at low power will, in general, be different than model errors at

high power. Since the response time of the engine is about 4 seconds, the

update scheme should be about this fast. Second, if the model is to be

used for dynamic control, then it should be actually faster than the engine

response time in order to successfully track transients. The problem,

in essence, becomes a multidimensional control problem, where the model

should follow the reference of the engine sensors in real time. The

criteria for judging the success of the dynamic update method are the

success in tracking the engine, the minimization of additional processor

time, and the minimization of additional through-put (variable names) to be

carried along.

Four methods with increasing levels of complexity are explored. The

first uses the inverted Jacobian Matrix as a gain matrix on the errors. It

is scaled as a function of power setting to insure stability. The second

is a method inspired by recursive least squares which penalizes both errors

between the engine and model and changes in the model parameters. The

third is the so-called constant gain extended Kalman filter (Ref. 8, 10,

48
12). The final is the extended Kalman filter. Each method is developed in

this section, and the results are shown in Section 4.

3.2.1 Inverse Jacobian Gain Matrix

The following method is deemed the simplest possible method of

updating the parameters in order to track the engine. Start with some

definitions. Let the vector r be the vector of changes in model

parameters.

r =d [AWIR, An t AWfJ* An f] 3.2.1-(1)

Let the vector Y be the vector of engine outputs.

Then the Jacobian is defined as the derivative of Y with respect to r

tY
-r
dr ij=dr.J(AYi
iJ = i
Io j) 3.2.1-(3)

and therefore

AY = J x Ar 3.2.1-(4)

Now, we define a vector e as the vector of output errors between the model

and the engine. To eliminate this error in the linear case, solve for Ar

that eliminates e

e Yr Y 3.2.1-(5)

Ar = J1 e

49
Since the engine and model are non-linear, iteration is necessary. This

iteration is handled by performing a time step integration of the vector

Ar. Thus

r (t + .01) = r(t) + a x (J~I x e) x .01 3.2.1-(6)

In the equation, alpha is a constant to scale the Jacobian to insure

stability.

This method ignores the dynamics of the model and engine. This fact

can be seen by examining the meaning of the Jacobian for this case, which

is the relation between r and Y steady state. Take a simple two degree of

freedom representation, and write out the dynamic relations to show X and Y

as functions of r.

1 a 11 a12 x1 b1
+ r
2j a 21 a 22 X2 b2j

3.2.1-(7)

Y= E
d C] 1

Lx2j

Then the transfer function from u to x in Laplace form is

X(s) = (sI - A)- Br(s)

= + 'r(s) 3.2.1-(8)
a2s + b2

a3s2 + b3s + C3

50
And the output is

1 1
Y(s) = c x (s) = c s + d x r(s) 3.2.1-(9)
a3s + b s + c
3 3

Thenthe Jacobian, defined for this case by(dy/dr) is the transfer

function at time equals infinity.

Ss
+ d = 3.2.1-(10)
a3 s s 3s 3 t=oo 3

It can be seen that the Jacobian ignores the dynamic terms in the transfer

function.

Stability of the update scheme was explored by performing a state

space analysis of the dynamics at a high power condition. The model

contains two states, N1 and N2 , plus the four dynamic variables become

states. The state space representation is derived in two steps. First, a

reference state and input condition are determined at sea level static

(SLS) IRP. The following are defined

States x A 6N
6N Perturbations from nominal
L61 21

Outputs Y A 6N
6N 2 Perturbations from nominal
6PS3C
66T5
-

51
Inputs
u A 6W
- = f
= PERTURBATIONS FROM NOMINAL
SA8

Then
k = Ax + Bu
3.2.1-(11)
Y = Cx + Du

is the state space representation. Now impose the condition that u = o.

Equation 3.2.1-(11) becomes

k = Ax
3.2.1-(12)
Y = Cx

These equations are applicable to the nominal value of the parameter vector

r, which is zero. Now allow r to vary, and form a new representation.

k = A x + B' r
3.2.1-(13)
Y = C x + D' r

The update equation is

. =-1
r_= axJ e

3.2.1-(14)
1
r -J x (Y g- Cx -D' r)
_a J =lx eng d
The new closed loop update method (for a fixed engine reference) is

A B' x 0

Y 3.2.1-(15)
-eng
Li -C - aJ1J '
]r . LaJ-
.

52
or X = A'X'+B'Y 3.2.1-(6)
- - - -eng 321(6

The closed loop eigenvalues of the A' matrix determine stability. A

multi-dimensional "root locus" for varying values of are shown in Figure

3.2.1-(1). A value for a of 2 was chosen at high power and .74 for low

power. The reason for the low value is that the Jacobian reflects static

relationships and must be reduced for stable dynamics. The results are

shown in Section 4.

3.2.2 Recursive Least Scuares Derived Gain Matrix

This next method is analyzed because of the ability to penalize large

changes in the parameters and, in effect, prevent corruption of the model.

The previous method will move the variables with no consideration of their

magnitude. Recursive Least Squares is used as a tool for model validation

when large amounts of data are available. The goal is to process the data

sequentially and carry the information from one pass to the next. This

avoids the complications of carrying large matrices that Classical Least

Squares employs.

The adaptation of Recursive Least Squares to this problem is shown in

the equations that follow. Start with a performance index J , and use a

linear model between the vector r and the error e. The performance index

penalizes errors as well as changes in parameters. The subscripts refer to

the indicated time step.

T T
J A e R e + (r. -r. R (r - r ) 3.2.2-(1)
pi i -i -1-- 2- --
-i

53
 POLE PLOT
SCALED INVERSE JACOBIAN F/B MATRIX

-1
r =r + ctJ e
Im

-6

A -4

-2

2 Poles Re
A aw A /I ..-..A 0
- - I 1 -4
RAD/~S -10 -8 -6 -4

-2
Symbols [I]
M
x: ct= 1
0(]: X: c = 2
3 A
-4

-6

FIGURE 3.2.1-1

54
Assume the relationship between e and r is linear using the Jacobian, J

e. (r) x (r. - r. 1) + e._1 3.2.2-(2)

where
dei
J A = JACOBIAN
dr.

For simplification let R in the performance index bet set to I, and let R2

be XI such that we are penalizing accuracy and parameters relatively

through Ot.

We seek to minimize the performance index with respect to r . Thus,

take the partial derivative and equate it to zero.

dpi d T T(r r 0
dr. dr r
r-. i*
-el -i + a(-i -_)(r - r
1I- )_) =
-

-1 -1

de.T 3.2.2-(3)
-1
2 d e. + 2 d (r. - r. = 0
dr. 1 -1

We wish to solve for a relation that updates r. from r , thus solve for

r_ -r _, and note the definition of J.

aI (r - r )=-Je

Employ equation 3.2.2-(2) for the definition of ei.

55
 oI (r - r ) = -J (J (r - r ))+ ej1 3.2.2-(4)

Thus (J + cI) (r -r ) -e

or r -r, -- (J + aI) Je 3.2.2-(5)

2 3.2.2-(6)
r r.i 1 +Ke 1 ; K = -[J + il-] 1

Therefore, in the linear case we have the above calculation of K in each

time step, to update the parameters r.

The non-linear case requires the derivation of the Jacobian in each

pass. However, for ease of the computational burden, the following

simplification can be made. The Jacobian can be preprogrammed as a

function of corrected fan speed. The reasons for this effect are two-fold.

First, corrected fan speed is indicative of the thermodynamic processes for

the fan shaft, and thus the derivatives for variables related to the fan

shaft, act consistently with corrected fan speed. Second, there is a tight

relationship between corrected fan speed and corrected core speed.

Therefore all parameters related to the core can be related to fan speed.

With this information, use the following logic as the update gain matrix.

Let a now be a vector penalizing the diagonal elements of R2 individually.

56
Let the Jacobian in each time step be a function of the corrected fan speed

JACOBIAN A J. = (N /v) = J (N //).

---- 1 -i 1 1 2 i

Then define the gain matrix K

i

K. A [J. 2 + UI] J 3.2.2-(7)

Finally update the parameter vector r once per time step

r (t + .01) = r (t) + K e (t)

In order to control the size of the Ar, let a be a function of ri such

that at high r , the Ar is small due to a large penalty from _t.

ct. = yr.

In other words, as r increases, the penalty on Ar increases.

In the actual iteration scheme, the inverse in equation 3.2.2-(7) is

not taken, in order to save computational time. The results are shown in

Section 4.

3.2.3 Constant Gain Extended Kalman Filter

The following method is based on the well-known Kalman-Bucy Theory

(Ref. 13) with significant modifications. In the basic form, the Kalman

57
filter is a minimum variance, unbiased state estimator for the continuous

linear system described by the following equations.

x(t) = F(t) x(t) + G(t) U(t) + W(t) W(t) % N(O, Q(t))

Y(t) = H(t) x(t) + V(t) V(t) % N(0, R(t)))

where

x(t) is the state vector

U(t) is the input vector

W(t) is the process noise-Gaussian with zero mean and covariance Q(t)

Y(t) is the output vector

V(t) is the measurement noise-Gaussian with zero mean and covariance R(t)

The best estimate of X, given by x is

= F(t) Z(t) + K(t) [Y(t) - H(t) X(t))] 3.2.3-(1)

where K(t) is the Kalman gain matrix.

The theory is further detailed in Section 3.2.4 which explores the

full equation set for the extended Kalman filter. Suffice to say here that

a significant computer burden may be relieved by precomputing a gain matrix

based on a time invariant linear model of the non-linear engine. This

approach has been suggested in several works (ref. 12, 13) and executed in

several others (Ref. 8, 10). Therefore the method is explored here, but

with the addition of the r parameters included as states.

58
 In the previous section, a state space representation of the engine

was derived for a high power condition. Reviewing that derivation, the

perturbations around the nominal point results in a representation

= Ax+B u
3.2.3-(2)
Y C x + D u

Let the control perturbations be zero, and introducing the parameter vector

r results in

_ = A x + B' r
3.2.3-(3)
Y = C x + D' r

Now, make the four variables in K states, whose nominal trajectory is zero.

r= 0

An alternative method is the so-called adaptive control reference model

approach as outlined in Ref 14.

Therefore, form a new system

k [A B' x
+ B 6 ; u = 0
Li 0 o rJ 0

Y =C x

3.2.3-(4)

59
Finally a Kalman filter may be employed. Let

x = x

The equations become

S= A' x'

3.2.3-(5)
Y = C' x'

And the Kalman filter becomes

x' = A' x' + K (Y - Y)

3.2.3-(6)

Y = C' S'

The above equations apply for the linear time invariant case. Now

introduce the non-linearities.

= f (9, t) X + K (Y - Y)

3.2.3-(7)
Y =h (x, t) X^

Here the f and h functions represent the non-linear real time model, which

is being used as the state estimator.

If the K matrix is derived at one condition, and allowed to be

constant, then we have the constant gain extended Kalman filter. Thus, the

method is to specify the Q and R matrices, and determine the steady state

gain, K.,0

Ko = POO H Roo 3.2.3-(8)

60
 In order to calculate the filter gains, the Q and R matrices must be

specified, and the steady state Ricatti equation solved. The filter gain

is computed from

K = PHT R~

Where 0 =FP + PFT + Q - PHTR~1HP 3.2.3-(9)

The guidance for specifying the Q and R matrices are from Ref. 10. The Q

matrix is a measure of process noise, which for this case implies modelling

error. It is assumed to be diagonal. For the elements corresponding to

N and N2 the terms were selected as representing the standard deviation

equal to 1% of maximum value. The remaining terms, corresponding to the r

parameters, were chosen to be 3% based on the results of the envelope

survey in Section 3.1. The R matrix (sensor noise) was chosen to be

similarly diagonal, with the elements corresponding to the known errors in

measurement. The errors are previously detailed in Table 3.1-3.

A check on the open loop and closed loop eigen values is shown. For

the open loop case, the poles are at -6.4, -3.2 and four zeros for the r

parameters. The closed loop poles are at -2.46, -5.01, -5.75 4.79j, and

-6.82 5.45j. The filter should be adequately fast. The implementation

and results are shown in Section 4. The resulting K matrix is

K = 8.18 0.54 1.18 0.7631

0.67 6.99 9.24 1.920
-0.098 -. 024 0.807 -.1288
0.009 .128 0.196 -.111
0.034 .014 0.764 .2868
0.109 -.036 0.475 -.1967_

61
The details of the specific derivation are shown below. The linearized

model at sea level static IRP is shown in the equations below.

States: x = [N, N2, AWlR, ATt f

Outputs: Y' = [N1 , N 2 , PS3C, T5]

Equations: k_ = F x'

Y' H x'

Definitions: dx.
F
1 (Ax. = 0, i = j)
I

dx.
H I (Ax. 0, i # j)

Detailed matrices:
F = -5.5 3.17 -325 -62 100 3771
0.63 -4. 08 -44 363 113 -105
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0]

H = 1 0 0 0 0 0
02 1 0 0 0 0
.005 .025 1.63 0 1.04 .82
.023 -. 055 -4 -5 0 -7

These matrices were determined using the full non-real time computer model.

62
3.2.4 Extended Kalman Filter

The extended Kalman filter is explored to give a reference in perfor-

mance for the "best" update algorithm. The following material is taken

heavily from Reference 13.

The continuous extended Kalman filter equations are summarized in

Table 3.2.4-1. Shown are the system model and measurement model, the

assumptions about initial conditions, the state estimate, the error covari-

ance and the gain equations. The execution of all the implied bases

requires considerable processor time. General simplifications may be made,

while still retaining the attractive features of the filter.

TABLE 3.2.4-1 SUMMARY OF CONTINUOUS EXTENDED KALMAN FILTER

(Ref. 13)

System Model k(t) = F(x(t),t)) + W(t); W(t) % N(O, Q(t))

Measurement Model Z(t) = h(x(t); t)) + V(t); V(t) % N(O, R(t))

Initial Conditions x(o) % N(X(o), Po)

Other Assumptions E[W(t) V T)] = 0 all t and T

State Estimate = F((t), t) + K(t) [Z(t) - h(^(t), t)]

Error Covariance P(t) = F x P(t) + P(t) x FT + Q(t) - P(t)HTR 1(t)H P(t)

Gain Equation K(t) = P(t) + H R1 (t)

Definitions
dF (x(t), t)
F = dx(t) x(t) = X(t)
dh (x(t), t)
H dx (t) x(t) = i(t)

63
 The following simplifications should not detract from the filter

algorithm, and are in essence, assumptions about the engine behavior and

sensor behavior. First, the Q and R matrices are assumed independent of

time and the state. The Q matrix is a measure of process noise or modelling

error. Although the model tends to be more accurate (relative to the truth

mode) at high power, this would lead to additional complication with

marginal results. The R matrix is a measure of sensor noise. For the two

speed signals, the noise is fairly independent of rotor speed. The PS3C

signal accuracy is essentially a percentage of full scale error, and thus

is independent of level. The same holds for the T56 signal. Thus the time

dependence on Q and R may be eliminated. The second major simplification

deals with the F and H matrices. As in Section 3.2.2, these will be

treated as functions of corrected fan speed (of the model) only. Thus they

are functions of one state, and the relationships may be preprogrammed.

