Linköping Studies in Science and Technology

Dissertations, No 1603

Model Based Diagnosis and Supervision

of
Industrial Gas Turbines

Emil Larsson

Department of Electrical Engineering

Linköping University, SE-581 83 Linköping, Sweden
Linköping 2014
Linköping Studies in Science and Technology
Dissertations, No 1603

Emil Larsson
[email protected]
www.vehicular.isy.liu.se
Division of Vehicular Systems
Department of Electrical Engineering
Linköping University
SE–581 83 Linköping, Sweden

Copyright © 2014 Emil Larsson, unless otherwise noted.

All rights reserved.

Larsson, Emil
Model Based Diagnosis and Supervision of Industrial Gas Turbines
ISBN 978-91-7519-312-0
ISSN 0345-7524

Typeset with LATEX 2ε

Printed by LiU-Tryck, Linköping, Sweden 2014
Abstract
Supervision of performance in gas turbine applications is important in order to achieve:
(i) reliable operations, (ii) low heat stress in components, (iii) low fuel consumption,
and (iv) efficient overhaul and maintenance. To obtain good diagnosis performance
it is important to have tests which are based on models with high accuracy. A main
contribution of the thesis is a systematic design procedure to construct a fault detection
and isolation (FDI) system which is based on complex nonlinear models. These models
are preliminary used for simulation and performance evaluations. Thus, is it possible to
use these models also in the FDI-system and which model parts are necessary to consider
in the test design? To fulfill the requirement of an automated design procedure, a
thermodynamic gas turbine package GTLib is developed. Using the GTLib framework, a
gas turbine diagnosis model is constructed where component deterioration is introduced.
In the design of the test quantities, equations from the developed diagnosis models are
carefully selected. These equations are then used to implement a Constant Gain Extended
Kalman filter (CGEKF) based test quantity. The number of equations and variables which
the test quantity is based on is significantly reduced compared to the original reference
model. The test quantity is used in the FDI-system to supervise the performance and the
turbine inlet temperature which is used in the controller. An evaluation is performed
using experimental data from a gas turbine site. The case study shows that the designed
FDI-system can be used when the decision about a compressor wash is taken. When
the FDI-system is augmented with more test quantities it is possible to diagnose sensor
and actuator faults at the same time the performance is supervised. Slow varying sensor
and actuator bias faults are difficult diagnose since they appear in a similar manner
as the performance deterioration, but the FDI-system has the ability to detect these
faults. Finally, the proposed model based design procedure can be considered when an
FDI-system of an industrial gas turbine is constructed.

iii
Populärvetenskaplig sammanfattning
Diagnostik och prestandaövervakning förekommer inom många industriella applika-
tioner. Detta område är viktigt att beakta för att: (i) upprätthålla hög tillförlitlighet,
(ii) undvika onödig belastning på komponenter, (iii) minimera energiförbrukningen,
och (iv) effektivt kunna planera underhåll. Eftersom prestandan i en applikation oftast
inte är direkt mätbar behövs metoder för att kunna skatta dessa prestandaparametrar
utifrån kända mätsignaler. Detta kan vara svårt eftersom: (i) sambandet mellan mätsig-
naler och prestandaparametrar kan vara komplicerat, (ii) mätsignaler innehåller brus,
och (iii) mätsignaler kan vara opålitliga och visa ett felaktigt värde. Dessa aspekter bör
beaktas när ett diagnos- och övervakningssystem utvecklas. Eftersom många system är
komplexa kan det vara nödvändigt att ha effektiva och automatiserade metoder för att
skatta prestanda och bestämma diagnoser.
Inom industrin finns det oftast bra modeller som används för att göra prestanda-
analyser och simuleringar över olika köruppdrag. För att göra diagnostik- och över-
vakningsanalyser används ofta andra typer av modeller som är enklare och inte lika
beräkningstunga. Dessa två områden är dock nära besläktade och ett gemensamt ram-
verk skulle kunna användas. Fördelarna med ett gemensamt ramverk är att: (i) endast
en modell behöver underhållas, (ii) fel i komponenter och prestandadegraderingar kan
bli enklare att modellera eftersom detta kan introduceras i modellen, (iii) diagnostest
med en given felkänslighet kan automatiskt genereras, och (iv) diagnostesten kan sedan
användas i det utvecklade diagnos- och övervakningssystemet för att bestämma möjliga
diagnoser utifrån vilka test som har reagerat.
I denna avhandling presenteras en automatiserad designmetodik för att på ett syste-
matiskt sätt konstruera ett diagnos- och övervakningssystem för en industriell gasturbin-
applikation där diagnosbesluten grundar sig på en fysikalisk modell. För industriella
gasturbiner är viktiga parametrar att övervaka: (i) verkningsgrader i komponenter,
(ii) massflöden genom komponenter, och (iii) temperaturer. En hög gasinloppstempera-
tur till turbinen kan medföra ökad belastning på materialet vilket leder till en förkortad
livslängd. Å andra sidan ger en hög temperatur en bättre verkningsgrad. Eftersom för-
bränningstemperaturen inte mäts med någon sensor är det viktigt att kunna skatta den
så noggrant som möjligt för att inte riskera att överskrida den tillåtna temperaturen.
För att bestämma om prestandan har försämrats eller förbättrats är en möjlig lösning
att bestämma avvikelsen för dessa parametrar från ett nominellt värde. Detta värde är
dock inte konstant utan varierar beroende på arbetspunkt och kan därför beskrivas
med en modell för det nominella fallet (vanligtvis den modell som används för prestan-
daanalyser). Avvikelsen för prestandaparametrarna från nominellt värde sägs, i någon
mening, representera gasturbinens hälsotillstånd och kallas därför hälsoparametrar.
Dessa hälsoparametrar skattas i de utvecklade diagnostesten.
Det finns ett antal olika typer av industriella gasturbiner med ett brett spektrum vad
gäller genererad effekt. De gasturbiner som har högst effekt används ofta för elproduktion
i ett kraftverk. Eftersom dessa används för att driva en generator är rotationshastigheten
konstant. Ett annat användningsområde är det s.k. mechanical drive vilket är vanligt
förekommande inom olje- och gasindustrin. För denna typ av applikation används den
genererade effekten för att driva en pump eller en extern kompressor för att exempelvis

v
vi Populärvetenskaplig sammanfattning

pumpa gas i en rörledning. Flödet i gasledningen kan variera med exempelvis tid på
dygnet vilket medför att det är en fördel om det går att variera effekt och rotationshastighet
på den påkopplade komponenten. Detta kan uppnås genom att använda en kraftturbin
som inte har någon mekanisk koppling med gasgeneratorn som driver kraftturbinen.
Eftersom både effekt och rotationshastighet kan varieras bör det ställas högre krav på de
modeller som används vilket även fångas upp av den föreslagna designmetodiken.
Avslutningsvis utvärderas diagnos- och övervakningssystemet genom applikations-
studier där data från en gasturbinsite studeras. En av studierna fokuserar på att skatta
prestandadegradering vilket kan hänföras till en nedsmutsad kompressor. Övervaknings-
systemet kan användas som underlag när beslut om att tvätta kompressorn fattas.
Acknowledgments
This work has been carried out at the Division of Vehicular Systems at the department
of Electrical Engineering, Linköping University. The research has been founded by the
Swedish Energy Agency, Siemens Industrial Turbomachinery AB, and GKN Aerospace
Sweden AB through the Swedish research program TURBOPOWER, who are gratefully
acknowledged.
First of all I would like to express my gratitude to my supervisors Jan Åslund, Erik
Frisk, and Lars Eriksson for all their support during these years as a Ph.D. Student at the
research group at the Vehicular Systems.
All colleagues at the Vehicular Systems are acknowledged for maintaining a pleas-
ant research atmosphere and interesting discussions during coffee breaks. Christofer
Sundström and Daniel Eriksson are both acknowledged for valuable research discussion,
especially topics regarding diagnosis and supervision.
Klas Jonshagen, chairman of the Processes and Diagnostics steering committee, is
thanked for valuable inputs regarding the work. Mats Sjödin at SIT is thanked for sharing
his expertise according to industrial gas turbine applications. Jesper Waldfelt, Lennart
Näs, Åsa Lovén, and Christer von Wowern from Siemens Industrial Turbomachinery
AB in Finspång are all acknowledged regarding issues according to measurement data
and developed models by the company.
Finally, I would like to express my gratitude to Marie for her support and encourage-
ment.

Emil Larsson
Linköping, May 2014

vii
 Contents

I Introduction 1
1 Background and Motivation 3
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Industrial Gas Turbine Background and Simulation Environment 7

2.1 Gas Turbine Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Brayton Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Mechanical Drive Application . . . . . . . . . . . . . . . . . . . 8
2.2 SGT-700 Gas Turbine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Measurement Signals . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Actuator Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Simulation Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Reference Gas Turbine Model . . . . . . . . . . . . . . . . . . . 14
2.3.2 Modelica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Summary of Contributions 17
3.1 Thesis Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

II Modeling and Design 23

4 Thermodynamic Concepts 25
4.1 Thermodynamic System . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1.1 Thermodynamic Quantities . . . . . . . . . . . . . . . . . . . . . 26
4.1.2 Thermodynamic Laws . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Thermodynamic Properties of Pure Substances . . . . . . . . . . . . . . 29
4.2.1 Specific Heat Capacity of Pure Substances . . . . . . . . . . . . 31
4.2.2 Standardized Enthalpy of Pure Substances . . . . . . . . . . . . 31
4.2.3 Standardized Entropy of Pure Substances . . . . . . . . . . . . . 32

ix
x Contents

4.2.4 Gibbs Free Energy . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.3 Combustion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3.1 Stoichiometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3.2 Chemical Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3.3 Comparison of the Heat Capacity Between the Stoichiometric
Gas Description and the Chemical Equilibrium Calculation . . 40
4.3.4 Mixing of Exhaust Gases with Different Lambda . . . . . . . . 40
4.4 Ideal Gas Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4.1 Thermodynamics Properties of Frozen Mixtures . . . . . . . . . 44
4.5 Energy Conservation of Thermodynamic Systems . . . . . . . . . . . . 44
4.5.1 Thermodynamic Differentials dU, dW, and dQ . . . . . . . . . 44
4.5.2 Energy of the Mixture of Frozen Ideal Gases . . . . . . . . . . . 45
4.6 Control Volume Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.6.1 Differential Form of the Ideal Gas Law . . . . . . . . . . . . . . 46
4.6.2 Lambda Concentration Differential dλ . . . . . . . . . . . . . . 48
4.6.3 Partial Derivatives of Gas Property Functions . . . . . . . . . . 48
4.6.4 State Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.6.5 Variation in Ambient Absolute Humidity . . . . . . . . . . . . . 49
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5 GTLib – Thermodynamic Gas Turbine Modeling Package 53

5.1 Gas Turbine Performance Characteristics . . . . . . . . . . . . . . . . . 54
5.1.1 Compressor Map . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1.2 Turbine Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2 Variation in Ambient Absolute Humidity . . . . . . . . . . . . . . . . . 57
5.3 GTLib Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.1 Global Environment Model . . . . . . . . . . . . . . . . . . . . . 58
5.3.2 Connectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3.3 Gas Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.4 Control Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.5 Compressor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.6 Turbine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3.7 Combustor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.8 Pressure Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 Gas Turbine Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Modeling, Analysis, and Transformation of the Diagnosis Model 71

6.1 Gas Turbine Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.1.1 Gas Path Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.1.2 Engine Health Monitoring . . . . . . . . . . . . . . . . . . . . . 75
6.2 Gas Turbine Diagnosis Model . . . . . . . . . . . . . . . . . . . . . . . . 76
6.2.1 Input Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.2.2 Output Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.2.3 Health Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Contents xi