The details of the extended Kalman filter result in the following

logic. Apply the above definitions:

F(X(t), t) = F(N / 2/
1 2

H(R(t), t) = H(N/
2)
3.2.4-(1)
Q(t) = Q

R(t) = R

The initial condition assumptions are straightforward. For X(o), use the

value to which the model converges at time zero. For the initial covariance,

assume that the initial conditions error standard deviation is equal to the

Q matrix.
The sequential processing tasks are then:

64
For time equal zero

K(o) = P(o) HT ( R7 3.2.4-(2)

X(o) = X(o)RTM + K(o) [Z(o) - Z(o)RTMI 3.2.4-(3)

P(o) = P(o) F(o) + FT (o) P(o) + Q

3.2.4-(4)
T -1
- P(o) H (o) R H(o) P(o)

For subsequent time steps

K(t) = P(t) HT (N /) R' 3.2.4-(5)

X(t) = x(t) RTM + K(t) [Z(t) - Z(t) RTM 3.2.4-(6)

P(t) = P(t) F (N /'e) + FT (NI/Me) P + Q

3.2.4-(7)

- P(t) HT (N/V6) R~1 H (N /v) P(t)

P(t + .01) = P(t) + P(t) x .01 3.2.4-(8)

X(t + .01) = x(t) + X(t) x .01 3.2.4-(9)

The last two are the simple time step integration of the time derivatives.

3.3 Robustness Issues

The robustness of the update can be judged in several ways, and the

analysis depends upon the update method used. As seen in the Results

65
Section, and in the Conclusions Section, the selected update method is the

Constant Gain Extended Kalman Filter. Therefore, the robustness analysis

is done for this method. A Bode plot of the singular values of the Kalman

filter judges the robustness of the update. The following information

reviews singular value theory from Ref. 15.

3.3.1 Singular Values

The singular values of the transfer function matrix indicate whether

the response of the system at a given frequency is "large" or "small."

They are equivalent to the gain at a frequency for the single input-single

output system, and may be plotted similarly on a Bode plot to determine the

multi-variable frequency response of any MIMO system.

Start by defining the complex conjugate transpose of a matrix F which

H
equals F

F =
F L3 -
,7 FH = [2 - 3j, 3 + j]
j
.

Now define the spectral norm:

MAX
JFZ 12
IFI 2 = 3.3.1-(1)
Z fo0 11 Z 12

Then

2 =
IF1 MAX
1 . {FFHF 3.3.1-(2)

66
Where Xi are the eigenvalues of FHF which are real, non-negative.

Now, define the singular values of F:

0 (F) = A H F 0 3.3.1-(3)
11

Then the maximum singular value is

a
'MAX (F) = M [a (F)] = IF 2
3.3.1-(4)

From equation 3.3.1-(2). Also the minimum singular value is

UMIN(F) = 1 F 2 3.3.1-(5)

Now apply these definitions and relations to control theory, as from Ref.

16. Consider a mimo system with input u(s), output Y(s), and transfer

function matrix G(s). These are related below, in typical fashion.

Y (s) = G(s) x u(s) 3.3.1-(6)

Take the complex conjugate transpose, and divide both sides by 11ul12

IY(s) 12 I IG(s) u(s) 112

3.3.1-(7)
1u(s)122 Iu(s)112

67
The magnitude of the RHS of this equation depends on the vector U(s). The

largest magnitude that the LHS of this equation can attain in calculated

from the singular values of G(s):

MAX j|G(s) u(s)112 3.3.1-(8)

u(s) # 0 = A (G(s)) = A G G
I1u(s)H 2

The same holds for minimum values.

Thus G(jw) is small when its maximum singular value is small, and

G(jw) is large when the minimum singular value is large. It is seen that a

plot of the singular values verse frequency yields the sensitivity as a

function of the input frequency, as in the SISO case. Thus, at low frequ-

ency, a large minimum singular value is desired, and at high frequency, a

small maximum singular value is desired.

3.3.2 Robustness Analysis

The assessment of robustness depends on the BODE plot of the singular

values of the closed loop constant gain Kalman filter. The singular values

of the filter algorithm are evaluated at SLS IRP power setting. The

linearized approximation from Section 3.2.3 is used to evaluate the closed

loop singular values. Recall that, for the linear Kalman filter, the

following equations apply:

Filter x = F2 + K(Z - Z)

Output Z = Hix 3.3.2-(1)

68
Reducing yields X = (F - KH) x + KZ 3.3.2-(2)

The singular values of the closed loop matrix (F-KH) determine robustness.

A controls design package determines the minimum singular value and

maximum singular value as a function of frequency for the closed loop

system. Figure 3.2.3-1 depicts the two singular values plotted against

frequency in radians/second. The typical high "gain" at low frequency is

exhibited along with the decreasing "gain" at high frequency. The 0.0 dB

points are at 1.5 rad/s for the minimum singular value at 13 rad/s for the

maximum singular value, and both fall off at about 20 dB per decade. Now

these characteristics must be judged against the unmodeled dynamics of the

engine and high frequency sensor noise.

The unmodelled dynamics include the gas volume dynamics within the

engine, and the sensor dynamics. The first type by referring to a study of

the Pratt & Whitney F100 Turbo-Fan engine from Ref. 17. That engine is

similar to the subject engine: it is a two spool, low-bypass augmented

engine, but of a higher thrust class due to its size. However, the dynamics

will be similar. A sixteen state model of that engine shows eigenvalues

ranging from -0.66 to -573.4 radians/second. The table below shows them in

detail.

TABLE 3.3.2-1 EIGENVALUES OF F100 ENGINE MODEL

1) -0.66 11) -39.71

2) -1.964 12) -47.52
3) -2.732 13) -49.38
4,5) -6.102 t 2.259j 14) -58.26
6,7) -18.62 t 5.253j 15) -177.7
8) -19.45 16) -573.4
9,10) -20.66 0.921j

69
40 SS to
0 S S

31-DEC-84 SINGULAR VALUES FOR

13:24 KALMAN FILTER
MAG
OPEN
- 40
GAIN=1.0
PH 100.
PL

CL DISPLAY:
WHAT NEXT?
- 20

10.- Robustness
Boundary
1.5 13
0-
1.0 RAD/S

.10 10. 100.

'Jo

.10 4-- 0

a .01 +-4 0
ni
dB
CO

CO
tO
 The first two relate to the "slow" heat storage in the hot section

turbines, the next three include rotor dynamics and the volume dynamics in

the afterburner. The remainder are temperatures and pressures that relate

to the fast volume dynamics of the engine. Therefore, the principle

concern are the filter response for the higher frequencies, which are

unmodeled. Before continuing, it should be noted again that the subject

engine is smaller in size and therefore, all gas dynamics will be faster

than this FINO engine, and tend to push the poles that are of the order -20

to -25 or faster based on airflow and volume ratios.

The magnitude of the effects of gas volume dynamics must be less than

about 8 dB at 25 rad/sec in order to have a robust filter. This is the

only frequency of concern, and further investigation may be required. The

higher frequency unmodelled dynamics are considered not to be of major

concern, due to the slope of the maximum singular value characteristic.

The sensor dynamics can be ignored if the frequency is high (as the

noise typically is). Another method to design in robustness to sensor

noise is to allow the filter to update the model only for residuals that

are greater than the known range of sensor errors.

Thus, for the NI
sensor, the update would be "turned off" if the difference in N was less

than 0.5% of full scale.

In summary for this section, the constant gain Kalman filter algorithm

is robust against unmodelled dynamics based on the known high frequency

dynamics typical of low bypass turbofan engines, and against sensor noise

due to not updating within the bounds of known sensor noise.

71
 4. NUMERICAL RESULTS

The results of this study are divided into two sections: steady state

and dynamic. The static results display the required parameter adjustments

that force the model to track the engine on a steady state basis. They

also show some of the capabilities of the update in terms of T41 synthesis,

condition monitoring and adaptability. These are compared to the model

without an update. The dynamic performance compares the four candidate

update methodologies against the model without an update.

4.1 Static Performance of the Update

The magnitude of the parameter changes required to match the "truth"

model determine the update's ability to detect sensor failures. If the

changes are small, then failures are more easily detected, but if they are

large, the model cannot detect small (or drift) failures. These results

are summarized for three types of tests. The first shows the comparison

between the engine model and the nominal "truth" model around the flight

envelope. This will show the initial "size" of the four dimensional box

that encloses possible nominal modelling error. The second test displays

the updating vector's reaction to normal engine to engine variations. The

third test shows the method's reaction to sensor errors. The important

judging criteria is the synthesis of turbine inlet temperature, T41,

compared against the model without an update.

The first step is to establish the baseline performance of the model

on a stand alone basis. The model results around the flight envelope are

72
shown in Table 4.1-1 for comparison to the truth model. The real time

model is given the same control inputs (fuel flow and exhaust nozzle area)

as the truth model, and the modelling errors on the four outputs plus

thrust and T41 are summarized.

TABLE 4.1-1 BASELINE PERFORMANCE OF THE MODEL

Sensors Performance

ALT XM PLA EN % EN % EP 3 ET 5% ET41 0 R EFN%

0 0 87 +0.9 -0.4 0 +0.0 -1* +3.3

0 1.1 87 -0.5 +0.1 -3.1 +1.1 +410 -3.8
10000 1.4 87 -0.5 0.0 -4.2 +2.6 +490 -8.3
15000 0 87 +2.1 +0.2 +1.4 -0.3 -3* +3.2
35000 0.8 87 +1.3 +0.2 +2.3 -0.9 -15* +1.3
35000 1.2 87 +0.9 -0.2 +0.5 +0.3 -10 +1.3
35000 1.6 87 +0.5 -0.4 -2.1 +1.0 +250 +0.6
35000 2.0 87 -1.0 -0.2 -4.3 +2.1 +400 +8.9
60000 1.0 87 +5.6 +2.2 +2.5 -1.2 -9* +0.7
60000 2.0 87 -1.2 -0.1 -3.7 +2.2 +370 +11.7
0 0 130 +1.4 -0.4 +0.6 +0.6 +150 +4.3
0 1.1 130 +0.5 -2.7 +1.3 +1.3 +380 -5.9
10000 1.4 130 +0.3 -3.9 +1.9 +1.9 +460 -11.1
15000 0 130 +2.7 +1.0 0.0 0.0 +90 +3.0
35000 0.8 130 +2.2 +1.9 -0.7 -0.7 -50 -1.6
35000 1.2 130 -0.5 0 +0.6 +0.6 +140 -4.3
35000 1.6 130 +0.8 -2.1 +0.8 +0.8 +260 -6.6
35000 2.0 130 0 -3.9 +1.5 +1.5 +360 -9.8
60000 1.0 130 +3.6 +3.7 -2.0 -2.0 -41* +3.3
60000 2.0 130 -0.4 -3.3 +1.5 +1.5 +330 -6.1

The average errors and standard deviations are summarized in Table

4.1-2. For the two speeds, PS3C, T5, and thrust, the errors are in

73
percent, whereas for T41 they are in degrees Rankine. The update method

should improve the predictions of T41 and thrust by eliminating the four

output errors. This information established the basis for judging the

effectiveness of the update.

TABLE 4.1-2 MODELLING ERROR STATISTICS

Parameter Average Error Standard Deviation

0.9% 1.7%
1
N 0.2% 0.7%
2
PS3C -1.0% 2.6%
T5 0.6% 1.2%
T41 +14.50 230
FN -0.8% 6.1%

The adaptation of the model for the parameter set of interest was

discussed in Section 3.1. This set is composed of an adder to fan effi-

ciency, an adder to high pressure turbine efficiency, an adder to fan

speed/flow relationship and a bias on the metering value measurement of

fuel flow. The results of the envelope survey are shown in the appendix in

Table A-2. They are summarized in Table 4.1-3 below.

TABLE 4.1-3 IMPROVEMENTS BY MODEL ADAPTATION

Average Error Standard Deviation

Parameter or Delta of Error or Delta
T41 -1.0* 130
FN -1.1% 6.4%
Fan Eff. 0.8 PTS 3.7 PTS
HPT Eff. 0.3 PTS 1.6 PTS
Fan Flow 1.9% 2.0%
Fuel Flow 0.2% 0.8%

74
 The table shows the average errors and standard deviations for T41 and

thrust. the T41 synthesis is significantly improved while the thrust

synthesis is marginally improved. The parameter deltas are "small" except

for fan efficiency. Thus around the flight envelope on a "nominal" engine,

the expected variations are of the order 4 points on fan efficiency, 2

points on turbine efficiency, 2 percent on airflow, and 1 percent bias on

fuel flow. These are the magnitudes of the modelling errors, and impose

the smallest four-dimensional box for identifying sensor errors. By this,

it is meant that if a sensor were drifting, it would have to cause an

efficiency change of greater than four percent before it could be declared

a fault. The section on sensor errors further amplifies the point.

The next set of results covers the effects of engine to engine

variations as discussed in Section 3.1 and shown in Appendix 1. The

expectation here is that the update method can distinguish component

quality through the adaptation of component parameters. If, for example,

high pressure turbine efficiency decreased by two points, it is desired

that the model would move this parameter (and this parameter only) by two

points on a relative basis. Thus, the time history of the update set (or

vector) would be indicative of engine component health.

The comparison is made by adjusting the truth model for the engine

variations, and then matching the resulting N, N 2 , PS3C and T5 at the same

fuel flow and exhaust nozzle area. The base point is at seal level static,

standard day intermediate rated power setting. Appendix 1 contains the

detailed information, which is summarized in Table 4.1-4. The Table

presents the base case plus the individual effects as deltas from the base

case. In the first variation, by changing fan efficiency of the truth

75
model by +0.6%, the vector changes model fan efficiency by 0.58%, turbine

efficiency by +0.04 points, fan airflow by -0.19 percent and fuel flow bias

by -0.09 percent. The method, therefore identifies the change in fan

efficiency almost exactly. This result reveals the health monitoring

capabilities of the update.