6.2.4 Component Faults . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6.2.5 Faults in Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2.6 Differential Algebraic Equation Form . . . . . . . . . . . . . . . 80
6.3 DAE-Index Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.3.1 Differential Index Reduction . . . . . . . . . . . . . . . . . . . . 81
6.4 Structural Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.4.1 Dulmage-Mendelsohn Decomposition . . . . . . . . . . . . . . 83
6.4.2 Investigation of Actuator Fault and Health Parameter Isolation 85
6.4.3 DAE-index 1 Conservation in the Over-Determined M  Part . 87
6.4.4 Diagnosis Test Equation . . . . . . . . . . . . . . . . . . . . . . . 90
6.5 Observability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.5.1 Structural Observability . . . . . . . . . . . . . . . . . . . . . . . 91
6.5.2 Removing of Unobservable Modes . . . . . . . . . . . . . . . . . 92
6.5.3 Number of Health Parameters in the Model . . . . . . . . . . . 93
6.6 Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.6.1 Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.6.2 Stationary Kalman Filters . . . . . . . . . . . . . . . . . . . . . . 96
6.6.3 Nonlinear Kalman Filters . . . . . . . . . . . . . . . . . . . . . . 96
6.7 Parsers for an Automatic Extraction of Sub Systems . . . . . . . . . . . 97
6.7.1 Automatic Extraction of the DAE Model . . . . . . . . . . . . . 97
6.7.2 Structural Model Parser . . . . . . . . . . . . . . . . . . . . . . . 98
6.7.3 Index Reduction Parser . . . . . . . . . . . . . . . . . . . . . . . 98
6.7.4 State Space Form Construction Parser . . . . . . . . . . . . . . . 99
6.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

III Application of Methodology 103

7 Performance Estimation in Industrial Gas Turbine Engines 105
7.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.1.1 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.2 Introductory Methods To Detect Compressor Fouling . . . . . . . . . . 110
7.2.1 Bell-Mouth Based Estimation . . . . . . . . . . . . . . . . . . . . 111
7.2.2 Pressure Ratio Based Mass Flow Estimation . . . . . . . . . . . 112
7.2.3 Power versus Mass Flow of Fuel . . . . . . . . . . . . . . . . . . 112
7.2.4 Performance Model Based Mass Flow Estimation . . . . . . . . 112
7.3 Measurement Delta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.4 Constant Gain Extended Kalman Filter . . . . . . . . . . . . . . . . . . . 117
7.4.1 Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.4.2 Observer Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.4.3 Filter Design Summary . . . . . . . . . . . . . . . . . . . . . . . 120
7.5 Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.5.1 Evaluation 1: Atmospheric Weather Condition Dependence . . 121
7.5.2 Evaluation 2: State Estimation for Control . . . . . . . . . . . . 122
7.5.3 Evaluation 3: State Estimation for Monitoring . . . . . . . . . . 124
xii Contents

7.5.4 Discussion of the Results in Evaluation 2–3 . . . . . . . . . . . . 127

7.5.5 Nonlinear versus Linear Estimator . . . . . . . . . . . . . . . . . 127
7.6 Summary of the Performance Estimation Techniques . . . . . . . . . . 128
7.6.1 Bell-Mouth Based Estimation . . . . . . . . . . . . . . . . . . . . 128
7.6.2 Measurement Delta . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.6.3 Constant Gain Extended Kalman Filters Constructed Using the
Proposed Equation Selection Procedure . . . . . . . . . . . . . 132
7.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

8 Investigation of Fault Diagnosis in the Startup and Shutdown Operating Modes 135
8.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.2 Test Construction Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.2.1 Fault Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.2.2 Diagnosis Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2.3 Test Quantity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.3 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.3.1 Fault Free Sequence . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.3.2 Component Faults – Leakage in the Compressor and Increased
Friction in Mechanical Bearings . . . . . . . . . . . . . . . . . . 139
8.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

9 Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines 145

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.1.2 Outline and Contributions . . . . . . . . . . . . . . . . . . . . . 148
9.2 Gas Turbine Diagnosis Modeling . . . . . . . . . . . . . . . . . . . . . . 148
9.2.1 Measurement Signals . . . . . . . . . . . . . . . . . . . . . . . . 149
9.2.2 Health Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.2.3 Sensor and Actuator Faults . . . . . . . . . . . . . . . . . . . . . 151
9.2.4 Constraints on Performance Deviation . . . . . . . . . . . . . . 152
9.2.5 Differential Algebraic Equation Form . . . . . . . . . . . . . . . 153
9.2.6 State Space Form . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.3 Test Quantity Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.3.1 Constant Gain Extended Kalman Filter . . . . . . . . . . . . . . 155
9.3.2 CUSUM Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.4 Fault Isolation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
9.4.1 No Fault Hypothesis H 0 . . . . . . . . . . . . . . . . . . . . . . . 159
9.4.2 Fault Hypothesis H i . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.4.3 Component Failure and Foreign Object Damage . . . . . . . . 159
9.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9.5.1 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9.5.2 Fault Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9.5.3 Fault Free Mode: FN F . . . . . . . . . . . . . . . . . . . . . . . . 162
9.5.4 Sensor Fault Modes: F p 3 , Ft 3 , and Ft 7 . . . . . . . . . . . . . . . 163
9.5.5 Actuator Fault Modes: Fm f and FPA . . . . . . . . . . . . . . . . 164
Contents xiii

9.5.6 Sensor Fault Mode: Fn C1 . . . . . . . . . . . . . . . . . . . . . . . 165

9.5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.6 Experimental Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . 167
9.6.1 Evaluation 1: Fault Detection and Isolation . . . . . . . . . . . . 168
9.6.2 Evaluation 2: Supervision Based on Fault Compensation in Test
Quantity T0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
9.6.3 Evaluation 3: Fault Tolerant Supervision of Performance . . . . 174
9.6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
9.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

10 Conclusion 177

References 179

A Mole/Mass Conversions 187

A.1 Mole/Mass Fraction Calculation . . . . . . . . . . . . . . . . . . . . . . . 187
A.2 Stoichiometry Matrix Expressed in Mass . . . . . . . . . . . . . . . . . . 188
A.3 Determination of Stoichiometric Air/Fuel Ratio . . . . . . . . . . . . . . 189

B Measurement Plots 191

B.1 Ambient Temperature T0 . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
B.2 Ambient pressure p 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
B.3 Shaft Speed n C1 of the Gas Generator . . . . . . . . . . . . . . . . . . . . 194
B.4 Generated Power PA by the Application . . . . . . . . . . . . . . . . . . 195
 Part I

Introduction
 Chapter 1

Background and Motivation

Diagnosis and supervision of industrial gas turbines are of vital importance since it
gives valuable information about: (i) process health, (ii) instrumentation failure, and
(iii) component faults. Air leakage in valves and bearing failures are examples of unde-
sired behaviors which can be represented by the general term: component faults. When
the actual health state of the gas turbine is known, it is easier to achieve: (i) more
reliable operations, (ii) lower heat stress in components, (iii) lower fuel consumption,
and (iv) more efficiently planed overhaul and maintenance. All these factors reduce the
environmental impact and improving the operation profitability for the customer. The
health state of an industrial gas turbine degrades gradually due to certain factors such as:
(i) environment air pollution, (ii) fuel content, and (iii) ageing to mention some of the
degradation factors. The compressor in the gas turbine is especially vulnerable against
contaminants in the air since these particles are stuck at the rotor and stator surface. The
loss in compressor performance, due to fouling, can partially be restored by an online or
offline compressor wash. Sensor and actuator faults which are left undetected affect the
engine’s operation point and complicate the determination of a suitable time when the
compressor should be washed. Thus, it is crucial to have a Fault Detection and Isolation
(FDI) system to detect and isolate those faults at an early stage.
The deterioration in components affect the diagnosis statements and estimates used
by the controller. For monitoring and control issues a good idea is to introduce parame-
ters which represent correction factors of performance from nominal baseline. These
parameters are denoted health parameters in the gas turbine diagnosis literature (Volponi,
2014). The motives to use those extra parameters are: (i) the health parameters are
indicators for overhaul and maintenance, and (ii) the health parameters compensate for
the deterioration in components which result in more reliable estimates of unmeasured
signals used by, e.g., the controller.
In gas turbine diagnosis and control field, it is crucial to have good estimation
of performance. Generally speaking, a model that has high accuracy gives smaller
prediction errors than a model with a lower accuracy. The same argument is also valid

3
4 Chapter 1. Background and Motivation

for model-based diagnosis statements. Thus, one valuable factor which increase the
diagnosis performance, i.e., increase the fault detection probability and lower the false
alarm probability, is the usage of a model with high accuracy within the diagnosis tests.
A common approach is to have diagnosis tests (or filters) which are based on a physical
engine model (Kobayashi et al., 2005), thus if any components in the model are removed
or replaced, also the filters need to be redesigned. The design of those filters might be a
complex task and involves a lot of manual work, especially when the model is a large
nonlinear differential algebraic equation (DAE) model. Thus, a systematic and automatic
design procedure of the filters is desirable when the FDI-system is constructed.
Another perspective is the ability to simplify the FDI-system design by using available
models, which are already developed, in the diagnosis tests. The industry partner Siemens
Industrial Turbomachinery AB (SIT) in Finspång has provided a reference gas turbine
model which is developed and evaluated during a long period of time. The reference
model is built from an in-house thermodynamic library SITLib which is implemented
in the modeling language Modelica. The reference model together with its surrounding
components is mainly used for performance analyses and other in-house tools are
considered for diagnosis and supervision statements. An overall idea with this work
is to integrate the performance model also in the design of the FDI-system. A lot of
work and money have been spent on development, validation, and maintenance of the
reference model. Thus, a good idea to reuse as much knowledge as possible from the
model used for performance analysis also in the FDI-system which is simplified by using
a structured and systematic approach when the diagnosis tests are constructed. With the
tool chain from the physical model to the performance monitoring that is proposed and
demonstrated in the thesis such a benefit can be achieved.

1.1 Problem Statement

The objective of this work is to investigate a model based approach for diagnosis and
supervision of industrial gas turbines. Since the available gas turbine fleet consists
of a large number of individuals, where all of them have their own properties and
are running under different ambient conditions, it is desirable that the design of the
diagnosis and supervision system is systematic. The intention with a systematic design
is: (i) the diagnosis tests for different gas turbine hardware configurations should be
generated easily, and (ii) the equations which are consider in the diagnosis tests should
be selected carefully from the model used for performance analysis. The systematic
design is especially important since the available reference gas turbine model, used for
performance analysis, is a large differential algebraic equation (DAE) model which is
nonlinear. Early investigations show that the reference model has unobservable state
variables which need to be removed when observer based diagnosis tests are constructed.
1.2. Thesis Outline 5

1.2 Thesis Outline

The thesis is divided into two main parts. The first part addresses the gas turbine modeling
and the systematic design of the FDI-system. In the second part, the FDI-system is
evaluated for different configuration using case studies with simulated and experimental
data.
The first part consists of Chapters 4, 5, and 6. Chapter 4 summarizes usable thermo-
dynamic concepts for gas turbine modeling. In Chapter 5, the developed gas turbine
modeling package is presented. Chapter 6 discusses how the diagnosis tests in the FDI-
system are constructed.
The second part consists of Chapters 7, 8, and 9. In Chapter 7, the FDI-system is
evaluated where the focus is on estimation of performance when the engine degrades.
Chapter 8 discusses how the methodology can be used to detect and isolate faults when
the gas turbine startup and shutdown. In Chapter 9, the FDI-system is extended with
more tests to also diagnose faults in sensor and actuator at the same time the performance
is supervised.
 Chapter 2

Industrial Gas Turbine Background and

Simulation Environment

2.1 Gas Turbine Theory

A gas turbine engine is preferably used in many application when mechanical power is
desired since it has: (i) a high power-to-weight ratio, (ii) a robust design (few moving
parts, one directional rotation, acceptable vibration, etc.), (iii) often high reliability,
(iv) low lubricating oil consumption, (v) an exhaust gas where most of the waste heat is
collected (which can be used in a combined cycle), (vi) low emissions of carbon monoxide
(CO) due to the excess of oxygen, and (vii) a wide spectrum of usable fuels to mention
some of the main advantages. Depending on the application, the mechanical power can
also be converted to other energy forms, e.g., electricity in a generator. Another working
area for gas turbines is the so-called mechanical drive where the applied load is, e.g., a
pump or an external compressor.
To create mechanical power in the gas turbine, the first step is to compress the
working fluid using the compressor. In the second step, fuel is burned in the compressed
fluid supplied by the compressor and the temperature is increased. In the final step, the
fluid is expanded through the turbine at the same time as the temperature and pressure
are decreased. In the gas turbine, the working fluid is in most cases atmospheric air. The
work that is left after the compression work is subtracted from the work generated by
the turbine is the produced mechanical power which is transferred to the application.

2.1.1 Brayton Cycle

The open gas turbine cycle is best described by the Brayton cycle see, e.g., Giampaolo
(2009); Horlock (2007); Saravanamuttoo et al. (2001). In the ideal Brayton cycle (ex-
pressed in temperature T and entropy s) the entropy is constant during the compression
(1–2) and the expansion (3–4) phases. A Brayton cycle with two turbines is shown in

7
8 Chapter 2. Industrial Gas Turbine Background and Simulation Environment

T K
p1
1400
3
Qin
1100
4
800 2
2s 4s p2
500 5s 5
Qout
200 1 s
2 4 6 8 1

Figure 2.1: An ideal (solid lines), and a non ideal (dotted lines) Brayton cycle of a 2-shafted gas
turbine is shown in the figure. In the non ideal cycle, the entropy in the compression and
the expansion phase is not constant. This means that more work needs to be supplied in the
compression phase and less heat is converted to work in the expansion phase, i.e., the entropy
increases. In the non ideal gas turbine cycle, no pressure losses in components are considered.
The numbers in the figure represent the gas path positions which are shown in Figure 2.4.

Figure 2.1. In the figure, the non ideal gas turbine cycle (dotted lines) is also shown where
the entropy increases during the compression and the expansion phases. This leads to
the fact that more work has to be supplied in the compression phase and less heat is
converted to work in the expansion phase. The increase in entropy results in a lower
efficiency of the engine. During the combustion process (2–3), the pressure is constant
and the amount of heat Qin is supplied by the fuel. Since the gas turbine is an open
system, the original state is not reached which is required for a thermodynamic cycle.
Ideally, the pressure in the inlet and outlet of the engine is the same. Thus, it can be
assumed that the amount of heat Qout is left the engine and the cycle can be closed.

2.1.2 Mechanical Drive Application

Gas turbines used for mechanical drive applications are often smaller than engines used
for power generation and have a typical power range of 1–50 MW. These engines are often
featured with a twin shaft design (2-shaft) which means that the power turbine is free to
move independently of the gas generator. The gas generator is used to produce hot gases
which are delivered to the power turbine. The gas generator consists of a compressor and
a turbine, which is also called a spool. Thus, a common configuration of an engine used
for mechanical drive is denoted 1-spool and 2-shafted gas turbine. Since a mechanical
connection between the gas generator and the power turbine is absent, it is possible to
transform between rotational speed and delivered torque for a given amount of output
power. Thus, this type of gas turbine is suitable for mechanical drive applications and at
these sites (often in the connection with the oil and gas industry), the driven component
is, e.g., a pump or an external compressor. The driven component can be used to pump
gas in a pipeline where the speed is adjustable.
2.1. Gas Turbine Theory 9

An important parameter to maintain high efficiency is to have a high flame tempera-

ture in the combustor cans. The hot gases which left the combustor lead to an increased
turbine inlet temperature (TIT). The turbine inlet temperature is often too high for the
material in the first turbine blades which may results in undesired heat stresses. To
reduce the heat stress in the turbines, compressed air is injected as a thin film at the
surface of the blades to protect the material. In modern gas turbines, about 20 % of the
inlet air flow is bled off to perform cooling.
In Figure 2.2, the SGT-750 (successor to SGT-700) developed and manufactured by
SIT in Finspång is shown. The typical application for the SGT-750 (SGT-700) 1-spool
and 2-shafted gas turbine is mechanical drive. The inlet guide vanes (IGVs) are used to
change the angle of the inlet mass flow to obtain as high efficiency as possible regardless
of the load.
axial-flow
compressor
combustor
variable guide cans
vanes

power
output
compressor
bleed valves turbine

free power
air intake turbine
Figure 2.2: Typical appearance of a 1-spool and 2-shafted gas turbine used for mechanical drive
applications. The engine shown in the figure is the SGT-750 (launched 2010 and successor to
SGT-700) developed and manufactured by SIT in Finspång. (Published with permission from
Siemens AG, http://www.siemens.com/press)
The engine studied in the thesis is a 1-spool and 2-shafted gas turbine. For this type
of engine, a gas generator which consists of a compressor C1 and a compressor-turbine
T1 is connected with a power turbine T0. The gas generator is used to produce hot gases
which are delivered to the power turbine. The work that is delivered to the application is
taken from the power turbine.
10 Chapter 2. Industrial Gas Turbine Background and Simulation Environment

2.2 SGT-700 Gas Turbine

For the case studies and model development performed in the thesis, the work is applied
to the SGT-700 gas turbine. The SGT-700 is chosen since models and experimental data
that are received from SIT are for this gas turbine type. The developed methodology
is general and is not just applied for the SGT-700 gas turbine. The SGT-700 is the
predecessor to the SGT-750 shown in Figure 2.2 and is a 1-spool and 2-shafted engine
used for mechanical drive applications. In Figure 2.3, the nominal performance chart of
the SGT-700 gas turbine for different inlet temperatures and power turbine speeds is
shown. According to the figure, the operating range is large. For the ambient temperature
15X C the speed of the power turbine and nominal shaft power can be varied between
3500–7000 rpm and 22–32 MW, where the efficiency increases with the power turbine
speed. The power used for the actual application can be lower than the nominal shaft
power shown in the figure.

2.2.1 Measurement Signals

A schematic representation of the SGT-700 gas turbine is shown in Figure 2.4 together
with a common set of measurement sensors. The sensors measure temperatures Ti
pressures p i , and shaft speeds n i throughout the gas path. The sensor position for the
measured quantity of temperature or pressure is located at different cross-sectional areas
i along the engine’s gas path. The measurement setup for the SGT-700 gas turbine is:
(i) temperature T2 and pressure p 2 before the compressor, (ii) discharge temperature T3
and pressure p 3 after the compressor, (iii) temperature T7 and pressure p 8 after the power-
turbine, (iv) speed of gas generator n C1 and the power-turbine n T0 , and (v) temperature
T0 , pressure p 0 , and relative humidity φ 0 of the ambient air. The subscript notation i in
the measured quantity describes at which cross-sectional area the sensor is located. A
low number represents the air entrance and a high index number represents the position
where the exhaust gas leaves the gas turbine. The sensor position of the gas turbine is
sketched in Figure 2.4. The other measured signals in the figure are: (vi) the mass flow
of fuel m f , (vii) the generated power PA , and (viii) the rotational speed n A of the driven
component. The power signal is estimated by the application and not directly measured.
In some of the cross-sectional areas, the quantity is measured with more than one
sensor. For example, the discharge pressure p 3 is measured with the three sensors p 3,1 ,
p 3,2 , and p 3,3 . The exhaust temperature T7 is measured with sensors in three rings
where each ring has 16 thermocouples. The total number of sensors that measure the
temperature T7 , at different location around the circumference of the three rings, is
48. The large number of thermocouples in the exhaust gas is used, e.g., to monitor the
burners in the combustion chamber to discover if any burner has, e.g., a poor flame.

2.2.2 Actuator Signals

Input signals to the gas turbine are the valve positions. The actuators adjust the positions
of the compressor bleed valves and combustor bypass valve. The bleed valves are usually
used during startup and shutdown sequences to avoid stall and surge in the compressor.
2.2. SGT-700 Gas Turbine 11

MW 
38
30X C
e 15X C
tu r
p era 0X C
t tem
34 b ien 38 %
am
15X C
nominal shaft power

30 37 %
30X C
36 %

26 35 %
45X C

34 %

22 33 %
32 %
31 %
30 %
18 r pm
3000 4000 5000 6000 7000
power turbine rotor speed

Figure 2.3: Nominal performance chart of the SGT-700 gas turbine for different inlet temperatures
and power turbine speeds when a mechanical drive application is considered. The performance
is described using the nominal shaft power and efficiency. (Source: SGT-700 Industrial Gas
Turbine Brochure, http://www.energy.siemens.com)
12 Chapter 2. Industrial Gas Turbine Background and Simulation Environment

At full load operations the bleed valves are fully closed. The combustor bypass valve is
used to control the emission and stability when the load is reduced.

Fuel
p0 T0 φ0 mf PA nA

CC
3.
2. 4.
C1 T1 T0 A

1.
5.
Air
Exhaust

T2 p2 T3 p3 n C1 T7 p8 n T0

Figure 2.4: An overview of the 1-spool and 2-shafted SGT-700 gas turbine is shown in the figure.
This gas turbine consists of a gas generator (compressor C1 and a compressor-turbine T1), a
power turbine T0, and an external application A. The gas generator supplies the power turbine
with hot gases and the power turbine delivers the work demanded by the application. The
cooling air, tapped from the compressor, is shown with dashed arrows in the figure.

2.3 Simulation Platform

For the work in the thesis, a simulation platform is provided by SIT. The simulation plat-
form includes the reference gas turbine which is built from an in-house thermodynamic
library SITLib. The provided simulation platform consists of: (i) a controller, (ii) a fuel
system, (iii) a starter motor, (iv) a transmission, and (v) a 2-shafted gas turbine (reference
model). The simulation platform and its components are shown in Figure 2.5. All of these
components are written in the modeling language Modelica and are simulated using the
tool Dynamic Modeling Laboratory (Dymola). The simulation platform can be used to
investigate: (i) different fuel setups, (ii) startup and shutdown effects, (iii) variation in
speed and power of applied load, and (iv) change in ambient environment conditions
to mention some of the simulation scenarios. The ambient conditions can be changed
according to: (i) the pressure p 0 , (ii) the temperature T0 , and (iii) the relative humidity
φ 0 of the incoming air. A variation of these ambient conditions can affect, e.g., the
efficiencies and mass flows along the engine’s gas path. The change in ambient conditions
has a direct impact on the efficiency and the mass flow since, e.g., the pressure ratio over
the compressor is affected. The change in ambient conditions also modifies the mixture
of substances in the ambient air, i.e., the amount of water steam in the air. This effect
 2.3. Simulation Platform 13

trip

f_sp
100000 fsp trip pv sov main pilot

mv
k=50 y pelsp T
droop
p_ramp sv p X
droop
k=4
load_control
LC
SGT-700 bv1 P
bv2
100000
Ceta = -4.16 fuelSource SGT-700
rh
Pa2kPa Valve LHV = 46.7
charac missing! bpv
p0 ign
Q nT1 Pel f p1 T2 T3 T75 mbm p3 p8

mass flow
k=1/1000

y
Controller
Fuel System

Gas Turbine
Transmission
Starter Motor C Hz MW

50 Hz
SGT-700 load
G

fixed
Pressure

SGT-700 SGT-700
period=1000

nT1 p1 p3 p8 T2 T3T75 mbm

set_load
Temperature

Ambient gain

p T RH k=d_convert

p offset=0
period=1000
T
X
T T
Humidity p p
AmbientAir X X

P P P

Air Comb
period=1000

Figure 2.5: The simulation platform used for, e.g., performance analysis consists of a controller, a
fuel system, a starter motor, a transmission, and a 2-shafted gas turbine. All of these components
are implemented in the Modelica language and are simulated using the tool Dymola. The input
signals are the ambient pressure p 0 , the ambient temperature T0 , the relative humidity φ 0 of
ambient air, and the demand power and frequency of the applied load.
14 Chapter 2. Industrial Gas Turbine Background and Simulation Environment

may also affect the performance parameters along the engine’s gas path.
The advantage with the simulation platform is the ability to study reliably performance
evaluations and parameter estimations throughout the engine’s gas path due to different
operational conditions. The input signals to the simulation platform are, e.g.,: (i) the
properties of the ambient air, and (ii) the demand power and frequency of applied load.
In Figure 2.5, the simulation platform is shown where the application is a 50 Hz electrical
generator. When a mechanical drive application is evaluated, the electrical generator is
replaced with a load where the speed is adjustable.