TABLE 4.1-4 EFFECTS OF ENGINE TO ENGINE VARIATION

Change to Vector Updates Relative to Base

"Truth" Model f-Pts t-PTS Airflow% W

%
BASE -1.55 0.81 2.03 -0.68

n f + 0.6 PTS +0.58 +0.04 -0.19 -0.09

Airflow +1.0% +0.04 -0.08 +1.10 -0.08
Ti + 0.7 PTS +0.17 +0.53 +0.04 -0.14
c
Comp. Airflow + 1.0% +0.89 -0.15 +0.17 -0.05

)t+ -0.8 PTS +0.27 +0.73 +0.04 -0.12

FF41 + 1.0% -0.73 -0.30 -0.84 -0.93

n2t + 1.0 PTS +0.46 +0.19 -0.39 -0.18

FF45 + 1.0% -0.33 +0.42 +0.49 +0.10
Cooling flow +0.4% -0.30 -0.30 -0.28 -0.27

These results are further analyzed item by item from the table. The

second item, fan airflow, is also correctly identified. The third item,

compressor efficiency, is not contained in the parameter vector. However

its result looks strongly like a turbine efficiency change. This is

because the changes in HPC and HPT efficiency are virtually indistinguish-

able without a measurement of compressor discharge temperature, T3. The

fourth item, compressor airflow is identified as a change in fan efficiency.

This is an unfortunate result. Perhaps the direction on fan airflow will

76
distinguish the two in the general case. Next, turbine efficiency is

correctly identified. Turbine flow function changes result in a negative

change in all parameters, which is good if it is unique to the other

possible changes. The remaining variations do not show a distinguishing

single characteristic, but perhaps the direction of the vector is unique.

In conclusion, for the three component characteristics in the update

vector, the adaptation responds in a unique direction, and in the correct

direction. The other directions of the vector could be explored further.

Finally, the effects of sensor errors are explored. Table A-14 in the

appendix displays the detailed results for IRP and maximum afterburner.

The results for the two power settings are the same, and shown in Table

4.1-5. The sensor errors are shown on the left-hand side, and the result-

ing parameter variations follow on the right.

TABLE 4.1-5 EFFECTS OF SENSOR ERRORS

Sensor Error An f PTS Ant PTS A Airflow, % AWf,

%
BASE -1.55 0.81 2.03 -0.68

Fuel Flow + 5.0% +0.01 -0.12 -0.01 -4.74

Nozzle Area + 4.0% -0.43 -0.43 +2.15 +0.67
N + 0.5% +0.60 +0.03 -1.24 -0.05
N2 + 0.5% -0.80 +0.54 -0.19 -0.15
PS3C + 1.1 PSI +0.21 0.00 +0.29 +0.32
T5 + 160 R -0.65 -0.51 -0.22 +1.07

In the first case, the control sensed fuel flow is 5 percent greater

than that actually entering the burner. This could be because of a leak,

for example, or the metering value position feedback could be in error.

77
The update correctly removes the bias on fuel flow without significantly

effecting the other three variables. The nozzle area sensor error effects

all four variables, but fan airflow most strongly. A fan speed error shows

up as efficiency and airflow. This result is logical. For detection, the

normal variations in airflow dictate the magnitude. Previously, the

modelling errors at high power are at most 2 percent on airflow. The

magnitude of airflow error bound is calculated by adding the normal model

errors (2%) plus the RSS effect of engine to engine variations (1.6%) and

the sensor error (2.5%). The last two are RSS'd together because of the

random nature. Thus the box in the airflow direction is equal to 5.0%

airflow variation. (Take the square root of (1.6)2 + (2.5)2 and add

2.0%.) Therefore, an N error of 4% (5.0/1.24) is detectable in a simple-

minded sense. The above calculation is an estimate of the detection

capability. There are many other considerations to detection, and the

analysis is beyond the scope of this work.

In summary of the static results, the magnitudes of parameter changes

have been established for three cases: modelling errors around the flight

envelope, engine to engine variations and sensor errors. None are exces-

sively large, permitting possible detection of small in range sensor

failures. The engine variation analysis establishes that the adaptation is

unique for the three parameters in the update vector, and may be unique for

those not included. The sensor error analysis shows positive detection of

fuel flow sensor error, and possible failure directions for other failures.

This could be done in a further study.

78
4.2 Dynamic Performance of the Four Strategies

The goal of this analysis is to check the tracking ability of the

update methods analyzed in Section 3. The engine transient is an accel-

eration from flight idle to IRP in about 3 seconds at sea level static,

standard day. This is a rapid time response, considering that thrust

increases by 40 times the flight idle setting. The truth model is the

large scale dynamic simulation which provides the control demanded fuel

flow and exhaust nozzle area. These are input to the model, which esti-

mates the four sensed parameters. The errors between the estimates and the

truth model outputs are converted to parameter adjustment through a gain

matrix, and the parameters are therefore dynamically updated.

A baseline for comparison is established by performing the accel-

eratIon without updating the four parameters. Figures 4.2-1 through 4.2-4

show the results. The first figure shows the comparison of the two rotor

speeds verses time. The two prefixed by "Z" are the reference speeds from

the large scale simulation. The maximum fan speed error is about 6% of

point at t=1.3 seconds, and the maximum compressor speed is 2% of point at

t=2.0 seconds. Near the end of the transient, the core speed prediction is

drifting away. The second figure shows the comparison of the remaining two

sensed variables: PS3C and T558. The PS3C estimate is fair, but the

temperature estimate is poor. The error is 8% at t=.8 seconds, and about

5.5% at the maximum temperature. The third figures shows the synthesis of

T41 and thrust. The former is more important, and the result is an under-

prediction of nearly 1000 at the peak temperature. This is important, as

turbine life is highly influenced by T41. The final figure summarizes the

79
r\.)
CD
N
-
LL
co rnLf
U)
I-J
C0 iz Cl,
zx z x
o
K
0A Ll i + N)
000' 91 0O T0s 0 olp I OOO ooo l 00011 00001 0006 oo8 0004 o009
Wdb 1NX
000' 9 oo1 000 1" 1o 000i 000h 00061 0I' 0od 006 0008 ood 00
W.ddi 1NXZ
000J a T 00021 00091 00091 o o00i, 000cl 00021 0001'1 00001' 0006( 0008
W.d~l HNX
00 081 0 0 ,L 00091 00091 ooo~l OCSil 0002,1 00011 0000' 0006 0008
LWdd HNXZ
FIGURE 4.2-1
80
 4

TS

Z-Z "7 HIMflIl

ZTSH DEG R
500.0 700.0 900.0 1100.0 1300.0 1500.0 1700.0 1900.0
I - i---- 2100.0 2300.0 2500.0

500.0
T658 DEG R
1
700.0 900.0 1100.0 1300.0 1500.0 1700.0 1900.0 2100.0
1 1 2300.0 2500.0
ZPS3C PSIA
50 100 150 200 250 300 350 400 450 500
PS3C PSIA
0 50 100 150 200 250 300 350 400 450 500
f I I I I

;vz
rr 0
- c

CD
w

D
-8

--I
~.~~~
-....... ~ .......

m 4
rn
M'i

co

4
 ...... . . . . ..... . . ...-
0,
. .......
). ...... -....... ---.-.-----....-..--...-.--.-..-..-.-.-..--.-.-.-...-.....-...- ~..... .............. L
I
L
U, L
(f)
Lfl
w
<r
LL CD cr)
0 z
cu
CL
z w~
o'oo rE 00EE 0'0018 00062 i'0E
0'0042 -' 0
0'0042 I'01 0'00E2 -'-4 -'09
0'0012 ,'0E
0'0061 -' -0
0oo021 '0 6
0o00st 1 0 I
8 030 1 ti
0'00
000 02 00081 00091 000*1 000i2 00001 0008 0009 0 001 ooo02 0
000 02 00081 00091 000*1 00021 00001 0008 0009 000' 0002 0
81l NZ
FIGURE 4.2-3
82
 O'
CC);
C
LL 0 m
0 r
w icu
D
01 8 9 $12 0 2- 9- 8- 01-
l~d HNXO
08 9 20 2- 9- 8- at I-
.L~d ESdO
01 8 9 2 0 2- -8- 01-
138 HSJ.O
FIGURE 4.2-4
83
percent errors for the four variables during the transient. These figures

establish the baseline comparison with which to judge the improvements.

The next set of figures show the results for the case where the gain

matrix multiplying the error vector is the inverse of the Jacobian.

Figures 4.2-5 through 4.2-8 relay information similar to the first set.

Figures 4.2-5 displays the comparison of the two rotor speeds. The differ-

ences between the truth model and the real time model are reduced by

correcting the modelling errors. Now the fan speed error is now 3.8% at

t=1.0 seconds. However, the tracking could be better. The next figure

shows the temperature and pressure sensor comparisons. The estimate of

PS3C is improved, and the T558 estimate is significantly improved--

especially at the peak over temperature. The T41 synthesis is now off by

only 200, as shown in Figure 4.2-7. The error summary figure (4.2-8) shows

significant improvement during the transient, plus the errors are forced to

be small at steady state. Finally figure 4.2-9 shows the parameter changes

during the accel. Note that at t=0, the model is not initialized to the

truth model. The modelling errors are significant in the first 3 seconds,

resulting in nearly 8 points in the two efficiencies. This result could

limit the ability to detect transient faults.

The results for the least squares derived gain matrix are shown in

Figures 4.2-10 through 4.2-14. Two results are worth noting with respect

to the previous method. First, in Figure 4.2-13, the error in PS3C is not

driven to zero. This unfortunate fact is due to the influence of the

penalty on individual parameter size. The effect is to cause the gain

matrix to change so that it cannot correct the error. (That is, the net

result of the gains times the error vector has no influence on PS3C as for

84
 F-
I II
a,
----4
LUj
.4
0
CD
- .....---....... ........... . ....
........
C
rLfl LU
z LuJ
z
0
-E LL
C
zzzz
x x x x
S+
LUJ
o I
L) LUI
cO 0
000 1 000SI 000*1 00081 00021 00011 0000T 0006 0008 000 0009
Wdd 1lNX
000 91 00051 000*1 OQOSI 000 1 00011 00001 0006 0008 oOOO 0009
Wd8 1lNXZ
000 81 OOOd1 00091 oDooS 000 1 co0i 00021 00011 00001 0006 0008
W.dN HNX
000 31 ooo41 00091 00051 000*1 00081 000a1 00011 00001 0006 0008
Wd8 HNXZ
FIGURE 4.2-5
85
9-Z' 1l~

ZTSH DEC R
5 100. 0 ?700.0 9 100 .0 1 1100.0 1300.0 1500.0 1700.0 1900.0 2100.0 2300.0 2~ ;o00.0
T558 DEC R
600.0 700.0 900.0 11100.0 1300.0 1600.0 1700.0 1900.0 2100.0 2300.0 2'600 .0
1111 -II - - - I I - - I - I
ZPS3C PSIA
0 so 100 ISO 200 260 300 350 400 450 SC 00
II - - IItA -I I -II
PS3C PSIA
05 10 100 IS0 200 260 300 360 400 450 SC 00

M UI)
m C)

+~ >

m z

) (4)
Cl(r
(A)mm

cz

I--u
 0'1
C
0
L.
u - U-
I-
-..-.--.-.- .- --. .... --.... .. - -. .+
al
0r
u LL
c"
cu
OOSE O O 00T 00062 0-0012 010OS2 00E2 a0T nlf-I n sj nnn CD
.0
8030 ItiJ
0*009 E 000 E 0*0018 0*0062 o Ooi2 ooos92 0-OOiL 0'001'2 0-0061 0-0,4 O-i S
8 030 t*iZ
000da 00081 00091 oooi" o002T 00oodt 0008 0009 ood1F 0062- 0
S8-1 N-4
000 02 00081 0009T 0 00*' 0002,1 00001 0008 0009 0 o0kt 0002 0
S 81 N. Z
FIGURE 4.2-7
----- 87 -
 SCALECU INVER 'E JACIJBIAN GOAIN UFOArTE OF iPA~RrAN ETER

.
KRL HL TIInEI IJEL

CU- cu CU (U

F- I--

Lfl WI Z Z

0 XNH!
A PS3J~
+ OT5H 1

33 C7 .1t 2-T 43
TIN SEC
00
 686

6-Z68

ADD [..FM PCT

-20 -16 -12 -8 -4 0 4 8 12 16 20
ADD 1..JR PCT
-20 -16 -12 -8 -4 0 4 8 12 16 20
1 - IL II - - - I I I I

PETAT PTS
- 120 - 116 - 112 -18 -14 0 4 8 12 16 20
- I - I -L - - - I -

AETAF PTS

r-n
D
rn z
i

-n~~ 71m
-~~~~I n ..
(D.-.D.
D
r-4

......... - .-.....-- *-......

-.-. .-.-. . . . - ..-. -. . - --.. - -

. . . ......
non-controllability). Secondly, in Figure 4.2-14, the fan airflow is

changed by almost 11%, which is undesirable. The T41 synthesis, shown in

Figure 4.2-12 is about the same as the previous method.

The constant gain Kalman filter shows the typical good tracking

ability of the algorithm in Figures 4.2-15 through 4.2-19. The first

figure shows a significant result: the rapid initialization of the filter,

plus the small tracking deviations. The improvement in PS3C and T558

estimation is shown in the next figure. In the final figure of the set

(Figure 4.2-19), the parameter changes are detailed. The stepping effect

on the turbine efficiency is the result of the decision not to update if

the errors are less than 0.3%.

The final set of figures, 4.2-20 through 4.2-24, show this transient

for the extended Kalman filter update of the four parameters. The results

are similar to the constant gain version. However, the initialization is

slower due to the initial condition on the covariance matrix P. See

Figures 4.2-20 and 4.2-21.

Finally Figure 4.2-25 displays a composite of the updates for estima-

tion errors. These lines are the boundaries of the estimates of the four

sensed variables. The constant gain extended Kalman filter shows the

smallest boundary, with the errors less than 1% at t=3 seconds. These

twenty-five figures are part of the basis for the final conclusions and

recommendations. The choice here is to determine which dynamic method is

the best. From these results, the two Kalman filters appear to be the best

choice.

90
 CD
0'
-
N
LL
0
...... ...
.- . . .....
.....
z L
U-
Hr
C
C
.0J JCD
<r z z z
>) LL x x x x
o El_ _+
ooo9 I 000ST 0 0 0If 00 0C1 00021 00011 00001 0006 0008 oo0i 00C 9p
Wd6 INX
00V9 ! 0005! 0001 00OE 0002! 0001 00001 0006 000a ooL 0009
WdN 1lNXZ
000O I OOOd4 00091 00041 000i 000Si 00021 00011 00001 0006 0008
WdN HNX
000 31 0O004 00091 000ST 000;T 0o01 00021 000!! 00001 0006 0008
Wd HNXZ
FIGURE 4.2-10
91
 ....
..
----- --.-
--------
a,
LUJ
I-
<L
0
ZD
z_ LL
r n U-
'zr
L-
LUJ
Li
JT
Cl,
mWLD Lr
LU
CJn 0 Lo) F-
a- N N
DO
Ele4+
Cu
I-
cr C
LU LU
-J
00S 05* 00 ose 00C 052 002 0ST 1,
001 05 C0
tISd OESd
00s 05* 00* ose 00E 052 002 051 001 0s
UISd OESdZ
0.00s2 0'0062 0'0012 0006T o00L1 0'00S1 000E1 0'0011 0'006 0'00Z 0*009
N 030 8S3J.
0-0052 0'0082 0'0012 0-0061 0-00Z1 0-001 0-0081 0'0011 0'006 0'004 0.005
d 030 HSiZ
FIGURE 4.2-11
92
 Lu
Z:
CN
L-
LUJ
...... . . . . . . . ... .. .... . . .-- ...
z _
crL
Li
oi
LU
LUJ 0
c
D LUJ 11Z L
c3u
<r
CC
Gi ci:
0 00o ~E 0 '00ft 0*0 0 1E 0 *0062 o04 00S 000 o~a o- o2 0102 0*00 oi 0*OS
N 030 IfrI
0 *oo ~ES 010 0 E 0 10 0 1 0-0062 0-0022 0-0042 0 -0 OE2 0 -001'2 0.0o 1 0-0021 oloosl
6J 030 TirlZ
00081od 00091 0O00,1
000C 00021 000 1 0009 0009 oolp 00020
0001 S8-1 NJ
FIGURE 4.2-12
93
 C C C)

LEAST SQUARE$ DERIVEO GOIN UPDOTE OF OARAMET RS

a) w w w
REAL TIME MOPEL

'D %D D %0

tu (U ru tu

F- F--
- U _ u u .. ..
........... . ............
C, a- C3, Q- C3

(Y)
Lf) U) z
x
C3 CU C3 cu- 0 cu- 0 cu ( .....
.. ....
.. .....
..... ...........

........
. .... ........... ..
......
..
....
.. DX . . .....
...
.......
DXNH!
DPS3;
+ DT5Hl1
ID a -D D- ....
...............
............................
.. ..........
................ .. -................
.................
.... ..
..... .........
......

OD - OD OD OD - . ...................
........ .... ...............
.....

C C C:) CD

3 4 9 10
TIMF qFr
 .........---.---.------.-- --------------.----.----.----------.-.--.-...-----.--...- 0
0'
. . .... .- ...... .. ~~-. -~- - .-.-.--...--- - .- ........ .. - . . - . .- -..- - -.-.-
<-
C
LU
z
- LUI
-'4
L
LUJ
LUJ
LUJ
JLU o ---- ---E-------
(Y
Q LUJ Li LL LL
LI) -4
N N
cu
UJ* -1
C C
LUJ LUJ
02 91 1 8 0 8- 21- 91T- 02-
Sid d V3V
02 91 1 0 8- 21'9- 02-
Sid 10 13~
02 91 1 ~8 0 a- 21- 91,- 02-
i :)d 61M mOO
02 91a 0 8- 21 91- 02-
13d WAJM ocu
FIGURE 4.2/--14
95
 LUi
a-
0
LUJ
LL
L)
u.,
(n
F- -
ui-
-J LL
_ ~
-_-_-------
.......
..... +.....
z zcz
o d x xx
uuc.Zu~ uuIM ~~
U UUUU
toU UU
Ann
WdN INX
000c ~! 00091 000~1 00061 00021 000h1 00001 006 0d oL 00
000~ 1t 0002. 01 000 91 000 1 ooovl o te 0002,1 000il ooodil 000,6 OOo
LWdb HNX
000~ 3 1 0002.1 00091 oos o 0o 00; ooiT ooo~l ooo' o ooodi00' 0006 000
W.dN HNXZ
FIGURE 4.2-15
96
 Q0'
LUJ
LL... .......--.----
LUL
:D
LL
z
)< LU
z K m cr) U.) L.)
C~a LU. ON-
LO z
z C
0 LU
u Q
0o* 098 008: 09 2 002 051 001 09
oc 0c ~I ISd OCSd
)s st 001, 098 00C 052 L 002 051 001 05 0
0 bdISd 36SdZ
O2 002 0 *0012 0-0061 00dOUT 000OS1 0.0061 0-0011 01006 o- O 000S
0* 0o0 8 030 8S9.L
32 o0008O2 0-0012 0 00(61 o0*001 0.00sl o 001 0-0011 0oo6. a 00?- 0,00s
N 030 HSIZ
FIGURE 4.2-16
97
 -~ ~ ~~~~~~~~~.... ...- . . -- - . - . . - - - - - . . -- - _ _. . . . N
a-
. ...
.-. ..... . .....
CD
Li. fl
Z L
C- ~C
C: LU
.- Ca - .... . ..........
z z -~ Z
v F LLN
oD z
cu
z C
0 L
L) a
0* 00 6 0los 001s00062 *Z *01 00U62 0~U *UU 0UU2 00 0 0 006 1 0 0 02T 010051-
8I 030 Tt'1
0* 00S e 0 '0068 00 01 E 0-0062 0 -0 Z2 0092 0 -00E2 0 001'2 0 006'1 00 021 0 00os'
6 030 TiriZ
0000 2 oood1 0009T 000*1 00021 00001 0008 000 0 oi' 0002 0
S81l NJ
0000~2 00081 o 0091 oo O~t, O001 00001 0008 0009 o oO' 000,2 0
S8-1 NJZ
FIGURE 4.2-17
98
 C3 c ) 0 - ... ...
.. ..... ..........

lCONST.. GAIN EXT. K LMAN FJL. UPOHIL Uh PORAMETER

CD- CD- co CD-

'REAL TIME MODEL

.
............

.
D ND D -D

.......
...

............
cu cu (U cu

u u u
c C)
Q- a_ a-

CY)
Lf) U) z
a_ x
0 cu- 0 cu- 0 (U- 0 CU .. ...
.....
....

...
....
...........
........ DXNL,
DXNH!
DPS3:
H
!

...
..............
-a. ....... ....
,D a- 10 ........... ...........

F-4 i

a) a)...
.. ..
...
......
....
.....
.........

=1 c C3 C:)
co
4 5 7 8 9 10
TIME SEC
 -- ~~.
0
- .- .-.....................
--- -.--.-. - - - - -...... ........
.......
......--
-7 -- - ----
L
c
LL .L Lf
z
Z E . LiJ LL U..
oD z
W-
z <1
0 ULJ
u)
Uo '7 cA L 0 U 2191 0 2
SiLd - 013 0
f I I I
02 9 18 0 8- 2- 91- 02-
Sid i V 130
02 91 21 8 ir0 t,8- 2 - 9T 02-
i~d ( TM oak
02 91 2,1 8 0 8- a - 9 - 2
l~d IWJdM COO
FIGURE 4.2-19
10-0
 CD
C,
C
N
....... . ....... . ... ..
C
C-
Lu L
u')
C-
L
I
I-
D- C
<1
z z m
-J CJ x
LUJ -j
Lu
.-.a......
a( TLUC t~
000
tWdN INX
000'
I~ ooo2l ooo~1 00oir 0002,1 0 aI' ooo Gd( 000, ood ood6 0008
0001
WdN HNX
0001 IT O00/-l 0009! 00051 000s"1 0G0iT 000i! 0006! 00001 o0006 0 008
WdN HNXZ
FIGURE 4.2-20
101
 . .........
U-)
wL
I..... ~ ______ _________
f')
wj
'E
....--. . . . ---
L-)
-i
J),
C-n C,,
wi
LL
z ~
-.i 0 COCm mT
C r-m() I f
C L aN I--
z cu
UL -J
<r
X UL
UL t
ads
La
a 00S 09,' ad, O~SE O S 0ST D0T 09 do
VISd 2JSSd
ads ad, 6, O~s ode Oa~ 0o2 051 001 O0s 0
UISd OESdZ
0-00S2 000E2 00OT2 0-006t 0-00i1 000S1o~ 000 EI 0*001T 01006 000oz 0.00S
l 030 BS31
oaos 2 oao6 2 0-0012 0aoa i 0oaoi oaost1 ooo 0100TI 0.006 - *0 *0
8i 030 HSiZ
FIGURE 4.2-21
102
0:,
LI
C
LlLU
C
LiLn
Z) W
~~J 0
LlU
o1 z
__~ LU
o d N- U-
z
CU~
______ ________________________
_______I
LU Q
i IIi
0 00oi c oooe 0-0016 0 0062 000 OL2 000osa o000E2 o001t2 0-0061 o0021
I
o00E I
CD
6 030 Tti
0 M~ c 0 00ES 0-0016 0 0062 0 -00d2 0O-00S, 0006~2 001 T' 0-0061 0-0021 0-00S1
S030 It, IZ
0001 02 00081 00091 o0o0*',1 00021 00001 0008 0009 0 00*' 0002 0
S831 N-4
0001 0'. 00081 00091 o0o0*'"1 0002,1 00001 0008 0009 o00* 0002. 0
S8-1 NAZ
FIGURE 4.2-22
103
 40 go 00 w LA w

EXTEND!ED KA~LMAN FIL TER UPDAiTE C. PrARAHETERS

CD co
REALIIIE [MODEL

'.0

cu cu

a_2
u u

ia
0
H 01

LI
2
1'~

O DXNHI
* DPS31
+ DISHI
'.0~

CDj
l,)
I'.
wx

9 10
10 I 2 3 4
78 I io
2 4
TTMF QFC'
 CD
0'
Li-)
S--..--.---.~~------ ..----- - -.- - -- - - _- - -
tN-
-LL- .......... .... ... . . . . . . . .... . . .-- ----.---..-.-.--..---.-..-.. .---....-.
u
LUI
I~O
LU
C _dL
aL LU I-
c,)
LJ 0j
a LL)LL
-I
LLI aa
Si~d J 0130
02 91 1 8 0 28- 902-
Sid 10 13 IJ
I f
02 91 21 8 0 8- 21- 91- 02-
.L~d WTM CO
FIGURE 4.2-24
105
 40 S
U-1 a

SUMMARY OF UPDATE METHODS' ERRORS

No Update
- - - Inv. Jacobian
.- Least Squares

+10_ - -- Const. Gain. EKF.

Ext. Kalman F/L

+ 8-

+ 6

+4-

+2-
1
r00
4
2 .5-- ____ * -7. _
TIME-SECONDS.------

.
-

-4-

6-
'7e

- 8-

- -10-
 5. PROCESSOR REQUIREMENTS

This subject has been ignored through the preceding analysis. It is,

however, a very important consideration in the design of the microcomputer

based model. The update time of the numerical integrations is 10 millisec-

onds, and thus, all processing of the equation set must be completed in

this time frame. In light of this requirement, the various update schemes

are analyzed for processor time on a reference computer, and then the

results are extrapolated for the real-world issues of the engine mounted

controller.

5.1 Required Processor Time-Current Computer

The computer used to analyze processing time on a relative basis is

the Honeywell model H600 computer. This is a very large scale machine, and

as such, is not representative of current microcomputer technology. It has

been in use at the General Electric Company in Lynn for the past five years

to handle most computer processing tasks. Although not a micro, it serves

as a basis for comparison between the four update schemes.

The ten second transients discussed in Chapter 4 are the foundation

for analyzing processor time per time step in the real time model. The

processor clock was started after the model had initialized at time zero,

so as to give a good indication of transient processor time. The results

are shown in Table 5.1.

107
 TABLE 5.5-1 PROCESSOR TIME

Method P-Time per Update (MS)

Base 7.83
Inverse Jacobian 8.41
Least Squares 11.55
Const. Gain EKF 8.83
Ext. Kalman Filter 19.91

The base case is for the real time model running an update. It does

run in real time! The burden added by the update (approximately 25 math-

ematical operations) using the inverse Jacobian uses an additional 0.6

milliseconds (8%) per update. The recursive least squares derived gain

adds a significant computer burden due to the calculation of the gain

matrix each time step. Processor time increases by 3.7 MS or 48%. The

efficiency of the constant gain extended Kalman filter shows an increase of

1 MS per update or 13%. Finally, the extended Kalman filter more than

doubles the processor time (12.1 MS added or 250%), due to updating the

covariance equation.

These results, along with the numerical results from Chapter 4 indi-

cate that the constant gain extended Kalman filter is the most likely

choice for the update method. The added processor burden is small, and the

accuracy during transients is nearly the best.

5.2 Practical Issues of the Embedded Model

A re-look at Table 5.1-1 reveals that an advance in microprocessors

increasing speed by only 200% would allow the extended Kalman filter

108
algorithm to be programmed into the model. Unfortunately, the task is not

so simple. Current technology does allow this capability (or shortly will)

by the use of VHSIC technology and parallel processing. However, these

prototype processors operate under well-controlled environmental conditions

of temperature and vibration.

The demands placed upon an airframe and/or engine mounted processor

are very strong. The machine must withstand engine bay temperatures from

sub-freezing to tropical temperatures. The bay pressures can range from 1

PSI to over 15 PSI. The maneuver loads can be as high as nine G's. Plus

there is a broad spectrum of vibrational characteristics. When compared to

the serene environment of the ground based, air conditioned computer used

to generate the results, the complexity of the task becomes obvious.

Further, the reliability requirements also impose a technology chal-

lenge. If the Honeywell computer on the ground "crashes," there is some

inconvenience to the user. If the engine controller "crashes" the airplane

may also. Of course, there will be a back-up system, but the processor

circuitry must be highly reliable in order to limit the high cost of each

failure.

Thus, in conclusion of this part of the analysis, the burden imposed

by practical considerations of the physical environment and reliability

constraints make the extended Kalman filter still years down the road, and

therefore the constant gain version is the choice.

109
 6. SUMMARY AND CONCLUSIONS

6.1 Summary

The results of this work show that an embedded real time jet engine

model can be logically adapted to successfully track the engine outputs

within the constraints of real time operation. Since the particular model

is predominantly algebraic equations, it does not lend itself to traditional

state space linear methods. The procedure used was to select component

characteristics, and adjust them dynamically to account for modelling

errors. This set a variables was found to consist of an adder to the

implied fan efficiency, an adder to the implied fan speed-flow relation-

ship, an adder to the high pressure turbine efficiency, and a bias the

control fuel flow measurement. Other component efficiencies and flows were

considered, but rejected based on the changes in these parameters to

correct model errors.