2.3.1 Reference Gas Turbine Model

A major part of a gas turbine model is the description of the gas media along the engine’s
gas path. The gas media is used to specify the thermodynamic properties of the fluid.
To simplify the handling of the thermodynamic properties in Modelica, they can be
encapsulated by a gas model. In the reference gas turbine model (developed by SIT), the
Modelica Media package is utilizes which is a part of the standard Modelica package.
The gas model in Modelica Media is flexible since substance mixtures of ideal gases with
an arbitrary concentration can be modeled. This approach gives a number of equations
in the gas turbine model that increases drastically with the number of substances in the
gas. The number of equations is increased since each substance in the gas is described
by a separate state variable in each control volume in the model. This result in, roughly
speaking, that the number of substances in the gas is proportional to the number of
equations in the overall gas turbine model. The reference model has about 2500 equations
and 60 state variables which are considered large. More details about the reference gas
turbine model, in an early stage, can be found in Idebrant and Näs (2003).

2.3.2 Modelica
Modelica is an equation based object oriented modeling language with focus on reusing
components and model libraries. In an equation based language the relationships between
variables are specified by the user simultaneously as the causality is left open. An open
causality means that the order to calculate the variables does not have to be specified
by the user. An example is a fluid which obey the ideal law. When the control volume
model is designed, the user specifies the relation of the involved variables according to
ideal law, i.e, the relationship p v R T in the model component. In this simple example,
the pressure p, the specific volume v, and temperature T can be calculated depending
on the available input signals or the surrounding variables. This, together with the object
oriented nature of the language simplifies the construction of component libraries since
models can be reused where the same base class model can be used in all the three control
volume cases.
Another advantage with the Modelica language is the concept of multi-domain
modeling which means that different kinds of physical domains can be encapsulated in
the same model. In the available simulation platform, shown in Figure 2.5, the considered
domains are: (i) the thermodynamic, (ii) the mechanical, and (iii) the electrical domain.
Between domains and components, information is only exchanged through special
2.3. Simulation Platform 15

connection points. Those connection points are called connectors in Modelica. There are
basically two kinds of quantities in a connector and these quantities are either defined
as a flow, or a non-flow quantities. In a connection point, flow variables are summed to
zero and non-flow variables are set equal.
In Modelica, state equations and algebraic constraints can be mixed which results in a
model that is in a differential algebraic equation (DAE) form. For a differential algebraic
equation model, the DAE-index of the model is an important property. For simulation
purposes, a state-space form of the system model is desirable and the DAE-index is one
measure of how easy or hard it is to obtain a state-space form. In general, higher index
problems are often more complicated than lower index problems to simulate. Simulations
of DAE systems are well described in Hairer et al. (1991); Ascher and Petzold (1998).
See the language specification at the webpage in Modelica Association (2007), or
the textbooks by Fritzson (2004); Tiller (2001) for a comprehensive description of the
Modelica language. In Casella et al. (2006), the Media library available in the standard
Modelica package is presented. The Media library is utilized in the reference gas turbine
model developed by SIT.
 Chapter 3

Summary of Contributions

The central topics in this work are the systematic design and the evaluation of a Fault
Detection and Isolation (FDI) system for an industrial gas turbine. For the FDI-system
design procedure, a thermodynamic modeling library GTLib is developed which can be
used to build up a physical model of an industrial gas turbine. This model is then used
for performance analysis and as a starting point when the diagnosis test quantities which
are implemented as Constant Gain Extended Kalman Filters (CGEKFs) are designed.
The designed CGEKFs have been given different diagnosis properties and together they
are used in an FDI-system to calculate the actual health state of the engine and possible
diagnoses. The systematic design procedure assures that relevant equations are chosen
when the test quantities are generated.
The constructed FDI-system is evaluated using case studies based on simulated and
experimental data. It is shown that the designed filters can be helpful when decisions
about a fouled compressor are taken. When the FDI-system consists of more than one
filter, sensor and actuator faults can be diagnosed at the same time as the performance
is supervised. Slow varying sensor and actuator bias faults are difficult to diagnose
(since they can appear in a similar manner as the performance deterioration), but the
FDI-system has the ability to detect these faults. With the proposed method, component
faults can be diagnosed during engine startup and shutdown sequences. Since loads and
stresses in components increase during startup and shutdown, it can be a good idea to
study component failure during these sequences.

3.1 Thesis Summary

The thesis is divided into two main parts, where the first part consists of the modeling
work of the gas turbine and the FDI-design. In the second part, the FDI-system is
evaluated for different configuration using case studies with simulated and experimental
data.

17
18 Chapter 3. Summary of Contributions

In Chapter 4, Thermodynamic Concepts, basic thermodynamic concepts that are

used to describe the gas are presented. The gas model is a part of the GTLib package
and is used along the engine’s gas path. In the chapter, the combustion of air and fuel
is introduced which is based on the assumption of a stoichiometric combustion. The
state of the gas in a control volume is specified using the three state variables: pressure
p, temperature T, and air/fuel ratio λ. When the state variable λ is known, the mass
fraction of species in the exhaust gas can be calculated. Here, the gas with substances:
argon (Ar), oxygen (O 2 ), nitrogen (N 2 ), carbon dioxide (CO 2 ), and water (H 2 O) are
considered. With the air/fuel ratio description, pure atmospheric air can be described
with an infinitely large air/fuel ratio λ.
In Chapter 5, GTLib – Thermodynamic Gas Turbine Modeling Package, the focus
is on the implementation of the gas turbine components described in Chapter 4. These
components are then used in an introductory control volume example where the objective
is on variation in ambient conditions. The constructed (in GTLib) gas turbine model
used for performance analysis is also shown in this chapter.
In Chapter 6, Modeling, Analysis, and Transformation of the Diagnosis Model,
the objective is to design the diagnosis model, and transform this model to a form
suitable for simulation. In the diagnosis model, a number of extra estimation parameters,
i.e., the health parameters are introduced. These parameters should capture deviation
in performance due to fouling, and other factors that can affect the performance. In
this step, also the sensor and actuator fault modes can be specified together with the
measurement setup. The equations, which are used in each diagnosis test, are then
selected by structural methods. Since observer based tests are derived in Chapter 7, the
state variables in the derived test equations must be observable. An observability analysis,
together with an index reduction are performed to get the test equations. A number of
parsers is presented in the chapter. These parsers are used to transform the diagnosis
model into runnable Matlab code. The Matlab environment is used here because of the
available tools for diagnosis analysis that are implemented in Matlab.
In Chapter 7, Performance Estimation in Industrial Gas Turbine Engines, three
studies are presented where techniques of performance deterioration estimations are
investigated. In the first study, a simple approach to calculate deterioration due to
compressor fouling is presented. In the next two studies, the gas turbine model is used as
a base for the estimation techniques. In the second study, the estimations are based on
so-called measurement deltas, which are generally the difference between the simulated,
and the measured gas path quantity. In the third study, a nonlinear Kalman observer
is evaluated on two test cases. In the first test case, simulated data from the reference
platform is evaluated for different operational points and different atmospheric weather
conditions. In the second test case, experimental data from a gas turbine mechanical
drive site in the Middle East is evaluated.
In Chapter 8, Investigation of Fault Diagnosis in the Startup and Shutdown Op-
erating Modes, a pilot study is presented where the objective is to investigate and develop
a diagnosis test quantity used for detection and isolation of component faults in the
startup and shutdown sequences of the gas turbine application. Since the engine goes
through many operating modes for this run, the potential of the nonlinear Kalman filter
3.2. Contributions 19

used for fault estimation and supervision can be utilized fully.

In Chapter 9, Diagnosis and Fault Tolerant Supervision of Industrial Gas Tur-
bines, the objective is to diagnose faults in sensor and actuator signals at the same time
the performance is supervised. To increase the fault detectability, limitations of the health
parameters are introduced. This limitation is introduced since the health parameters,
for some fault modes, may capture the fault which results in residuals which can not
be discriminated from the non-faulty case since measurement noise is present in the
signals.

3.2 Contributions
Figure 3.1 gives an outline of the main methodological contributions of the thesis and
the dashed line shows the boundaries between this work and material provided by SIT.
Analysis of the gas turbine model provided by SIT shows that it is unnecessary complex
for diagnosis purposes, especially with respect to the gas medium. Thus, the first step
(a) is to replace the SITLib with the GTLib. The backward arrows symbolize that the
model constructed using GTLib can be used together with the provided simulation
platform. The GTLib based model and the SITLib based model (reference model) give
the same outputs when: (i) air, fuel, and exhaust gases are used as the working fluids,
and (ii) ambient conditions are not changed. In step (b), the diagnosis properties of
the test quantity are specified. Since Modelica is an object oriented language, model
specifications from GTLib based model can be inherited. In the final step (c), the test
quantity is automatically generated using developed parsers. The overall benefit with the
design procedure is to ensure that the accuracy of the diagnosis test quantity in step (c)
is the same as in the model used for performance analysis in step (a).
The main contributions are summarized in:

• A thermodynamic modeling library GTLib implemented in Modelica is developed.

Using GTLib, a gas turbine model can be constructed which and used for perfor-
mance analysis and diagnosis modeling. In GTLib, the gas medium behavior is
more efficiently modeled but still accurately describe machine behavior. In the
diagnosis model, parameters to estimate component deteriorations and faults are
introduced. These extensions are necessary to be able to design a model based
diagnosis system.

• Based on the diagnosis models, which are of considerable size and complexity, a
second key contribution is an automated FDI-system design method. The method,
described in Section 6, aims to keep the required human work to a minimum.
Important is also that the design technique is not gas turbine specific, but is
applicable to any Modelica based model.

• The third contribution is the case study in Section 7 where the automated design
procedure is applied to a gas turbine diagnosis model and experimentally validated
using measured data from an industrial gas turbine site.
 20 Chapter 3. Summary of Contributions

Experimental
Platform Sensor
Processing

Mechanical Resample
Data
Drive Site Remove idle conditions

Simulation Gas Turbine Diagnosis Diagnosis

Platform Model Model(s) Test(s)
D1 T1

GTLib
Reference   
Based
Gas Turbine (b)
(a) Gas turbine Model
Model Diagnosis
modeling: Di Ti
Equation modeling:
(c) Test
reduction Analysis quantity design:
Previous Work: Lambda Health Structural
Reference GT model concept parameters methods
Simulation platform No component Fault DAE-index
Equation based simplification modeling reduction
SITLib Simulation Sensor State space form
platform selection
Reliable DAE and Kalman filter
nonlinear models compatible Component
simplification Estimates: x̂ i , ŷ i

Figure 3.1: An overview of the experimental and simulation platform together with the key con-
tribution of the systematic test quantity design procedure is shown. The FDI-system design
consists of the three main steps: (a) gas turbine modeling, (b) diagnosis modeling, and (c) test
quantity generation. The goal with the design is to have the same performance and accuracy in
the FDI-system as in the gas turbine model used in the simulation platform simultaneously as
the design procedure is simple enough.
3.3. Publications 21

• The forth contribution is a fault tolerant FDI-system which is presented in Section

9. The proposed FDI-system can be used to detect and isolate a single sensor or
actuator fault at the same time the performance is supervised.

3.3 Publications
The thesis is based on the work presented in the following publications.

Journal Papers
• Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Gas turbine modeling
for diagnosis and control. Journal of Engineering for Gas Turbines and Power, 136
(7):17 pages, July 2014a. doi:10.1115/1.4026598

Submitted
• Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Diagnosis and fault
tolerant supervision of industrial gas turbines. Submitted for journal publication,
2014b

Conference Papers
• Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Fault isolation for an
industrial gas turbine with a model-based diagnosis approach. ASME Conference
Proceedings, 2010(43987):89–98, 2010. doi:10.1115/GT2010-22511
• Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Health Monitoring in an
Industrial Gas Turbine Application by Using Model Based Diagnosis Techniques.
ASME Conference Proceedings, 2011(54631):487–495, 2011. doi:10.1115/GT2011-
46825
• Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Fault Tolerant Supervision
of an Industrial Gas Turbine. ASME Conference Proceedings, 2013(55188), 2013.
doi:10.1115/GT2013-95727

Licentiate Thesis
• Emil Larsson. Diagnosis and Supervision of Industrial Gas Turbines. Technical
report, Department of Electrical Engineering, 2012. LiU-TEK-LIC-2012:13, Thesis
No. 1528, http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-75985
 Part II

Modeling and Design

 Chapter 4

Thermodynamic Concepts

The objective of this chapter is to present thermodynamic concepts that are useful in the
development of a physical based gas turbine model used for: (i) performance analysis,
(ii) supervision of components, and (iii) diagnosis statements. A central part in the gas
turbine model is the description of the gas. During the combustion, fuel molecules react
and break up into smaller molecules simultaneously as heat is released. Since different
gas concentrations appear before and after the combustion, also the thermodynamic
properties (e.g., enthalpy and heat capacity) of the fluid are changed. To reduce heat
stresses in the first blades of the turbine, compressed fresh air is injected as a thin film at
the surface of the blades to protect the material. These two factors result in a gas model
which should be able to handle: (i) ambient air, (ii) exhaust gas, and (iii) a mixing between
the two gases when a stoichiometric combustion using a hydrocarbon fuel is considered,
i.e., when all hydrocarbons are completely oxidized to the products of carbon dioxide
and water. Each fluid is described by a number of substances. The concentration of, e.g.,
the substances in the exhaust gas depends on the ambient conditions, the hydrocarbon
fuel, and the air/fuel ratio. In Modelica standard library, thermodynamic properties
such as heat capacity, enthalpy, and entropy are tabulated using the well known NASA
polynomials (McBride et al., 2002) for a large number of species. Since these polynomials
are provided by the standard library it is a good idea to incorporate this knowledge into
the gas model.

Outline of the Chapter

The main focus in the chapter is to present the thermodynamic concepts which are used
to describe the fluid throughout the gas path of the engine. In Section 4.1, a summary of
basic thermodynamic relations (e.g., thermodynamic laws) and quantities (e.g, internal
energy, and heat capacity) are presented which are used in the development of the gas
model (Section 4.4). In Section 4.2, the description of the thermodynamic quantities such
as: enthalpy, heat capacity, and entropy is introduced. These quantities are modeled using
polynomial interpolation. Two types of combustion processes are presented in Section 4.3

25
26 Chapter 4. Thermodynamic Concepts

which are based on: (i) a chemical equilibrium calculation, and (ii) a stoichiometry
combustion process. A comparison study between these two combustion models for
different temperatures and air/fuel ratios is performed. The concept of a stoichiometric
combustion is then incorporated in the gas model using the air/fuel ratio. The benefit
with using the air/fuel ratio as a state variable is the reduction of model equations in the
gas turbine model, which gives a model which is less computationally demanding and
easier to handle in practise. In Section 4.4, the gas model is presented which is used in
all thermodynamic components. In Section 4.5, the energy conservation for a mixture
of ideal gases is derived, which leads to the specification of the state equations in the
control volume model presented in Section 4.6.

4.1 Thermodynamic System

The intention with the present chapter is to introduce important relationships and give
an introductory insight for models that are utilized when a thermodynamic system is
designed. A thermodynamic system is defined as an amount of space together with a
surrounding boundary against its environment where relationships between heat and
work are studied. The thermodynamic system can either be open or closed. In an open
system, the boundary lets mass, heat, and work passing through. In a closed system, the
boundary only allows heat and work to be transferred. A gas turbine can be defined
as a thermodynamic system since it converts heat (which is released from the fuel) to
mechanical work (which is used to drive, e.g, a generator). Since a gas turbine exchange
masses across the boundaries it is an open system, thus only open systems are studied in
the thesis. An important thermodynamic component is a control volume. The control
volume is central since it stores energy and keeps track of the state of the thermodynamic
part of system. For a more comprehensive thermodynamic survey, see, e.g., Eastop and
McConkey (1993); Heywood (1988); Borman and Ragland (1998); Turns (2006); Öberg
(2009).

4.1.1 Thermodynamic Quantities

The state of a thermodynamic system can be described by a number of quantities. The
most commonly occurring quantities are: temperature T, pressure p, specific volume v,
mass m, enthalpy h, and and internal energy u. For a gas that occupy a volume V , the
state of the gas can be described with an independent pair of thermodynamic quantities.
Depending on this choice, the appearance of the described system equations are different.
An example of an independent pair of state variables, for a known volume, is the pressure
p and temperature T. From the state variables, all the other thermodynamic quantities
can be derived, e.g., mass m and enthalpy h. In a thermodynamic system it is often
possible to measure both the pressure and the temperature and therefore are these
quantities called measured quantities since they are measurable. When a thermodynamic
system is analyzed, it can be convenient to introduce quantities that are not directly
measurable. These type of variables are called intermediate quantities. Internal energy u
and enthalpy h are examples of such quantities. The relationship between internal energy
4.1. Thermodynamic System 27

u and enthalpy h is defined:

h u  pv (4.1)
where p is the pressure, and v is the specific volume of the gas. In open systems, the
enthalpy is suitable to consider since it encapsulates both the internal energy and the
work applied to the system. The applied work comes from the flowing fluid which affects
the control volume.

Intensive and Extensive Properties

Thermodynamic quantities which do not vary with the size of the system is classified as
an intensive property. Temperature T, pressure p, and density ρ are examples of intensive
properties. Quantities which do vary with the size of the system has an extensive property.
Volume V , and total energy U are examples of extensive properties. For each extensive
property, an intensive property can be obtained when the quantity is divided by, e.g., the
mass. Mass specific quantities are denoted with lower case letters in the thesis. Upper
case letters are usually applied to denote the total amount of a certain quantity in the
system. In some cases it is more suitable to consider the mole specific quantities. In the
thesis, the tilde convention over the corresponding mass specific quantity is used. The
total energy can, e.g., be expressed either in mass or in mole according to:

U mu nũ (4.2)

where m is the total mass, and n is the total number of moles in the gas. Since the molar
mass is defined M m~n, the relationship between the internal energy in (4.2) is:

ũ
u (4.3)
M

Specific Heat Capacities

To describe the amount of energy that is needed to increase the temperature of the fluid
one degree, for a unit mass, the specific heat capacities are used. Since the amount of
energy that is required for a system that undergoes a constant-volume or a constant-
pressure thermodynamic process is different, two specific heat capacities cv and c p are
defined according to:
’ ∂q “ ’ ∂q “
cv , cp (4.4)
” ∂T • ” ∂T •
v p

where q is the supplied heat, v denotes a constant-volume process, and p denotes a

constant-pressure process. An example of a constant-volume process is a fluid in a bomb
calorie meter and an example of a fluid that undergoes a constant-pressure combustion
process is the fluid in a bunsen burner. The combustion chamber in a gas turbine is
typically described by a bunsen burner. A fluid which goes from an initial state to a final
state is said to be reversible if the fluid can be restored to their original state at the same
time no change in its surroundings is caused. The heat capacities of a reversible process
28 Chapter 4. Thermodynamic Concepts

can be written:
’ ∂u “ ’ ∂h “
cv , cp (4.5)
” ∂T • ” ∂T •
v p
where the first law of thermodynamics (4.7) for a reversible thermodynamic process
(4.8) together with the differential of the enthalpy definition (4.1) are considered. The
ratio of the specific heat capacities γ is defined:
cp
γ (4.6)
cv
The specific heat ratio is frequently used when an isentropic compression process or an
expansion process is considered.

4.1.2 Thermodynamic Laws

The first law of thermodynamics states that the energy in a system that undergoes a
closed thermodynamic cycle cannot either be created, or destroyed. The energy is
merely converted between thermal energy (heat) and mechanical energy (work). For
a thermodynamic cycle that is open, the intrinsic energy of the fluid can increase or
decrease. The first law of thermodynamics is written:
dU dQ  dW (4.7)
where U is the internal energy, Q is the supplied heat, and W is the supplied work.
The sign conventions of the energy flows are shown in Figure 4.1, where positive flow
directions point into the control volume.

dQ
control volume

in flows out flows

dm i dU dm j

boundary

dW

Figure 4.1: Sign conventions for an open thermodynamic system are shown in the figure. Positive
flow directions point into the control volume. For a time interval dt, the amount of heat dQ is
added to the system, the work dW is done on the system, the mass dm i is added to the system,
and dm j is removed from the system.

When the system undergoes a reversible thermodynamic process, the supplied work
is dW  p dV and the first law of thermodynamics is rewritten:
dU dQ  p dV (4.8)
4.2. Thermodynamic Properties of Pure Substances 29

where p is the pressure and V is the volume. The second law of thermodynamics is
written:
dQ
B dS (4.9)
T
where S is the entropy, and T is the temperature. The equality in (4.9) holds for all
reversible processes.

4.2 Thermodynamic Properties of Pure Substances

A gas media, used in a thermodynamic system, can either consist of a pure substance or
a mixture of substances. These substances are called species. The atmosphere air media
consists of the species, e.g., nitrogen, oxygen, and argon. Once the composition of the
gas mixture is known, thermodynamic properties such as enthalpy, entropy, and heat
capacity can be determined either on a mass basis, or on a mole basis as shown in:

h Σx i h i , h̃ Σ x̃ i h̃ i (4.10a)
s Σx i s i , s̃ Σ x̃ i s̃ i (4.10b)
cp Σx i c p,i , c̃ p Σ x̃ i c̃ p,i (4.10c)

where x i is the mass concentration, and x̃ i is the mole concentration of substance i.

In this section, the thermodynamic properties for the species in (4.10) will be pre-
sented. To describe these thermodynamic properties of an ideal gas, tabulated data can
be used. The NIST-JANAF thermochemical tables in Chase (1998) consist of tabulated
data for a large number of species. The NIST-JANAF tables are well known, and the
thermodynamic data is available in a wide range of pressures and temperatures with high
accuracy. Since the data is in a tabular form, it can be necessary to interpolate between
the points depending on the application.
Another method to describe gas properties is to use polynomial curve fitting tech-
niques. The main advantage with using polynomials is the ability to encapsulate a large
amount of thermodynamic data with only a few polynomial coefficients. Since polyno-
mials are continuous, they can be differentiated easily and no interpolation is necessary,
thus the simulation time can be reduced.
An early chemical equilibrium program contribution named CEC71 is presented in
Gordon and McBride (1971). In CEC71, the heat capacity is described by a fourth-order
polynomial with constant coefficients a 1 , . . . , a 5 . These coefficients are approximated
with a least-square technique (McBride and Gordon, 1992). To describe enthalpy and
entropy the heat capacity coefficients are extended with the coefficients a 6 and a 7 . For
every species, two sets of coefficients are available. These sets are divided into a low
temperature interval 200 to 1 000 K and a high temperature interval 1 000 to 6 000 K.
The chemical equilibrium program (CEA) presented in Gordon and McBride (1994)
is an extension of the previous developed CEC71 program. In the new program, the
thermodynamic heat capacity data is represented by two more coefficients. An additional
temperature interval 6 000 to 20 000 K is also added for some of the substances. A
summary of the NASA Glenn least-square coefficients and the tabulated thermodynamic
30 Chapter 4. Thermodynamic Concepts

data are shown in McBride et al. (2002). In the paper, the enthalpy of formation ∆ f h o and
the difference in enthalpy H0 between the datum state temperature To and temperature
at 0 K are also tabulated.