Several methods were tried to perform the dynamic update of the

variables. The selected method was to augment the two dynamic states (the

rotor speeds) by creating an augmented state vector that included the

parameters and forming the constant gain extended Kalman filter. It is so

called because the evaluation set is non-linear, but the filter gains are

precomputer.

The adaptation algorithm can be used to monitor the condition of the

engine as run time accumulates. This is accomplished by storing a refer-

ence variable vector at a reference power condition, and studying the time

history of the vector. This is made possible by the fact that the update

110
responds in a unique (and correct) direction for changes in the engine of

those parameters in the model update vector.

The update provides a significant improvement in synthesizing important

unmeasurable engine parameters. Of particular interest is turbine inlet

temperature, T41. The potential for synthetic control is present in this

case, where the engine is controlled to the output on the model rather than

a sensed parameter.

6.2 Conclusions

Specific conclusions are listed below.

1. Modelling errors that result in output differences between the

engine and the model are best corrected by updating component

representations in the model in real time for models that are

non-state space representations.

2. A vector of parameters consisting of fan efficiency, high pressure

turbine efficiency, fan speed flow relationship and a fuel flow

bias is the best set to change in order to eliminate modelling

errors.

3. A constant gain extended Kalman filter, which includes the two

model dynamic states, and uses the vector as augmented states is

the best way to dynamically update the model in consideration of

real time constraints.

4. The resulting algorithm can successfully track an analytical model

of the dynamic engine system, and provide insight into engine

component health.

111
6.3 Recommendations for Further Study

Several questions were suggested during the results section. These

concerns are listed here. The first area of study is the uniqueness of the

vector direction for certain engine variations and sensor errors. It has

been established that for a fuel flow sensor error, the direction is

unique--it changes the fuel flow bias only. It has also been established

that the vector responds in the correct direction for fan and turbine

efficiency and fan air flow. The question is whether the direction for the

N sensor failure or error is unique. This information would further

enhance failure detection and health monitoring.

The above analysis further suggests the failure detection logic should

be studied. A detailed study of all failures could be undertaken to

establish threshold levels. Then, in the event of failure detection, is

the filter workable without the feedback of the failed sensor. A simple

way to address this would be to look at the closed loop filter with one

column set to zero. This would result in the particular error of the

failed sensor having no effect on the update.

Another potential area of study is to analyze the update using engine

test data. This would complicate the update by introducing noisy sensors

and real uncertainty on measured parameters. In addition, data from an

endurance test would highlight the ability to monitor engine health.

Finally, an investigation into the synthetic control capability is

warranted. It is guessed that the update introduces phase shift into the

control loop due to the integration. If the update is occurring near the

control loop frequency, poor control may result. This may force a faster

or slower update relative to the control frequency response.

112
 REFERENCES

1. Linebrink, K.L., and Vizzini, R.W., "Full Authority Digital

Electronic Control (FADEC) - Augmented Fighter Engine Demonstration,"
SEA Paper 821371 Oct. 1982.

2. "Digital Electronic Engine Control (DEEC) Flight Evaluation on an F15

Airplane," NASA CP-2298, 1984.

3. Reukauf, P.J., and Burcham, F.W., "Propulsion System/Flight Control

Integration for Supersonic Aircraft," NASA CP-001, 1976.

4. Merrill, W.C., "Sensor Failure Detection for Jet Engines Using

Analytical Redundancy," NASA TM 83695, AIAA-84-1452, June 1984.

5. Wallhagen, R.E., and Arpasi, D.J., "Self Teaching Digital Computer for
Fail-Operational Control of a Turbojet Engine in a Sea-Level Test
Stand," NASA TM X-3043, 1974.

6. Hrach, F.J., Arpasi, D.J., and Bruton, W.M.,"Design and Evaluation of

a Sensor Fail-Operational Control System for a Digitally Controlled
Turbojet Engine," NASA TM X-3260, 1975.

7. Sahgal, R.K., and Miller, R.J., "Failure Accommodation in Gas Turbine

Engines with Application Fan Turbine Inlet Temperature
Reconstruction," Proceddings of Joint Automatic Control Conference,
Dencer, Co., June 17-21, 1979, pp. 381-386.

8. Spang, H.A., and Corbey, R.C., "Failure Detection and Correction for
Turbofan Engines," Joing Automatic Control Conference, San Francisco,
Ca., June 22-24, 1977.

9. French, M.W., "Development of a Compact Real-Time Turbofan Engine

Dynamic Simulation," SAE paper 821401, Oct. 1982.

10. Pisano, A.D., "Failure Indication and Corrective Action for Turboshaft
Engines," AHS paper HPS-16, Nov. 1979.

11. Williams, L.W., "Instruction Manual for Transient Computer Program

82086D" General Electric Co. Memo SCO 82086D, 1983.

12. Sofanov, M.G., and Athans, M., "Robustness and Computational Aspects
of Non-Linear and Stochastice Estimators and Regulators," Proccedings
of IEEE Conference Decision and Control, Dec. 7-9, 1977.

13. Gelb et al., Applied Optimal Estimation, MIT Press, 1974.

113
14. Landau, Y.D., Adaptive Control, The Model Reference Approach, Marcal
Dekker, New York, 1979.

15. Stewart, G.W., Introduction to Matrix Computations, Academic Press,

New York, 1977.

16. Doyle, J.C., and Stein, G., "Multivariable Feedback Design Concepts
for a Classical/Modern Synthesis," IEEE trans. on automatic control,
Apr. 1981.

17. Messerole, J.S., "Detection Filuters for Fault Tolerant Control of

Turbofan Engines," Ph.D. Thesis, Massachusetts Institute of
Technology, 1981.

18. Sage, A.P., Optimum Systems Control, Prentice Hall, N.J., 1968.

19. Kwakernaak, H., and Sivan, R., Linear Optimal Control Systems, Wiley,
New York, 1972.

114
 APPENDIX 1

The following information shows the detailed results from Section 3.1.

The purpose is to determine how far various parameters must be changed in

order to eliminate modelling errors. The truth model in this case is the

full range complex reference engine simulation. For each of the sets in

Table 3.1-2, the model tracks the engine outputs of N1 , N 2 , PS3C, and T56

by varying the four (or three) selected parameters of the set. This is a

static test, so the method used to update dynamically is of no consequence.

The exploration of the flight envelope includes altitudes to 60,000 feet

and flight mach numbers to 2.0. It covers two power settings:

intermediate rated power (IRP) at a power lever angle (PLA) of 87* for

maximum thrust non-afterburning and maximum afterburner at a PLA of 130*.

The twelve tabulations show individual parameter changes at each of

the ten flight conditions for the two power settings. The headings are

translated as summarized below.

ALT Altitude in feet

XM Flight mach number

PLA Power lever angle - degrees

ZAETAF Adder to implied model fan efficiency - points

ZAETAT Adder to implied model high pressure turbine eff. - points

ZAFFM Adder to implied mode fan speed-flow - percent

ZAFF41 Adder to turbine nozzle flow function - percent

DWFMC Adder (bias) to main fuel flow - percent

DASC Adder (bias) to exhaust mozzle area - percent

115
 EFN Error in calculated thrust - percent

ET41 Error in calculated T41 -*R

EPX3C Error in calculated PS3C - PCT

ENH Error in calculated core speed - PCT

For example, with set 1 at sea-level-static, IRP power setting, the

modelling errors (relative to the truth model) are equivalent to -1.2

points in fan efficiency, +1.3 points in turbine efficiency, 2.6 percent in

fan speed-flow, and 0.7% in turbine nozzle area.

The magnitude of the improvement in predicting non-measured parameters

judges the "goodness" of the update vector set. The non-measured

parameters for the 4 element vectors are T41 and net thrust, FN. The

elimination of a variable allows a comparison of N 2 syhtnesis and PS3C

synthesis for the 3 element sets. The improvements are detailed in the

Table A-13. It shows the set numbers along with the average synthesis

errors (X) and the standard deviation (o) around the errors for the 20

flight condition/power settings. Set #0 is the base model without

parameter adjustments.

116
 w 1*V

37952LV/ FP/ SET-I 01/25/85 3:34 PM PAGE 1 SET-i PAGE

TABLE A-1
SET 1: WIR,ETA-F,ETA-TFF41

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.203 1.278 2.626 0.744 0. 0. 3.89 -8. -0.00 -0.00
0. 1.1 87. 2.663 -0.864 0.428 -2.068 0. 0. -2.76 19. 0.00 -0.00
10000. 1.4 87. 5.220 1.727 3.284 -0.428 0. 0. -4.89 -24. 0.00 0.00
15000. 0. 87. -2.360 -0.144 0.488 0.804 0. 0. 2.56 6. 0.00 0.00
35000. 0.8 87. -2.219 -0.344 -0.695 1.058 0. 0. 0.05 8. 0.00 -0.00
35000.. 1.2 87. -1.027 0.955 0.842 0.933 0. 0. 1.48 -9. -0.00 0.00
35000. 1.6 87. -0.064 1.388 2.409 -0.555 0. 0. 2.13 -4. -0.00 -0.00
35000. 2.0 87. 4.314 1.365 1.577 -1.095 0. 0. 14.89 -16. 0.00 -0.00
60000. 1.0 87. -4.098 -3.779 0.179 -0.137 0. 0. -0.65 23. -0.00 0.00
60000. 2.0 87. 4.945 1.820 1.382 -0.129 0. 0. 18.77 -26. -0.02 -0.00
0. 0. 130. -1.544 1.055 3.396 0.459 0. 0. 5.14 -1. -0.00 0.00
0. 1.1 130. 3.731 -0.250 3.477 -1.244 0. 0. -3.79 4. 0.00 0.00
10000. 1.4 130. 4.163 0.693 3.891 -1.196 0. 0. -4.08 -8. 0.00 0.00
15000. 0. 130. -2.087 -0.442 1.177 0.622 0. 0. 2.98 9. 0.00 0.00
35000. 0.8 130. -2.195 -0.761 0.129. 0.766 0. 0. -2.48 12. 0.00 0.00
35000. 1.2 130. -1.183 0.770 1.591 0.754 0. 0. -3.61 -3. 0.00 0.00
35000. 1.6 130. -0.408 0.896 2.720 -0.909 0. 0. -5.54 8. -0.00 -0.00
35000. 2.0 130. 3.369 0.474 2.735 -1.662 0. 0. -6.41 -4.. 0.00 -0.00
60000. 1.0 130. -3.904 -3.141 -0.910 0.440 0. 0. 0.86 14. 0.00 0.00
60000. 2.0 130. 4.042 0.812 2.002 -0.781 0. 0. -2.31 -13. -0.01 -0.00

-
-

37952LV/ FP/ SET-1 01/25/85 3:34 PM PAGE I OE- 1 PAG14 1

 w w w IV

37952LV/ FP/ SET-2 01/25/85 3:34 PM PAGE 1 SET-2 PAGE

TABLE A-2

SET 2: WIR.ETA-FETA-T,DELTA WF

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.547 0.814 2.032 0. -0.68 0. 3.03 -7. -0.00 -0.00
0. 1.1 87. 4.613 0.583 2.687 0. 1.57 0. -0.20 9. 0.00 0.00
10000. 1.4 87. 5.568 1.979 3.698 0. 0.39 0. -4.39 -24. 0.00 0.00
15000. 0. 87. -2.739 -0.639 -0.164 0. -0.73 0. 1.67 7. 0.00 0.00
35000. 0.8 87. -2.707 -1.000 -1.535 0. -0.94 0. -1.09 9. 0.00 -0.00
35000. 1.2 87. -1.424 0.370 0.130 0. -0.85 0. 0.35 -8. -0.00 -0.00
35000. 1.6 87. 0.201 1.712 2.840 0. 0.52 0. 2.97 -5. -0.00 -0.00
35000. 2.0 87. 5.328 2.030 2.659 0. 0.98 0. 17.40 -18. 0.00 0.00
60000. 1.0 87. -4.052 -3.705 0.303 0. 0.13 0. - -0.48 23. 0.01 0.01
60000. 2.0 87. 4.949 1.831 1.445 0. 0.32 0. 18.92 -24. -0.02 -0.01
0. 0. 130. -1.751 0.774 3.025 0. -0.43 0. 4.61 -1. -0.00 -0.00
0. 1.1 130. 4.844 0.523 4.805 0. 1.07 0. -2.45 1. 0.00 0.00
F- 10000. 1.4 130. 5.113 1.356 5.054 0. 1.15 0. -2.87 -8. 0.00 0.00
H 15000. 0. 130. -2.375 -0.820 0.677 0. -0.57 0. 2.31 9. 0.00 0.00
0 35000. 0.8 130. -2.567 -1.232 -0.500 0. -0.68 0. -3.23 13. 0.00 -0.00
35000. 1.2 130. -1.494 0.306 1.010 0. -0.69 0. -4.37 -2. 0.00 0.00
35000. 1.6 130. 0.024 1.426 3.434 0. 0.85 0. -4.52 7. -0.00 -0.00
35000. 2.0 130. 4.885 1.395 4.422 0. 1.61 0. -4.31 -5. 0.00 0.00
60000. 1.0 130. -4.061 -3.398 -1.228 0. -0.40 0. 0.43 14. 0.01 0.00
60000. 2.0 130. 4.710' 1.240 2.748 0. 0.77 0. -1.34 -14. -0.01 -0.01