Datum State
For thermodynamic systems which consist of a mixture of substances, the energy which
is stored in chemical bonds must be included in the enthalpy description. Therefore, a
datum state is defined for the temperature and the pressure. The reference state of the
NASA Glenn polynomials (McBride et al., 2002) is represented by the datum temperature
To 298.15 K, and the datum absolute pressure p o 1.01325 bar 1 atm. In the present
section, the reference datum state is denoted with the super-script o . In other parts of
the thesis, the datum state notation is omitted for simplicity.

Reference Elements and Enthalpy of Formation ∆ f h o

To each tabulated substance, a value called enthalpy of formation ∆ f h o is assigned. The
enthalpy of formation is defined to be the energy that is released when the substance is
split up to its reference elements in the datum state. An example of reference elements are:
argon Ar (g), carbon C (c), hydrogen H 2 (g), nitrogen N 2 (g), and oxygen O 2 (g). The
symbol (g) indicates that the element is in a gaseous phase and the symbol (c) indicates
that the element is in a condensed phase. For the reference elements the enthalpy of
formation is equal to zero, hence:
∆ f h o ˆTo  0
for the datum state temperature To .

Standardized Enthalpy
The standardized enthalpy h o ˆT , defined in Turns (2006), is the sum of the enthalpy of
formation ∆ f h o ˆTo  and the sensible enthalpy h so ˆT . The enthalpy of formation takes
into account the energy associated with the chemical bonds and the sensible enthalpy
takes into account only the changes in temperature.

h o ˆT  h o ˆTo   h o ˆT   h o ˆTo  (4.11)

where the sensible enthalpy is defined:
T
h so ˆT 
h o ˆT   h o ˆTo  ∫ To
c p ˆτ dτ (4.12)

according to the definition (4.4) of the heat capacity c p . For all substances at the datum
state, the enthalpy of formation is arbitrary assigned the same value as the enthalpy:
∆ f h o ˆTo 
h o ˆTo  (4.13)
The standardized enthalpy (4.11) is now written:
T
h o ˆT  ∆ f h o ˆTo   ∫
To
c p ˆτ dτ (4.14)
4.2. Thermodynamic Properties of Pure Substances 31

When another reference state is used, e.g., the reference state Tô 0 K it is possible
to adjust (4.11) with the tabulated constant bias term H0:
H0 h o ˆTo   h o ˆ0
since:
T
h ô ˆT 
h o ˆT   H0 ∆ f h o ˆ298.15  ∫
0
c p ˆτ dτ (4.15)
when the standardized enthalpy (4.14) is considered at the right hand side of (4.15).

4.2.1 Specific Heat Capacity of Pure Substances

The NASA polynomial for the specific heat capacity c̃ p of a pure substance i has the
structure:
c̃ p,i 1 1
a 1,i 2  a 2, i  a 3,i  a 4, i T  a 5,i T 2  a 6,i T 3  a 7, i T 4 (4.16)
R̃ T T
where the constants a j,i are the tabulated NASA Glenn Coefficients. The left hand side
of (4.16) is a dimensionless quantity, hence it is possible to reformulate it as:
c̃ p,i c p, i
(4.17)
R̃ R
where R̃ is the universal gas constant, and R is the specific gas constant. The relation
between the gas constants is R̃ MR, where M is the mole mass. Equation (4.17) shows
that both the mass and the molar specific quantities can be calculated from the same
polynomial coefficients.

4.2.2 Standardized Enthalpy of Pure Substances

The enthalpy is related thermodynamically to the heat capacity as follows:
h ˆT  ∫ c p ˆτ dτ b1
 (4.18)
RT RT T
where b 1 is an integration constant and the heat capacity c p is integrated with respect
to the temperature T. To obtain the standardized enthalpy of the pure substances, the
integration of (4.16) is performed to get:

h oi 1 lnˆT  a 4,i a 5,i 2

 a 1,i  a 2,i  a 3,i  T T 
RT T2 T 2 3
a 6,i 3 a 7,i 4 1
 T  T  b 1,i (4.19)
4 5 T
where the constant b 1,i is chosen to match (4.14) for the reference temperature T To .
Thus, the enthalpy of formation is included in the NASA polynomials by default. When
the sensible enthalpy is needed, the enthalpy of formation is subtracted from the NASA
polynomial calculations. The coefficients a j, i are the same as for the heat capacity in
(4.16).
32 Chapter 4. Thermodynamic Concepts

4.2.3 Standardized Entropy of Pure Substances

The entropy is related thermodynamically to the heat capacity as follows:

s ˆT  c p ˆτ dτ
R ∫ RT
 b2 (4.20)

where b 2 is an integration constant. To obtain the standardized entropy of the pure

substance i, the integration of c p ~T in (4.16) is performed to get:

s oi a 1,i 1 1 a 5,i 2
  a 2,i  a 3,i lnˆT   a 4, i T  T 
R 2 T2 T 2
a 6,i 3 a 7,i 4
 T  T  b 2,i (4.21)
3 4
where b 2,i is an integration constant. The coefficients a j,i are the same as for the heat
capacity in (4.16).

Entropy for an Ideal Gas

For an ideal gas, the entropy depends on the temperature and the pressure. When the
first (4.7) and second (4.9) law of thermodynamics are combined, together with the
enthalpy definition (4.1) and an assumption of a reversible thermodynamic process (4.8),
the entropy differential can be written:

cp R
ds dT  d p (4.22)
T p

where the relation dh c p dT of an ideal gas is introduced. To get an expression for the
entropy, the differential (4.22) is integrated to get:

T c p,i ˆτ  p p
s i ˆT, p ∫ To τ
dτ  R ln Š o 
p
s oi ˆT   R i ln Š
po
 (4.23)

where the entropy is calculated by the integration of c p ~T, i.e., the expression (4.21).

4.2.4 Gibbs Free Energy

When the temperature T and the pressure p are given, Gibbs free energy is feasible
to consider in the determination of mixture concentration of substances which are in
chemical equilibrium. Gibbs free energy is minimized when substances in a mixture are
in chemical equilibrium. The definition of Gibbs free energy is:

g ˆT, p h ˆT   TsˆT, p (4.24)

where h is the enthalpy, and s is the entropy defined in (4.23).

4.3. Combustion 33

4.3 Combustion
For the modeling work, the combustion process is one of the central parts to consider in
the gas turbine model. During the combustion, chemical energy in the fuel is released
through dissipation of heat in the fluid simultaneously as the reactants break up into
smaller substances. The composition of substances in the exhaust gas is different from
the composition of the unburned mixture, which results in different thermodynamic
properties of the gas before and after the combustion. With an assumption of a stoichio-
metric combustion, the concentration of the exhaust gas mixture is not dependent on
the pressure and the heat which is released during the combustion. In a general case, the
composition of the substances in the exhaust gas reacts with each other. The reacting
rate depends on the temperature and the pressure of the fluid, i.e., so-called dissociation.
Usually, a higher temperature and a higher pressure give a higher reacting rate. The
number of dissociation products that have to be considered is also larger for higher
temperatures and pressures. For example, nitrogen oxide NO x molecules appear in a
lean exhaust gas for high temperatures. The dissociation products have a significant effect
on the heat capacities, but a disadvantage with the consideration of a dissociation term
in the gas model is the increasing complexity. For lower temperatures (@ 1500 K) a good
simplification is to assume that the exhaust gas composition is frozen, i.e., independent
of pressure and temperature so the dissociation terms can be neglected. This gives an
assumption of a stoichiometric combustion.
A main goal here is to develop a gas model that can be used both for the compressed
air and for the exhaust gas. This introduces the combustion into the gas model, where
the concentration of the exhaust gas mixture is described using the air/fuel ratio λ (see
Section 4.3.1). The ambient air (or the air before the combustor) is described using a very
large air/fuel ratio, i.e., λ AA 1. In the gas model, dissociation effects are not described
which result in an absence of NO x molecules. It will be shown that the dissociation
effect for higher temperatures have a significant influence on the heat capacities (see
Example 4.2). The advantages with the stoichiometric modeling approach are that the
combustion can be described using only an extra state variable, the thermodynamic
properties (e.g., enthalpy and heat capacity) vary with the air/fuel ratio, and it is simple
to change the concentration of the reactants, i.e., the ambient air and the hydrocarbon
fuel.

Adiabatic Flame Temperature of a Constant-Pressure Combustion Process

For an ideal gas turbine cycle, it is often assumed that the pressure during the combustion
process is constant. For an adiabatic constant-pressure combustion process, it holds that
the enthalpy h u before the combustion and the enthalpy h b after the combustion are the
same. At least locally, h b is invertible which gives a solution for the flame temperature:

Tb h b 1 Šh u ˆTu  (4.25)

where Tu is the temperature of the unburned mixture, and Tb are the temperature of
the exhaust gas. Depending on the chosen pair of independent thermodynamic state
variables the expression (4.25) needs to be evaluated or not evaluated in each control
34 Chapter 4. Thermodynamic Concepts

volume. When the temperature is chosen as a state variable, the enthalpy needs to be
calculated in each control volumes and (4.25) has to be solved with a numerical solver.
When the enthalpy is instead chosen as a state variable, it is not necessary to find an
explicit solution of the temperature (and solve (4.25)) in each control volume. Solving
(4.25) can be time consuming and can be avoided if the enthalpy is chosen as a state
variable instead of temperature.

4.3.1 Stoichiometry
When sufficient oxygen is available in the air, a hydrocarbon fuel can be completely
oxidized with the rest products of carbon dioxide and water. For the amount of air that
just converts all hydrocarbons to carbon dioxide and water, a stoichiometric air/fuel ratio
is defined. The stoichiometric air/fuel ratio can either be expressed in moles ˆA~F s̃ or
in masses ˆA~F s . The actual air/fuel ratio λ, expressed in a mole basis of the mixture is
defined:
n a ~n f
λ (4.26)
ˆ A~F s̃

where n a is the mole number of air, and n f is the mole number of fuel. The air/fuel ratio,
corresponding to an expression in mass basis, is defined:

m a ~m f
λ (4.27)
ˆ A~F s

where m a is the mass of air, and m f is the mass of fuel. In general, stoichiometric
air/fuel ratios expressed in mass and mole are not equal, i.e., ˆA~F s̃ x ˆA~F s . A simple
reaction formula for a hydrocarbon fuel, with emphasis on combustion, is presented in
the following example.

Example 4.1
A hydrocarbon fuel propane C 3 H 8 is combusted with air that consists of oxygen and
nitrogen. A simple air model assumption is that for each oxygen molecule O 2 , 3.773
nitrogen molecules N 2 are available. This gives the following reaction formula, in mole
basis, for the combustion:
Fuel Air
­ ³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ·¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹µ
C 3 H 8 n ˆO 2  3.773N 2  Ð 3CO 2  4H 2 O  ˆn  5O 2  n 3.773N 2

where n is the number of available oxygen molecules in the unburned mixture where
one propane molecule is present. When n 5 the oxidization is complete, i.e., all
hydrocarbons have been converted to carbon dioxide and water. If n @ 5, there are not
enough hydrocarbons in the combustion so the formula has an excesses of oxygen. In the
present example, it is assumed that n C 5, but for the case with a n @ 5, it is not enough
oxygen molecules for the carbon dioxide molecule formations. Instead, the formula has
to be extended with carbon oxide molecules CO to preserve the number of atoms in the
4.3. Combustion 35

chemical reaction. Since the combustion in a gas turbine has an excesses of oxygen, this
case is not considered in the gas model.
To simplify the presentation of the chemical reaction, it can be written in a matrix
form according to:

Air < 0 = <3= < CO =

Fuel @ A @ A @ 2A
­ ³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ·¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹µ @ A @ A @ A
@ 0 A @4A T
C 3 H 8 n ˆO 2  3.773N 2  Ð Š@
@ n A @5A
A@ A
@H 2 O A
@
@ O2 A
A (4.28)
@ A @ A @ A
@3.773n A @ 0 A @ N2 A
> ? > ? > ?
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ±
n air S̃

where the stoichiometric matrix S̃ consists of the coefficients of the hydrocarbon molecule
combustion. The advantage with the introduced stoichiometric matrix is the ability to
use fuels with several hydrocarbon molecules in a compact manner. The right hand side
of (4.28) describes the number of constructed molecules in the combustion.
For n 5, the oxidation is complete and the stoichiometric air/fuel ratio can be
calculated (i) in mole: ˆA~F s̃ 23.87, and (ii) in mass: ˆA~F s 15.68 according to the
definition of air/fuel ratio (4.26) and (4.27).

The chemical reaction formula in Example 4.1 can be extended to capture more general
descriptions of the ambient air and the hydrocarbon fuel. Here, it is assumed that the
fuel consists of the following molecules: methane CH 4 , ethane C 2 H 6 , propane C 3 H 8 ,
carbon dioxide CO 2 , and nitrogen N 2 . The chemical reaction for these species is written:

CH 4  2O 2 CO 2  2H 2 O
C 2 H 6  3.5O 2 2CO 2  3H 2 O
C 3 H 8  5O 2 3CO 2  4H 2 O (4.29)
CO 2 CO 2
N2 N2

where the species of CO 2 and N 2 are unaffected by the combustion. The corresponding
substances for the air are the following: argon (Ar), carbon dioxide (CO 2 ), water (H 2 O),
nitrogen (N 2 ), and oxygen (O 2 ). To get a more flexible description, also the concentra-
tions of respective substance in the mixtures are considered. The mole concentration
vectors of air x̃ a and fuel x̃ f are expressed according to:

’ x̃ a,Ar “ ’ x̃ f ,C H 4 “
– x̃ a,C O 2 — – x̃ f ,C 2 H 6 —
– — – —
x̃ a – x̃ a,H 2 O —, x̃ f – x̃ f ,C 3 H 8 — (4.30)
– — – —
– x̃ a,N 2 — – x̃ f ,C O 2 —
” x̃ a,O 2 • ” x̃ f ,N 2 •

where the sums of the elements are i x̃ a,i i x̃ f ,i 1. When the stoichiometric P P
reaction formula in (4.29) is combined with the method described in Example 4.1, the
36 Chapter 4. Thermodynamic Concepts

following combustion formula can be written:

< Ar =
@ A
@ A
T @ CO 2 A
x̃ f  n x̃ a Ð @
Šn x̃ a  S̃ x̃ f  @H 2 O A
@
A
A
(4.31)
@ N2 A
@ A
@ O A
> 2 ?

where the stoichiometric matrix (Stephanopoulos et al., 1998):

’0 0 0 0 0“
– 1 2 3 1 0—
– —
S̃ –2 3 4 0 0— (4.32)
– —
–0 0 0 0 1—
”2 3.5 5 0 0•

is expressed in moles. The element ˆ i, j in the matrix symbolize the number of air
species x̃ a,i that are created/depleted from each species x̃ f , j in the fuel. In practice, it
is more suitable to have the description in masses instead of moles. The conversion
procedure is shown in Appendix A.2. The stoichiometric matrix can be rewritten:

’ 0 0 0 0 0“
MC O2 M M
–
– M C H4
2 M CC OH2 3 M CC OH2 1 0—
—
2 6 3 8
– MH O M M —
S –2 2
3 M CH 2HO 4 M CH 2HO 0 0— (4.33)
– M C H4 —
– 2 6 3 8 —
–
–
0 0 0 0 1—
—
MO M M
”2 M 2
C H4
 3.5 M C OH2  5 M COH2 0 0•
2 6 3 8

where M i is the mole mass of substance i. The chemical reaction between the two gases
air and fuel, can now be written:

ma xa  m f x f m a x a  m f Sx f (4.34)

where m a is the mass of air, and m f is the mass of fuel. The mass fraction of the burned
gas arise when the right hand side of (4.34) is normalized. This gives an expression of
the mass fraction of substances in the exhaust gas x b according to:
m a x a  m f Sx f ˆ A~F s λx a Sx f 
x b ˆ λ (4.35)
ma P i x a,i  m f P
i Sx f Srow i ˆ A~F s λ  1

P P
where x a,i 1, and Sx f Srow i 1 because normalized mass flow concentrations are
used and the number of atoms are conserved. To receive the final expression (4.35), the
air/fuel ratio (4.27) is considered in the determination. The stoichiometric air/fuel ratio
ˆ A~F s is calculated according to:

x x x
ma 2 Mf ,CC HH 4  3.5 Mf ,CC 2HH 6  5 Mf ,CC 3HH 8
ˆ A~F s (4.36)
4 2 6 3 8
Š  x a,O 2
mf s M O2
4.3. Combustion 37

where the concentration of oxygen in the exhaust gas in (4.34) is equal to zero and the
calculations are shown in Appendix A.3. Equation (4.35) states that the mass fraction of
the exhaust gas can be expressed only in the scalar variable λ for the two gases air and
fuel.

4.3.2 Chemical Equilibrium

For performance analysis, a good approximation is to assume that the substances pro-
duced by the combustion is in equilibrium. Equilibrium here means that the dissociation
between the species occurs with equal rate, e.g., the number of O 2 molecules dissociated
into two O atoms is the same as the number of two O atoms that are dissociated into O 2
\
molecules, i.e., O 2 2O.
This dissociation rate depends highly on the temperature, and increases with the
increased temperature (Heywood, 1988). To calculate the chemical equilibrium for a
specific exhaust gas at a given combustion temperature, a chemical equilibrium program
can be used. A well known program is the NASA equilibrium program, presented in
Gordon and McBride (1994). Here, the chemical equilibrium program CHEPP developed
in Eriksson (2004) is utilized with some modification. In the original CHEPP version,
hydrocarbons in the form: C a H b OH, together with atmospheric air in the form: ˆO 2 
3.773N 2  are considered as reactants. In the present work, the interface to CHEPP is
modified to handle hydrocarbons, and atmospheric air in the form showed in (4.30).
The modified CHEPP is used to check how well the thermodynamic properties of a
stoichiometric chemical reaction in (4.34) harmonize with a gas in chemical equilibrium.
In Figure 4.2, a demonstration of CHEPP’s ability to calculate the mole concentration
in the exhaust gas for an isooctane fuel C 8 H 18 at three different temperatures is shown.
As the figure indicates, the dominant species for λ A 1 are: nitrogen N 2 , carbon dioxide
CO 2 , water H 2 O, and oxygen molecule O 2 when the temperature is low. For higher
temperatures, the species of nitrogen oxide NO increase in concentration. For λ @ 1, the
shortage of oxygen molecules results in a reduction in carbon monoxide CO.
An introduction of how the equilibrium is calculated will follow in the remaining
part of the section. For a more comprehensive explanation about equilibrium calculation
see, e.g., Heywood (1988); Turns (2006); Öberg (2009). A commonly occurring method
to calculate the chemical equilibrium is to minimize the thermodynamic potential of the
mixture, e.g., the sum of Gibbs free energy in (4.24) for each substance in the mixture,
i.e., G ˆn P i g̃ i n i . For the calculation, the temperature T and pressure p have to be
specified. To find that n which minimizes G ˆn an optimization procedure is performed
together with a number of atom balance constraints. The optimization problem is written:

minimize G ˆn
n
l m
subject to Q Q ã i j n j b̃ i (4.37)
i 1 j 1

nj C 0

where l is the number of unique atoms, m is the number of substances in the mixture
38 Chapter 4. Thermodynamic Concepts

T 1750 K T 2250 K T 2750 K

100 100 100
N2 N2 N2
H2 O CO H2 O CO H2 O CO
101 101 101
O2 H2
102 102 102
H2 CO2 H2 CO2 CO2
NO H O OH
NO
103 103 103
Mole fraction

OH O2
104 H 104 H NO 104

105 OH 105 O 105

O2
106 106 106
NO
H
107 107 107
O2
O
108 108 108
0.5 1 1.5 2 0.5 1 1.5 2 0.5 1 1.5 2
ˆF ~ A ratio ϕ ˆF ~ A ratio ϕ ˆF ~ A ratio ϕ

Figure 4.2: In the figure, the combustion product substances of an isooctane fuel C 8 H 18 at chemical
equilibrium for the temperatures 1750 K, 2250 K, and 2750 K and pressure 1 atm is shown. The
figure is generated in the chemical equilibrium program CHEPP, where the ten most affectable
product species are considered. The species are plotted versus the fuel/air ratio ϕ, where ϕ λ1 .
For lean combustions (ϕ @ 1) at temperature 1750 K the dominated species are: oxygen O 2 ,
carbon dioxide CO 2 , water H 2 O, and nitrogen N 2 . For lean combustions at higher temperatures
the nitrogen oxide NO and carbon monoxide CO substances increase in concentration.
4.3. Combustion 39

which is in equilibrium, and G ˆn i g̃ i n i is the cost function. The first step to set up P
the optimization problem is to select which product substances that should be used in
the exhaust gas. This means that the structure of the concentration vector x̃ b has to be
specified. In the second step, the balance equations of the atom conservation are set up.
In the conservation equations of atoms, l different atoms and m different substances in
the exhaust gas is considered. In mole basis, these constraints are written à n b̃. The
à > R l m matrix describes which (and how many) atoms each combustion substance
consists of. The solution to the optimization problem is the equilibrium gas composition,
which is x̃ b nn when the optimization is performed. In Example 4.2 the setup of the
S S

optimization problem is presented, and the results from the equilibrium calculation are
shown in Figure 4.3.

Example 4.2
A hydrocarbon fuel with concentration x̃ f is combusted with atmospheric air with
concentration x̃ a . These two concentration vectors were introduced in (4.30). The
exhaust gas concentration vector x̃ b is calculated using the chemical equilibrium program
CHEPP and is the solution to the optimization problem presented in (4.37). The product
species in the exhaust gas are chosen to be: hydrogen H, nitrogen N 2 , oxygen O, carbon
monoxide CO, carbon dioxide CO 2 , water H 2 O, oxygen O 2 , nitrogen monoxide NO,
hydroxyl OH, hydrogen H 2 , and argon Ar. These substances are summarized in vector
form:
T
x̃ b Šx̃ H , x̃ N 2 , x̃ O , x̃ C O , x̃ C O 2 , x̃ H 2 O , x̃ O 2 , x̃ N O , x̃ OH , x̃ H 2 , x̃ Ar  . (4.38)

The chemical reaction formula between air and fuel can in this case be written:
Air Fuel
³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ·¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ µ ³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ·¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ µ
n a ˆx̃ Ar Ar  x̃ C O 2 CO 2  x̃ O 2 O 2  . . .  n f ˆx̃ C H 4 CH 4  x̃ C 2 H 6 C 2 H 6  . . . Ð
n 1 H  n 2 N 2  n 3 O  . . .  n 11 Ar (4.39)
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Exhaust Gas

where n a is the number of mole air, n f is the number of mole fuel, and n i , i 1, . . . , 11
depends on the temperature and the pressure. To get the concentration vector x̃ b , n is
normalized. In this case, unique atoms in the reactants are: hydrogen H, nitrogen N,
oxygen O, carbon C, and argon Ar. Thus the coefficients ã i j for the balance equations
can be summarized in the matrix:

H N2 O CO CO 2 H2 O O2 NO OH H2 Ar
H ’ 1 0 0 0 0 2 0 0 1 2 0 “
N – 0 2 0 0 0 0 0 1 0 0 0 —
– —
à O – 0 0 1 1 2 1 2 1 1 0 0 —
– —
C – 0 0 0 1 1 0 0 0 0 0 0 —
Ar ” 0 0 0 0 0 0 0 0 0 0 1 •
40 Chapter 4. Thermodynamic Concepts

and the vector b̃ is constructed according to:

T
b̃ ˆx̃ a , x̃ f  Šn H , n N , n O , n C , n Ar 

where the number of moles in b̃ depends on the amount of air and fuel. The relative
measure between the atoms can be controlled by the λ parameter. In Figure 4.3, the result
of the CHEPP calculation (in mass concentration) is compared with a mass fraction
calculation based on the stoichiometric combustion in (4.34) for the most dominant
substances. The more general equilibrium calculation was shown in Figure 4.2 but
without including argon (Ar) in the mixture.