F
37952LV/ FP/ SET-2 01/25/85 3:34 PM PAGE 1 SET-2 PAGE 1
 37952LV/ FP/ SET-3 01/25/85 3:34 PM PAGE 1 SET-3 PAGE

TABLE A-3
SET 3: ETA-F,ETA-T,FF41,DELTA A8

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -0.360 2.773 0. 2.008 0. -5.77 3.76 -32. 0.00 -0.00
0. 1.1 87. 2.659 -0.834 0. -2.079 0. -0.70 -2.84 19. -0.00 0.00
10000. 1.4 87. 6.737 3.291 0. 0.396 0. -8.27 -3.62 -41. 0.01 0.00
15000. 0. 87. -2.320 0.055 0. 0.985 0. -0.85 2.42 3. 0.02 -0.00
35000. 0.8 87. -2.284 -0.672 0. 0.752 0. 1.33 0.12 13. -0.01 0.00
35000. 1.2. 87. -0.850 1.400 0. 1.307 0. -1.79 1.76 -16. -0.00 -0.00
35000. 1.6 87. 2.244 4.318 0. 1.552 0. -9.10 8.37 -48. -0.00 -0.00
35000. 2.0 87. 4.983 1.974 0. -0.870 0. -3.92 18.38 -22. 0.01 -0.00
60000. 1.0 87. -4.069 -3.703 0. -0.068 0. -0.38 -0.61 21. 0.00 0.00
60000. 2.0 87. 5.082 2.324 0. 0.112 0. -3.05 21.28 -24. -0.13 0.05
0. 0. 130. -0.311, 2.977 0. 2.101 0. -9.00 5.05 -33. 0.00 0.00
0. 1.1 130. 4.070 0.316 0. -1.129 0. -7.29 -4.69 0. 0.00 0.00
10000. 1.4 130. 6.914 3.439 0. 0.506 0. -13.56 -3.25 -43. 0.01 0.00
15000. 0. 130. -1.969 0.052 0. 1.064 0. -2.49 2.55 1. 0.00 0.00
35000. 0.8 130. -2.226 -0.756 0. 0.766 0. -0.17 -2.57 12. -0.01 0.00
35000. 1.2 130. -0.791 1.596 0. 1.456 0. -4.05 -3.64 -17. 0.00 -0.00
35000. 1.6 130. 2.330 4.362 0. 1.572 0. -12.43 -1.45 -43. -0.00 0.00
35000. 2.0 130. 5.239 2.142 0. -0.769 0. -9.43 -2.91 -23. 0.00 -0.00
60000. 1.0 130. -4.040 -3.522 0. 0.104 0. 2.26 1.09 21. 0.00 -0.00
60000. 2.0 130. 5.246 2.175 0. 0.011 0. -6.87 0.42 -26. -0.07 0.03

37952LV/ FP/ SET-3 01/25/85 3:34 PM PAGE 1 SET-3 PAGE

 w w v IV wV

37952LV/ FP/ SET-4 01/25/85 3:34 PM PAGE 1 SET-4 PAGE 1

TABLE A-4

SET 4: ETA-FETA-T,DELTA WF,A8

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.352 1.180 0. 0. -1.31 -3.36 2.22 -21. -0.00 -0.00
0. 1.1 87. 4.680 0.920 0. 0. 1.49 -4.36 -0.69 7. 0.00 0.00
10000. 1.4 87. 6.346 2.952 0. 0. -0.30 -7.57.-4.10 -39. 0.00 0.00
15000. 0. 87. -2.729 -0.662 0. 0. -0.68 0.24 1.76 8. -0.00 0.00
35000. 0.8 87. -2.552 -1.200 0. 0. -0.56 2.19 -0.41 15. 0.00 -0.00
35000. 1.2 .87. -1.419 0.393 0. 0. -0.89 -0.22 0.33 -9. -0.00 0.00
35000. 1.6 87. 1.140 2.922 0. 0. -0.86 -6.68 5.04 -34. -0.00 .0.00
35000. 2.0- 87. 5.945 2.745 0. 0. 0.67 -5.63 21.95 -25. 0.00 0.00
60000. 1.0 87. -4.069 -3.656 0. 0. 0.03 -0.44 -0.55 20. 0.02 -0.01
60000. 2.0 87. 5.370 2.277 0. 0. 0.04 -3.27 21.47 -31. -0.01 -0.01
0. 0. 130. -1.358 1.311 0. 0. -1.36 -6.05 3.48 -21. 0.00 0.00
0. 1.1 130. 5.158 1.276 0. 0. 0.80 -9.48 -3.88 -6. 0.00 0.00
10000. 1.4 130. 6.406 2.998 o0 0. -0.37-12.56 -3.66 -40. 0.00 0.00
15000. 0. 130. -2.415 -0.724 0. 0. -0.74 -1.15 1.92 6. 0.00 0.00
35000. 0.8 130. -2.519 -1.295 0. 0. -0.57 0.76 -2.93 14. 0.01 -0.00
35000. 1.2 130. -1.428 0.482 0. 0. -0.99 -1.98 -4.72 -9. 0.00 0.00
35000. 1.6 130. 1.213 2.956 0. 0. -0.88 -9.56 -3.60 -30. -0.00 0.00
35000. 2.0 130. 6.098 2.821 0. 0. 0.61-11.19 -1.34 -26. 0.00 0.00
60000. 1.0 130. -4.076 -3.595 0. 0. -0.07 2.38 1.03 21. 0.01 0.00
60000. 2.0 130. 5.599 2.237 0. 0. -0.05 -7.33 0.73 -31. -0.01 -0.00

r
CET-4 PAGE
01/25/85
ix 37952LV/ FP/ SET-4 3:34 PM PAUt I
W 0 a a w V w w w

37952LV/ FP/ SET-5 01/25/85 3:34 PM PAGE 1 SET-5 PAGE I

TABLE A-5
SET 5: WiR,ETA-F,ETA-T

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.085 O.811 2.651 0. 0. 0. 3.94 -7. 0.68 -0.00
0. 1.1 87. 2.798 0.587 0.851 0. 0. 0. -2.25 9. -1.55 0.00
10000. 1.4 87. 5.162 1'.980 3.270 0. 0. 0. -4.91 -24. -0.39 0.00
15000. 0. 87. -2.250 -0.644 0.505 0. 0. 0. 2.59 7. 0.73 0.00
35000. 0.8 87. -2.078 -1.006 -0.678 0. 0. 0. 0.08 9. 0.96 -0.00
35000. 1.2 87. -0.884 0.364 0.865 0. 0. 0. 1.52 -8. 0.84 0.00
35000. 1.6 87. -0.134 1.716 2.390 0. 0. 0. 2.08 -5. -0.52 -0.00
35000. 2.0 87. 4.214 2.031 1.586 0. 0. 0. 14.93 -18. -0.97 -0.00
60000. 1.0 87. -4.126 -3.700 0.179 0. 0. 0. -0.65 23. -0.12 0.00
60000. 2.0 87. 4.865 1.881 1.327 0. 0. 0. 18.80 -26. -0.11 -0.00
0. 0. 130. -1.467 0.770 3.414 0. 0. 0. 5.17 -1. 0.42 0.00
0. 1.1 130. 3.646 0.530 3.556 0. 0. 0. -3.71 1. -1.05 0.00
10000. 1.4 130. 3.950 1.368 3.810 0. 0. 0. -4.17 -9. -1.13 0.00
15000. 0. 130. -1.998 -0.825 1.190 0. 0. 0. 3.01 10. 0.57 0.00
35000. 0.8 130. -2.088 -1.237 0.140 0. 0. 0. -2.45 13. 0.70 0.00
35000. 1.2 130. -1.062 0.299 1.611 0. 0. 0. -3.57 -2. 0.68 0.00
35000. 1.6 130. -0.520 1.435 2.700 0. 0. 0. -5.59 7. -0.84 -0.00
35000. 2.0 1.30. 3.081 1.415 2.629 0. 0. 0. -6.54 -5. -1.56 -0.00
60000. 1.0 130. -3.818 -3.403 -0.896 0. 0. 0. 0.90 14. 0.40 0.00
60000. 2.0 130. 3.900 1.261 1.954 0. 0. 0. -2.38 -14. -0.72 0.00

~IL 01/2S/8S 3:34 PM PAGE 1 ~jtI~~ VM~2~ 1

37952LV/ SET-S
FP/ SET-5
37952LV/ FP/ 01/25/85 3:34 PM PAGE 1 bET:I- Z kmt
PAG 1
I
 w

37952LV/ FP/ SET-6 01/25/85 3:34 PM PAGE 1 SET-6 PAGE

TABLE A-6

SET 6: WIR,ETA-T,DELTA WF

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 0.805 4.086 0. 1.54 0. 5.99 -7. 2.23 -0.00
0. 1.1 87. 0. 0.597 -2.001 0.. -2.56 0. -5.58 9. -4.07 0.00
10000. 1.4 87. 0. 2.044 -2.330 0. -5.58 0. -11.92 -26. -5.88 0.00
'15000. 0. 87. 0. -0.668 3.546 0. 3.21 0. 6.64 8. 3.92 0.00
35000. 0.8 87. 0. -1.042 2.133 0. 3.07 0. 3.81 10. 3.97 0.00
35000. 1.2 87. 0. 0.359 2.049 0. 1.33 0. 3.39 -8. 2.19 -0.00
35000. 1.6 87. 0. 1.715 2.569 0. 0.20 0. 2.45 -5. -0.31 -0.00
35000. 2.0 87. 0. 2.050 -2.538 0. -3.99 0. 5.44 -18. -4.90 -0.00
60000. 1.0 87. 0. -3.782 5.950 0. 6.39 0. 7.44 23. 6.22 0.00
60000. 2.0 87. 0. 1.810 -3.337 0. -4.16 0. 7.24 -22. -4.67 0.00
0. 0. 130. 0. 0.756 5.395 0. 2.11 0. 8.00 -0. 2.53 -0.00
0. 1.1 130. 0. 0.563 -0.292 0. -3.46 0. -7.80* 0. -4.47 0.00
10000. 1.4 130. 0. 1.426 -0.521 0. -4.25 0. -8.92 -10. -5.28 0.00
15000. 0. 130. 0. -0.848 3.881 0. 2.87 0. 6.58 10. 3.43 0.00
35000. 0.8 130. 0. -1.273 2.908 0. 2.94 0. 0.82 13. 3.59 0.00
35000. 1.2 130. 0. 0.290 3.051 0. 1.61 0. -1.64 -2. 2.32 -0.00
35000. 1.6 130. 0. 1.427 3.401 0. 0.81 0. -4.57 7. -0.03 -0.00
35000. 2.0 130. 0. 1.451 -0.472 0. -2.90 0. -10.45 -6. -4.38 -0.00
60000. 1.0 130. 0. -3.489 4.319 0. 5.85 0. 7.83 15. 6.26 0.00
60000. 2.0 130. 0. 1.275 -1.919 0. -3.49 0. -7.84 -14. -4.39 0.00

37952LV/ FP/ SET-6 01/25/85 3:34 PM PAGE 1 SET-6 PAGE

to q 90 0 L

-
37952LV/ FP/ SET-7 01/25/85 3:34 PM PAGE I SET-7 PAGE

TABLE A-7

SET 7: DELTA A8,ETA-F,ETA-T

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -0.139 1.434 0. 0. 0. -5.72 3.75 -28. 1.74 -0.00
0. 1.1 87. 2.863 0.705 0. 0. 0. -1.50 -2.35 9. -1.52 0.00
10000. 1.4 87. 6.710 3.010 0. 0. 0. -8.17 -3.68 -40. 0.32 0.00
15000. 0. 87. -2.189 -0.560 0. 0. 0. -0.87 2.45 4. 0.91 -0.00
35000. 0.8 87. -2.171 -1.133 0. 0. 0. 1.30 0.19 13. 0.72 0.00
35000. 1.2 87. -0.654 0.567 0. 0. 0. -1.84 1.80 -14. 1.18 0.00
35000. 1.6 87. 2.418 3.376 0. 0. 0. -9.19 8.54 -46. 1.45 -0.00
35000. 2.0 87. 4.980 2.557 0. 0. 0. -4.07 18.73 -24. -0.70 -0.00
60000. 1.0 87. -4.086 -3.664 0. 0. 0. -0.38 -0.63 21. -0.06 0.00
60000. 2.0 87. 5.793 2.405 0. 0. 0. -3.85 22.53 -32. 0.16 0.02
0.0. 130. -0.079 1.573 0. 0. 0. -8.91 5.04 -29. 1.82 0.00
0. 1.1 130. 4.177 1.148 0. 0. 0. -7.75 -4.52 -5. -0.83 0.00
10000. 1.4 130. 6.873 3.076 0. 0. 0. -13.41 -3.31 -41. 0.40 0.00.
15000. 0. 130. -1.831 -0.612 0. 0. 0. -2.51 2.58 2. 0.97 0.00
35000. 0.8 130. -2.114 -1.230 0. 0. 0. -0.20 -2.53 12. 0.72 0.00
35000. 1.2 130. -0.568 0.670 0. 0. 0. -4.10 -3.59 -15. 1.31 0.00
35000. 1.6 130. 2.519 3.416 0. 0. 0. -12.55 -1.32 -42. 1.47 0.00
35000. 2.0 130. 5.237 2.658 0. 0. 0. -9.58 -2.78 -25. -0.63 -0.00
60000. 1.0 130. -4.021 -3.584 0. 0. 0. 2.26 1.09 21. 0.09 0.00
60000. 2.0 130. .5.539 2.235 0. 0. 0. -7.32 0.71 -29. -0.02 0.02

r 1
sjPAGEV
3795LV/FPISET- U1LET-7
37952LV/ FP/ SET-7 01/25/115 3:34 FM PA~z I
 37952LV/ FP/ SET-8 01/25/85 3:34 PM PAGE 1 SET-8 PAGE

TABLE A-8

SET 8: ETA-T.DELTA A8,DELTA WF

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 1.454 0. 0. 0.18 -5.99 3.92 -29. 1.96 -0.00
0. 1.1 87. 0. 0.317 0. 0. -2. 8 3.44 -5.22 11. -3.97 0.00
10000. 1.4 87. 0. 1.227 0. 0. -3.85 4.16-11.30 -6. -5.09 0.00
15000. 0. ' 87. 0. -0.222 0. 0. 2.87 -5.36 4.79 -5. 4.46 0.00
35000. 0.8 87. 0. -0.749 0. 0. 2.83 -3.28 3.06 2. 4.36 -0.00
35000. 1.2 87. 0. 0.726 0. 0. 0.75 -3.20 3.08 -19. 2.22 -0.00
35000. 1.6 87. 0. 2.506 0. 0. -1.68 -4.28 2.02 -24. -1.30 -0.00
35000. 2.0 87. 0. 1.263 0. 0. -2.53 4.74 4.06 -4. -4.18 -0.00
60000. 1.0 87. 0. -2.917 0. 0. 5.31 -8.92 5.77 -5. 6.67 0.00
60000. 2.0 87. 0. 0.650 0. 0. -2.01 6.30 6.03 5. -3.54 0.01
0. 0. 130. 0. 1.584 0. 0. 0.10 -9.10 5.14 -30. 1.95 -0.00
0. 1.1 130. 0. 0.512 0. 0. -3.40 0.60 -7.71 1. -4.43 0.00
H 10000. 1.4 130. 0. 1.243 0. 0-. -3.85 1.07 -8.71 -5. -5.10 0.00
15000. 0. 130. 0. -0.333 0. 0. 2.37 -6.34 4.32 -6. 3.92 0.00
35000. 0.8 130. 0. -0.903 0. 0. 2.78 -4.73 -0.79 4. 4.23 0.01
35000. 1.2 130. 0. 0.806 0. 0. 0.64 -5.45 -2.84 -20. 2.21 0.00
35000. 1.6 130. 0. 2.511 0. 0.. -1.74 -6.51 -5.64 -19. -1.37 -0.00
35000. 2.0 130. 0. 1.276 0. 0. -2.55 0.95-10.50 -2. -4.25 -0.00
60000. 1.0 130. 0. -3.039 0. 0. 4.80 -7.00 5.49 -5. 6.21 0.00
60000. 2.0 130. 0. 0.496 0. 0. -1.95 3.80 -7.61 8. -3.63 0.01

01/25/85 3:34 PM PAGE SET-8 PAGE

379521V/ FP/ SET-8
x
(9 lS 40 90 4 w

37952LV/ FP/ SET-9 01/25/85 3:34 PM PAGE 1 SET-9 PAGE 1

TABLE A-9

SET 9: WIR,FF41,ETA-T

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 0.311 2.735 0.504 0. 0. 4.13 -10. 0.00 -0.59
0. 1.1 87. 0. 1.280 0.329 -1.211 0. 0. -2.45 15. -0.00 0.69
10000. 1.4 87. 0. 5.109 2.582 0.444 0. 0. -5.11 -15. 0.00 1.27
15000. 0. 87. 0. -1.930 0.779 0.345 0. 0. 3.10 1. 0.00 -1.19
35000. 0.8 87. 0. -1.930 -0.365 0.675 0. 0. 0.45 1. 0.02 -1.00
35000. 1.2 87. 0. 0.098 0.929 0.726 0. 0. 1.63 -i1. 0.00 -0.69
35000. 1.6 87. 0. 1.339 2.411 -0.567 0. 0. 2.14 -4. -0.00 -0.02
35000. 2.0 87. 0. 4.406 1.245 -0.041 0. 0. 15.15 -14. -0.00 1.03
60000. 1.0 87. 0. -6.674 0.586 -0.799 0. 0. 0.06 14. 0.01 -1.50
60000. 2.0 87. 0. 5.486 1.160 1.226 0. 0. 19.40 -29. -0.00 1.28
0. 0. 130. 0. -0.142 3.490 0.175 0. 0. 5.53 -4. 0.00 -0.62
0. 1.1 130. 0. 2.435 3.260 -0.302 0. 0. -3.99 4. 0.00 0.93
10000. 1.4 130. 0. 3.200 3.367 -0.578 0. 0. -4.70 -0. 0.00 0.99
15000. 0. 130. 0. -1.979 1.356 0.254 0. 0. 3.54 4. 0.00 -1.09
35000. 0.8 130. 0. -2.322 0.375 0.399 0. 0. -1.95 5. 0.01 -1.07
35000. 1.2 130. 0. -0.163 1.655 0.540 0. 0. -3.35 -5. 0.00 -0.79
35000. 1.6 130. 0. 0.591 2.734 -0.979 0. 0. -5.46 7. 0.00 -0.13
35000. 2.0 130. 0. 2.238 2.213 -1.220 0. 0. -7.05 2. -0.00 0.71
60000. 1.0 130. 0. -5.821 -0.601 -0.133 0. 0. 1.87 4. 0.01 -1.29
60000. 2.0 130. 0. 2.656 1.333 -0.410 0. 0. -2.97 -8. -0.03 0.89

r
-3 2ET- PAGEFP/
E
2 37952LV/ FP/ SET-9 01/25/85 3:3J4 PM PAt . I
0 0 0 v v w w

01/25/85 3:34 PM PAGE 1 SET-i0 PAGE 1

37952LV/ FP/ SET-10

TABLE A-10

SET 10: WIR,ETA-T,DELTA WF

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. -0.141 2.382 0. -0.42 0. 3.64 -10. 0.00 -0.67
0. 1.1 87. 0. 2.656 1.281 0. 0.71 0. -1.20 10. -0.00 0.92
10000. 1.4 87. 0. 4.673 2.264 0. -0.37 0. -5.55 -15. 0.00 1.20
15000. 0. 87. 0. -2.235 0.534 0. -0.29 0. 2.78 1. 0.00 -1.28
35000. 0.8 87. 0. -2.548 -0.849 0. -0.51 0. -0.13 2. 0.00 -1.21
35000. 1.2 87. 0. -0.545 0.438 0. -0.61 0. 0.85 -11. 0.00 -0.87
35000. 1.6 87. 0. 1.841 2.804 0. 0.49 0. 2.90 -4. -0.00 0.06
35000. 2.0 87. 0. 4.453 1.275 0. 0.02 0. 15.21 -14. 0.00 1.03
60000. 1.0 87. 0. -6.079 1.149 0. 0.71 0. 0.84 14. 0.00 -1.39
60000. 2.0 87. 0. 3.942 0.215 0. -0.37 0. 16.70 -19. -0.05 0.97
0. 0. 130. 0. -0.294 3.363 0. -0.15 0. 5.36 -4. 0.00 -0.65
0. 1.1 130. 0. 2.754 3.503 0. 0.21 0. -3.74 3. 0.00 0.99
10000. 1.4 130. 0. 3.745 3.803 0. 0.49 0. -4.24 0. 0.00 1.09
15000. 0. 130. 0. -2.202 1.172 0. -0.21 0. 3.32 4. 0.00 -1.16
35000. 0.8 130. 0. -2.682 0.072 0. -0.29 0. -2.27 t. 0.00 -1.21
35000. 1.2 130. 0. -0.633 1.277 0. -0.46 0. -3.82 -6. 0.00 -0.91
35000. 1.6 130. 0. 1.441 3.430 0. 0.85 0. -4.53 7. -0.00 0.01
35000. 2.0 130. 0. 3.379 3.135 0. 1.04 0. -5.86 3. -0.00 0.92
60000. 1.0 130. 0. -5.730 -0.509 0. 0.13 0. 1.98 4. 0.00 -1.27
60000. 2.0 130. 0. 3.026 1.564 0. 0.43 0. -2.98 -5. -0.03 0.90

37951V/FP/ ET- 01/5/8 3:3 PMPAGE 1 ~II~J1I42~

r
IS
37952LV/ FP/ SE T -10 01/25/85 3 :34 PM PAGE I E_ Tv- IV PAGE"r I
 0 . 4 4r

37952LV/ FP/ SET-il 01/25/85 3:34 PM PAGE 1 SET-11 PAGE I

TABLE A-l

SET 11: FF41,ETA-T,DELTA A8

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 2.429 0. 1.902 0. -5.79 3.80 -32. 0.00 -0.24
0. 1.1 87. 0. 1.321 0. -1.207 0. -0.55 -2.49 15. -0.00 0.70
10000. 1.4 87. 0. 9.597 0. 2.691 0. -8.25 -2.68 -43. -0.01 1.89
15000. 0. 87. 0. -1.603 0. 0.623 0. -1.41 2.83 -3. -0.00 -1.20
35000. 0.8 87. 0. -2.390 0. 0.361 0. 0.88 0.36 4. -0.01 -1.02
35000. 1.2 87. 0. 0.739 0. 1.171 0. -1.95 1.92 -18. 0.00 -0.56'
35000. 1.6 87. 0. 6.135 0. 1.964 0. -8.82 7.89 -42. -0.00 0.70
35000. 2.0 87. 0. 5.876 0. 0.619 0. -3.59 18.91 -23. -0.01 1.24
60000. 1.0 87. 0. -6.286 0. -0.513 0. -1.28 0.14 9. 0.02 -1.45
60000. 2.0 87. 0. 3.962 0. 0.050 0. -0.25 17.50 -7. -0.02 0.86
0. 0. 130. 0. 2.683 0. 2.011 0. -9.00 5.09 -33. 0.00 -0.21
0. 1.1 130. 0. 3.754 0. 0.227 0. -7.36 -4.66 -4. 0.00 1.09
10000. 1.4 130. 0. 9.790 0. 2.802 0. -13.82 -3.30 -44. -0.00 1.93
15000. 0. 130. 0. -1.357 0. 0.767 0. -2.84 2.99 -5. -0.00 -1.08
35000. 0.8 130. 0. -2.564 0. 0.281. 0. -0.41 -2.24 5. -0.04 -1.16
35000. 1.2 130. 0. 0.984 0. 1.324 0. -4.16 -3.47 -18. 0.00 -0.53
35000. 1.6 130. 0. 6.246 0. 2.003 0. -12.24 -2.01 -37. -0.00 0.72
35000. 2.0 130. 0. 6.090 0. 0.712 0. -9.14 -3.17 -23. -0.01 1.28
60000. 1.0 130. 0. -6.069 0. -0.323 0. 1.43 2.06 8. 0.02 -1.32
60000. 2.0 130. 0. 5.611 0. 1.119 0. -5.85 -0.59 -20. -0.03 1.19

r
rT-44 nAGE

&
37952LV/ FP/ SET-1l 01/25/85 3:34 PM PAGE I aE
 w

37952LV/ FP/ SET-12 01/25/85 3:34 PM PAGE 1 SET-12 PAGE

TABLE A-12

SET 12: ETA-T,DELTA WF,DELTA AB

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 0.428 0. 0. -1.19 -3.86 2.62 -25. 0.00 -0.61
0. 1.1 87. 0. 2.831 0. 0. 0.65 -2.10 -1.44 9. -0.00 0.93
10000. 1.4 87. 0. 5.825 0. 0. -1.34 -4.70 -5.40 -31. 0.00 1.34
15000. 0. 87. 0. -2.187 0. 0. -0.41 -0.78 2.50 -1. 0.00 -1.32
35000. 0.8 87. 0. -2.614 0. 0. -0.29 1.24 0.22 6. 0.00 -1.12
35000. 1.2 87. 0. -0.457 0. 0. -0.74 -0.71 0.78 -14. 0.00 -0.86
35000. 1.6 87. 0. 3.562 0. 0. -0.98 -6.28 4.46 -30. -0.00 0.31
35000. 2.0 87. 0. 4.985 0. 0. -0.28 -2.68 17.07 -19. -0.00 1.10
60000. 1.0 87. 0. -5.863 0. 0. 0.39 -1.86 0.53 7. 0.00 -1.38
60000. 2.0 87. 0. 4.045 0. 0: -0.44 -0.48 17.01 -21. -0.05 0.99
0. 0. 130. 0. 0.559 0. 0. -1.24 -6.54 3.94 -26. 0.00 .-0.57
0. 1.1 130. 0. 3.465 0. 0. -0.13 -7.02 -4.79 -3. 0.00 1.05
10000. 1.4 130. 0. 5.873 0. 0. -1.40 -9.59 -4.99 -32. 0.00 1.35
15000. 0. 130. 0. -2.080 0. 0. -0.50 -2.00 2.65 -2. 0.00 -1.23
35000. 0.8 130. 0. -2.677 0. 0. -0.31 -0.12 -2.31 5. 0.00 -1.21
35000. 1.2 130. 0. -0.359 0. 0. -0.85 -2.48 -4.28 -14. 0.00 -0.86
35000. 1.6 130. 0. 3.630 0. 0. -1.00 -9.12 -4.10 -26. -0.00 0.34
35000. 2.0 130. 0. 5.087 0. 0. -0.34 -7.90 -4.19 -20. -0.00 1.12
60000. 1.0 130. 0. -5.822 0. 0. 0.27 0.99 2.23 7. 0.00 -1.28
60000. 2.0 130. 0. 4.106 0. 0. -0.61 -4.03 -2.02 -22. -0.03 1.03

x 37952LV/ FP/ SET-12 01/25/85 3:34 PM PAGE 1 SET-12 PAGE

 TABLE A-13

Envelope Survey of Parameter Sets

Set # AFN, % AT41, *R EPS3C, % EN 2,

%
0 (base) x -0.8 +15 -1.0 0.2
6.1 23 2.6 0.7

1 x +0.8 1.9 0 0
a 6.4 13 0 0

2 x +1.1 -1.1 0 0
6.4 13 0 0

3 x 1.7 -11.3 0 0
0 7.3 21 0 0

4 x 1.7 -9.3 0 0
0 6.9 25 0 0

5 x 0.8 -1.4 -.15 0

a 6.4 13 .85 0

6 x 0.3 -1.2 -.10 0

7.0 13 4.23 0

7 x 2.1 -14 .48 0

a 7.2 22 .92 0

8 x -0.6 -7 .07 0
6.1 12 4.14 0

9 x 1.0 -2.3 0 -.10

a 6.7 11 0 1.0

10 x 1.0 -1.8 0 -.13

6.2 9 0 1.0

11 x 1.9 -15 0 .13

6.2 19 0 1.1

12 x 1.0 -13 0 -.06

a 6.3 15 0 1.1

129
 The large deviations on T41 synthesis eliminate sets 3 and 4 (which

include an A8 bias, in the parameter set). There is no improvement on the

prediction versus the base case set 0. The same applies to sets 7 and 11.

Further eliminations are accomplished according to Figure A-1. It

shows the sequential logic for judging the performance of the various sets.

The remaining sets were judged upon the magnitude of variations required to

track the engine. Variations on A8 bias eliminated sets 8 and 12,

variations in turbine nozzle area eliminated sets 1 and 9, and the fan
efficiency variations for set 11 deleted it from the 1st. At this point

four sets remain.

The results from the sensor error sensitivities are shown in Tables

A-14 to A-17. To evaluate these, the root sum square of the effects is

taken for each set. Root sum square (RSS) is used because the sensor

errors are assumed to be random and uncorrelated.

The root sum square is defined as

dx.
a. = >Z((d ) Ay
)

y1 j-C

130
 FLOWPATH FOR VARIABLE SET SELECTION

Sets 1-12

REJECTED
Accuracy of
T41 3 , 11
5,
Synthesis

,2,5,6,8-10,12

Parameter
8, 12
Variations
A8

,121,5,6,9,10,11

Parameter
1, 9 Variations Set 11
FF41
TIF

2,5,6,10 :
Sensor
Error Sets 5, 6
Sensitivities

Sets 2, 1D

Engine -Engine
Variation -,C Set 10
Sensitivities

FIGURE A-1
Set 2
131
9 9 9 9p a w

37952LV/ FP/ SET-2SEN 01/25/85 3:34 PM PAGE 1 SET-2SEN PAGE

TABLE A-14

SET 2:SENSOR ERRORS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.547 0.814 2.032 0. -0.68 0. 3.03 -7. -0.00 -0.00
0. 0. 87. -1.538 0.692 2.025 0. -5.42 0. 3.05 -8. 0.00 0.00
0. 0. 87. -1.977 0.388 4.185 0. -0.01 0. 3.74 8. 0.00 0.00
0. 0. 87. -0.945 0.847 0.790 0. -0.73 0. 3.02 -8. -0.00 -0.00
0. 0. 87. -2.351 1.355 1.849 0. -0.83 0. 2.71 -6. -0.00 -0.00
0. 0. 87. -1.331 0.813 2.321 0. -0.36 0. 3.46 -7. -0.00 -0.00
0. 0. 87. -2.202 0.307 1.811 0. 0.39 0. 3.09 17. -0.00 -0.00
0. 0. 130. -1.751 0.774 3.025 0. -0.43 0. 4.61 -1. -0.00 -0.00
0. 0. 130. -1.739 0.611 3.023 0. -5.18 0. 4.62 -1. 0.00 0.00
0. 0. 130. -2.184 0.416 4.770 0. 0.12 0. 5.04 12. 0.00 0.00
0. 0. 130. -1.142 0.806 1.778 0. -0.47 0. 4.59 -1. -0.00 -0.00
0. 0. 130. -2.546 1.300 2.875 0. -0.56 0. 4.26 1. -0.00 -0.00
0. 0. 130. -1.536 0.772 3.319 0. -0.10 0. 5.04 -1. -0.00 -0.00
0. 0. 130. -2.421 0.251 2.912 0. 0.67 0. 4.53 24. -0.00 -0.00

CrTy-2ENi DAGE 1
37952LV/ FP/ SET-2SEN 01/25/85 3:3 PqM PAk ~I
 9 S S S 5 9 9 9 S S

37952LV/ FP/ SET-5SEN 01/25/85 3:34 PM PAGE 1. SET-5SEN PAGE

TABLE A-15

SET 5: SENSOR ERRORS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.085 0.811 2.651 0. 0. 0. 3.94 -7. 0.68 -0.00
0. 0. 87. 2.593 0.662 7.433 0. 0. 0. 10.58 -7. 5.68 0.00
0. 0. 87. -1.971 0.387 4.194 0. 0. 0. 3.75 8. 0.01 0.00
0. 0. 87. -0.-476 0.844 1.424 0. 0. 0. 3.99 -8. 0.73 -0.00
0. 0. 87. -1.801 1.351 2.598 0. 0. 0. 3.81 -6. 0.83 -0.00
0. 0. 87. -1.085 0.811 2.651 0. 0. 0. 3.94 -7. 0..36 -0.00
0. 0. 87. -2.453 0.310 1.464 0. 0. 0. 2.58 17. -0.38 -0.00
0. 0. 130. -1.467 0.770 3.414 0. 0. 0. 5.17 -1. 0.42 0.00
0. 0. 130. 2.142 0.562 8.217 0. 0. 0. 11.86 0. . 5.41 0.00
0. 0. 130. -2.267 0.416 4.654 0. 0. 0. 4.87 12. -0.12 0.00
0. 0. 130. -0.844 0.801 2.190 0. 0. 0. 5.22 -1. 0.47 -0.00
0. 0. 130. -2.179 1.295 3.384 0. 0. 0. 5.01 1. 0.56 -0.00
0. 0. 130. -1.467 0.770 3.414 0. 0. 0. 5.17 -1. 0.10 0.00
0. 0. 130. -2.856 0.257 2.308 0. 0. 0. 3.63 24. -0.66 -0.00

37952LV/ FP/ SET-5SEN 01/25/85 3:34 PM PAGE 1 SET-5SEN PAGE

'C
 0 V 9 S P

37952LV/ FP/ SET-6SEN 01/25/85 3:34 PM PAGE 1 SET-6SEN PAGE -1

TABLE A-16

SET 6:SENSOR ERRORS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. 0.805 4.086 0. 1.54 0. 5.99 -7. 2.23 -0.00
0. 0. 87. 0. 0.685 4.064 0. -3.32 0. 5.97 -8. 2.21 0.00
0. 0. 87. 0. 0.370 6.933 0. 2.87 0. 7.59 8. 2.86 0.00
0. 0. 87. 0. 0.841 2.061 0. 0.72 0. 4.95 -8. 1.45 -0.00
0. 0. 87. 0. 1.341 4.997 0. 2.57 0. 7.22 -5. 3.41 -0.00
0. 0. 87. 0. 0.805 4.086 0. 1.54 0. 5.99 -7. 1.90 -0.00
0. 0. 87. 0. 0.294 4.740 0. 3.57 0. 7'.28 17. 3.16 -0.00
0. 0. 130. 0. 0.756 5.395 0. 2.11 0. 8.00 -0. 2.53 -0.00
0. 0. 130. 0. 0.595 5.372 0. -2.80 0. 7.97 -1. 2.50 0.00
0. 0. 130. 0. 0.394 7.781 0. 3.29 0. 9.32 13. 3.14 -0.00
0. 0. 130. 0. 0.792 3.345 0. 1.29 0. 6.95 -1. 1.76 -0.00
0. 0. 130. 0. 1.273 6.349 0. 3.15 0. 9.22 1. 3.71 -0.00
0. 0. 130. 0. 0.756 5.395 0. 2.11 0. 8.00 -0. 2.20 -0.00
0. 0. 130. 0. 0.225 6.201 0. 4.21 0. 9.20 25. 3.49 -0.00

PAGE 1 SEI-~SEN F'A~.it 1

K
x FP/ SET-6SEN
37952LV/ FP/ SET-6SEN 01/25/85
01/25/85 3:34 PM
3:34 PM PAGE I SET-6SEN PAGE I
 Iop 0 SSS J

37952LV/ FP/ SET-10SN 01/25/85 3:34 PM PAGE 1 SET-10SN PAGE 1

TABLE A-17

SET 10:SENSOR ERRORS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. -0.141 2.382 0. -0.42 0. 3.64 -10. 0.00 -0.67
0. 0. 87. 0. -0.253 2.372 0. -5.18 0. 3.65 -11. 0.00 -0.66
0. 0. 87. 0. -0.846 4.633 0. 0.32 0. 4.51 4. 0.00 -0.73
0. 0. 87. 0. 0.246 0.988 0. -0.57 0. 3.40 -10. 0.00 -0.48
0. 0. 87. 0. -0.141 2.382 0. -0.42 0. 3.64 -10. 0.00 -1.16
0. 0. 87. 0. -0.007 2.625 0. -0.14 0. 3.98 -10. 0.00 -0.60
0. 0. 87. 0. -1.032 2.311 0. 0.74 0. 3.95 12. 0.00 -0.76
0. 0. 130. 0. -0.294 3.363 0. -0.15 0. 5.36 -4. 0.00 -0.65
0. 0. 130. 0. -0.445 3.358 0. -4.92 0. 5.36 -5. 0.00 -0.64
0. 0. 130. 0. -0.922 5.187 0. 0.47 0. 5.97 8. 0.00 -0.71
0. 0. 130. 0. 0.088 1.979 0. -0.29 0. 5.09 -4. 0.00 -0.47
0'. 0. 130. 0. -0.294 3.363 0. -0.15 0. 5.36 -4. 0.00 -1.14
0. 0. 130. 0. -0.164 3.619 0. 0.14 0. 5.69 -4. 0.00 -0.58
0. 0. 130. 0. -1.209 3.382 0. 1.05 0. 5.56 19. 0.00 -0.75

x 1I IU3PU
I -- JI~~I rM~2L
E~AI'~ I4
37952LV/ FP/ SET-1OSN 01/25/85 3:34 PM PAGE I ar. '-'IVZal rm-
These are shown in Table A-18 below.

TABLE A-18

RSS of Sensor Error Sensitivities

Set # AnF-PT Ant-PT AWIR % AWF % EFN % ET41 0 EN2 EPS3C

2 1.29 0.86 2.51 1.34 0.92 28 -- --

5 4.12 0.86 5.31 -- 6.78 28 -- 5.2

6 -- 0.87 3.65 2.77 2.61 28 -- 1.8

10 -- 1.20 2.64 1.43 1.01 26 .23 --

The parameter variations for sets 5 and 6 caused by sensor errors are

too large and therefore the sets are eliminated. Finally, the two

remaining sets are subjected to typical "known" engine to engine

variations. The parameter changes for these variations are detailed in

Tables A-19 and A-20, and summarized in Table A-21.

136
 9 W

37952LV/ FP/ SET-2E-E 01/25/85 3:34 PM PAGE 1 SET-2E-E PAGE

TABLE A-19

SET 2:ENG-ENG VARIATIONS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. -1.547 0.814 2.032 0.- -0.68 0. 3.03 -7. -0.00 -0.00
0. 0. 87. -0.971 0.859 1.840 0. -0.77 0. 2.88 -8. -0.00 -0.00
0. 0. 87. -1.503 0.738 3.139 0. -0.60 0. 3.06 -6. -0.00 0.00
0. 0. 87. -1.381 1.347 1.989 0. P-0.82 0. 2.94 -9. -0.00 -0.00
0. 0. 87. -0.662 0.655 2.206 0. -0.63 0. 3.29 -10. -0.00 0.00
0. 0. 87. -1.278 1.546 2.074 0. -0.80 0. 3.03 -14. -0.00 -0.00
0. 0. 87. -2.281 0.509 1.190 0. -1.61 0. 1.72 -2. -0.00 -0.00
0. 0. 87. -1.093 1.008 1.645 0. -0.86 0. 2.65 -13. -0.00 -0.00
0. 0. 87. -1.873 1.230 2.519 0. -0.58 0. 3.48 -7. -0.00 -0.00
0. 0. 87. -1.845 0.521 1.750 0. -0.95 0. 2.61 -9. -0.00 -0.00

37952LV/ FP/ SET-2E-E 01/25/85 3:34 PM PAGE I ETl-2E-E PAGE- I

 37952LV/ FP/ SET-10EE 01/25/85 3:34 PM PAGE 1 SET-10EE PAGE

TABLE A-20

SET 10:ENG-ENG VARIATIONS

ALT XM PLA ZAETAF ZAETAT ZAFFM ZAFF41 DWFMC DA8C EFN ET41 EPS3C ENH
0. 0. 87. 0. -0.141 2.382 0. -0.42 0. 3.64 -10. 0.00 -0.67
0. 0. 87. 0. 0.268 2.066 0. -0.61 0. 3.26 -10. 0.00 -0.47
0. 0. 87. 0. -0.191 3.486 0. -0.35 0. 3.64 -9. 0.00 -0.67
0. 0. 87. 0. 0.482 2.314 0. -0.59 0. 3.48 -11. 0.00 -0.81
0. 0. 87. 0. 0.255 2.360 0. -0.52 0. 3.55 -11. 0.00 -0.30
0. 0. 87. 0. 0.736 2.378 0. -0.59 0. 3.53 -16. 0.00 -0.75
0. 0. 87. 0. -0.904 1.701 0. -1.23 0. 2.62 -7. 0.00 -0.76
0. 0. 87. 0. 0.339 1.900 0. -0.68 0. 3.08 -16. 0.00 -0.56
0. 0. 87. 0. 0.046 2.948 0. -0.26 0. 4.21 -11. 0.00 -0.97
0. 0. 87. 0. -0.620 2.167 0. -0.64 0. 3.34 -13. 0.00 -0.64

37952LV/ FP/ SET-1OEE 01/25/85 3:34 PM PAGE 1 SET-10EE PAGE

 TABLE A-21
Effects of Engine to Engine Variations

Set 2: AWIR, AnfP Ant, AWF

I Ant AWIR AWF EFN ET41

BASE -1.55 0.81 2.03 -0.68 3.03 -7
n 0.6 pts -0.03 0.04 -0.19 -0.09 -0.15 -1
WIR 1.0% 0.04 -0.08 +0.11 +0.08 +0.03 +1
c 0.7 pts 0.17 0.53 -0.04 -0.14 -0.09 -2
W25R 1.0% 0.89 -0.16 0.17 0.05 +0.23 -3
t 0.8% 0.27 -0.07 0.04 -0.12 0.00 -7
FF41 1.0% -0.73 -0.31 -0.84 -0.93 -1.31 +5
n2 t 1.0 pts 0.45 0.19 -0.39 -0.18 -0.38 -6
FF45 1.0% -0.33 0.42 0.49 +0.10 0.45 0
Cooling .4% -0.30 -0.29 -0.28 -0.27 -0.42 -2
RSS 1.35 0.84 1.12 1.02 1.52 11

Set 10: AWIR, Ant, AWF, Matching PS3C

Ant AWIR AWF EFN ET41 EN2

BASE -0.14 2.38 -0.42 3.64 -10 -.67
nf 0.6 pts 0.41 -0.32 -0.19 -0.38 0 +0.20
WIR 1.0% 0.05 0.10 0.07 0 +1 0
nc 0.7 pts 0.62 -0.07 -0.17 -0.64 -1 -0.14
W25R 1.0% 0.40 -0.02 -0.10 -0.09 -1 +0.37
t 0.8% +0.08 0 -0.17 -0.11 -6 -0.03
FF41 1.0% -0.76 -0.68 -0.81 -0.02 +3 -0.09
n2t 1.0 pts 0.48 -0.48 -0.26 -0.56 -6 +0.11
FF45 1.0% 0.19 0.57 0.16 +0.57 -1 -0.30
Cooling .4% 0.48 -0.22 -0.22 -0.30 +3 +0.03
RSS 1.34 1.09 0.95 0.96 9.7 0.62

139
 In the last table, some explanation is necessary. The first part

depicts the results for set 2. The first row shows the base deltas on the

four parametes that correct the modelling errors. This applies to the
"nominal" engine. The next row shows the deltas if the engine is 0.6 pts.

better in fan efficiency than the nominal engine. The mondel responds by

changing its fan efficiency by .57 points and the other variables by a

small amount. In essence, the model identifes the difference in fan

efficiency and then corrects itself accordingly. The third row shows a

similar characteristic for fan speed flow changes, and the sixth for

turbine efficiency changes. The root sum square of the random effects are

summarized at the bottom.

The same information for set 10 is shown in the lower part. This set,

with only two performance variables can only identify the two component

changes. The thrust and T41 synthesis are about the same for the two

methods. In conclusion set 2 is selected for its better ability in tracking

performance changes due to the extra variable. Finally, the effects of

engine to engine variations and sensor errors are combined as a final test.

TABLE A-22

Final Summary

Af A t AWIR AWF EFN ET41 EN 2

Set 2:
Sensors 1.29 0.86 2.51 1.34 0.92 280 0
E-E var 1.35 0.84 1.12 1.02 1.52 110 0
RSS 1.87 1.20 2.75 1.68 1.78 30* 0

Set 10:
Sensors -- 1.20 2.64 1.43 1.01 260 0.23
E-E var -- 1.34 1.09 0.95 0.96 100 0.56
RSS -- 1.80 2.86 1.72 1.39 280 0.61

140