4.3.3 Comparison of the Heat Capacity Between the Stoichiometric Gas

Description and the Chemical Equilibrium Calculation
The most commonly occurring thermodynamic properties of a gas are, e.g., the enthalpy
h, the entropy s, the internal energy u, the density ρ, the gas constant R, and the heat
capacities c p , cv . An idea is to check how well these properties are described for different
gas description approaches. Here, the behavior of the stoichiometric combustion (4.34)
and the chemical equilibrium calculation (4.37) presented in Example 4.2 is investigated
for the heat capacity c p . In Figure 4.3, the mass concentrations of substances in the two
modeling approaches are shown. The main difference between these two approaches is
the increase in concentration of the nitrogen oxide molecules. The chemical equilibrium
calculation is strongly connected to the increase in temperature which leads to a higher
amount of nitrogen oxide molecules. The change in mass concentration of species affects
the thermodynamic properties mentioned previously, and in Figure 4.4 the heat capacity
is compared for the two approaches. In the figure, three different combustion temper-
atures: T 1000 K, T 1500 K, and T 2000 K are compared for the two modeling
approaches. The figure indicates that the mismatch between the two cases appears to
increase for large temperatures and for temperatures above 2000 K the mismatch is larger
than 8 %. For temperatures below  1500 K, the stoichiometric description agrees with
the chemical equilibrium calculation. In all three cases the results are as expected.

4.3.4 Mixing of Exhaust Gases with Different Lambda

It is important to have the potential to mix two (or more) gases with different air/fuel ratio.
For example, a way to increase the efficiency is to increase the combustion temperature.
To reduce the heat stresses which appear in the first blades of the turbine, compressed
fresh air is injected as a thin film at the surface of the blades to protect the material.
After some time, the injected air and the exhaust gas are mixed which changes the
thermodynamic gas properties of the fluid. Therefore it is important to have a model
library that can handle the mixture between exhaust gases with different air/fuel ratios
and fresh cooling air. The mass concentration x b of the gas, that is a mixture of the two
4.3. Combustion 41

xAr % x O 2 %
1.30 20

1.25
10
1.20

1.15 λ   0 λ  
1 3 5 1 3 5
(a) Ar (b) O 2

x H 2 O % x C O 2 %
20 20

10 10

0 λ   0 λ  
1 3 5 1 3 5
(c) H 2 O (d) CO 2

x N O % x N 2 %
1 76

0.5 73

0 λ   70 λ 

1 3 5 1 3 5
(e) NO (f) N 2

Figure 4.3: Mass fraction of the substances argon (Ar), oxygen (O 2 ), water (H 2 O), carbon dioxide
(CO 2 ), nitrogen oxide (NO), and nitrogen (N 2 ) in the exhaust gas when the combustion is lean,
the temperature is 2000 K, and the pressure is 1 atm. The dashed lines in each figure represent
substances for equilibrium calculations performed in CHEPP, and described in Example 4.2.
Solid lines represent calculations made according to stoichiometric combustion presented in
(4.34). For these two cases, the same air x a , and fuel x f mass concentration vectors are used.
The nitrogen oxide NO is not considered in the stoichiometric combustion, and is not shown
here. The other substances that x̃ b consists of, showed in (4.38) are not showed here since they
are too small. The main difference between these two cases is the appearance, and the increase
of nitrogen oxide in the CHEPP calculation.
42 Chapter 4. Thermodynamic Concepts

c p kJ ~kgK 
1.65
CHEPP
GTLib
1.55

1.45

1.35 2000 K

1.25 1000 K
1500 K
1.15 λ  
1 2 3 4 5

Figure 4.4: In the figure, the heat capacity c p is calculated for the two calculation methods: stoichio-
metric combustion, and chemical equilibrium calculation. For these two methods, the combus-
tion is lean and three temperatures are studied. The temperatures are: T 1000 K, T 1500 K,
and T 2000 K. As the figure indicates, the difference between these two cases appears to
increase for high temperatures. For temperatures above 2000 K, the mismatch between the two
calculation methods is larger than 8 %. For temperatures below  1500 K, the stoichiometric
description agrees with the chemical equilibrium calculation.

fluids 1 and 2, is:

m1 xb ˆλ1   m2 xb ˆλ2 
x b ˆ λ (4.40)
m1  m2
where m i is the mass of fluid i, and λ i is the air/fuel ratio of fluid i.
Solving (4.40) gives an analytic solution of the air/fuel ratio λ in the mixed gas as:
ˆm 1 λ 1 m 2 λ 2   λ 1 λ 2 ˆA~F s ˆm 1  m 2 

λ (4.41)
ˆm 1 λ 2  m 2 λ 1 ˆ A~F s  ˆm 1  m 2 

where the mass concentration vector in (4.35) is used in the determination. The expression
(4.41) is used in the turbine component in the GTLib package where the cooling air is
mixed with the exhaust gas. The masses in (4.41) can directly be translated to masses per
unit time, i.e., mass flows.

4.4 Ideal Gas Model

The gas model is accessible in each thermodynamic component in the gas turbine pack-
age where the gas properties are calculated. The state variables: pressure p, temperature
T, and air/fuel ratio λ are considered as input variables to the gas model. The ther-
modynamic properties which are always calculated by the gas model are: the specific
4.4. Ideal Gas Model 43

internal energy u, the density ρ, the specific enthalpy h, and the gas constant R. The
gas model can be extended to also calculate, e.g., the heat capacity c p , and the entropy s.
To simplify the model representation, the gas model is divided into two parts: (i) an
equation based part, and (ii) a function based part. The basic idea is that the equation
based part consists of fundamental constraints and the function based part is used for
calculations of thermodynamic properties, e.g., typically described by look-up tables.
The same procedure is utilized in the original Modelica Media package. The structure of
the gas model is shown in Figure 4.5 where the dotted arrows represent quantities which

Gas Model
p Equations: u
p ρRT
ρ
h uRT
T Functions: h
xb Required
R
Mass
cp
Fraction
λ Optional s


Figure 4.5: Structure of the gas model.

are not calculated by default. The equation based part relies on the ideal gas law (4.42a),
and the relation between enthalpy and internal energy (4.1) of an ideal gas. The equation
based part is:

m
p RT ρRT (4.42a)
V
h u  pV u RT (4.42b)

where p is the pressure, m is the mass, V is the volume, R is the gas constant, T is
the temperature, ρ is the density, u is the specific internal energy, and h is the specific
enthalpy.
Since the gas is described by a mixture of ideal gas substances, the mass fraction
calculation must be encapsulated by the gas model. The mass fraction calculation is linked
together with the thermodynamic properties of the gas and form the function based
part. The thermodynamic property of each substance in the function based part relies
on the NASA Glenn coefficients McBride et al. (2002). The mass fraction calculation is
based on (4.35) where the mass fraction of air x a , the mass fraction of fuel x f , and the
stoichiometric air/fuel ratio ˆA~F s (ˆA~F s is a function of x a and x f ) are all assumed
to be constant since the ambient conditions are not changed. Thus, those quantities are
not input signals to the gas model and are treated as constant values. When the ambient
conditions change, those constant quantities have to be updated.
44 Chapter 4. Thermodynamic Concepts

4.4.1 Thermodynamics Properties of Frozen Mixtures

In the gas model, a stoichiometric combustion is considered together with an assumption
of a frozen mixture of ideal gases. For a frozen mixture, the gas composition is indepen-
dent of the pressure and temperature, see, e.g., Heywood (1988). Thus, the temperature
dependent part can be separated from the air/fuel ratio dependent part as shown in
(4.43a)–(4.43d) using the mass fraction vector in (4.35). The thermodynamic properties
calculated by the function based part in Figure 4.5 are:

h ˆT, λ Q h i ˆT xb,i ˆλ (4.43a)

i
R ˆ λ Q R i xb, i ˆλ (4.43b)
i
c p ˆT, λ Q c p, i ˆT xb, i ˆλ (4.43c)
i
s o ˆT, λ Q s oi ˆT xb, i ˆλ (4.43d)
i

and the entropy (4.23) for an ideal gas is written:

sˆT, p, λ Q s i ˆT, p i xb,i ˆλ

i

Q s oi ˆT xb,i ˆλ Q R i ln Š ppoi xb, i ˆλ



i i
p
s o ˆT, λ  Rˆ λ ln Š
po
 Q R i ˆλxb,i ˆλ ln Šxb, i ˆλ (4.44)
i

where the partial pressure p i p x b,i is introduced according to Dalton’s law. The index
i denotes the thermodynamic property for the pure substances. For an entropy change
(∆s s 2  s 1 ) where the substances in the mixture do not change the last term in (4.44)
can be removed.

4.5 Energy Conservation of Thermodynamic Systems

The first law of thermodynamics (4.7) states that energy cannot be created or destroyed.
The energy can only be transformed between different states of the fluid. In open ther-
modynamic systems, the transformation is between thermal energy dQ, mechanical
work dW, and intrinsic energy dU of the fluid. The objective with this section is to
derive a relation between differentials of the fluid using the state variables pressure p,
temperature T, and air/fuel ratio λ.

4.5.1 Thermodynamic Differentials dU, dW, and dQ

The thermodynamic differentials dU, dW, and dQ are summarized in this section.
4.5. Energy Conservation of Thermodynamic Systems 45

Internal Energy Differential dU

The internal energy of the gas before and after the mixing occurs can be denoted U 0 and
U ∆ according to:

U0 mu ˆT0 , λ 0   ∆m i uˆTi , λ i 
(4.45)
U∆ ˆm  ∆m i u ˆT∆ , λ ∆ 

where the mass ∆m i is added, the subscript ∆ denotes the mixing properties, and the
subscript i denotes the incoming fluid i. The difference between the two fluid states
can be described by a Taylor series expansion at point ˆ p 0 , T0 , λ 0 . The Taylor series
expansion of U ∆ is:

∂u ∂u
U∆ ˆm  ∆m i Šu ˆT0 , λ 0   ∆T  ∆λ  O ˆ∆2  (4.46)
∂T ∂λ
where O ˆ∆2  captures all the second order, and higher terms. The change in internal
energy ∆U U ∆  U 0 can now be written:

∂u ∂u
∆U mŠ ∆T  ∆λ  ŠuˆT0 , λ 0   uˆTi , λ i ∆m i (4.47)
∂T ∂λ
where second order terms and higher are removed. The definition of the differential dU,
together with (4.47), gives:

∆U ∂u ∂u
dU lim Š  mŠ dT  dλ  ŠuˆT0 , λ 0   uˆTi , λ i dm i (4.48)
t 0 t ∂T ∂λ

Work Energy Differential dW

The gas stream that is flowing into the control volume does work on the thermodynamic
system, thus the differential dW had to be split into the two work contributions:

dW d W̃  pν i dm i (4.49)

where d W̃ is the external mechanical work, and pν i dm i is the work performed by the
mass differential dm i . When no external work is applied to the system the differential
d W̃ 0.

Thermal Energy Differential dQ

The thermal energy differential dQ is assumed to be known and when the container is
perfectly insulated the differential dQ 0.

4.5.2 Energy of the Mixture of Frozen Ideal Gases

The first law of thermodynamics (4.7) can be written together with (4.48) and (4.49):

∂uˆT, λ
mŠcv ˆT, λdT  d λ  uˆT, λdm i dQ  d W̃  h ˆTi , λ i dm i (4.50)
∂λ
46 Chapter 4. Thermodynamic Concepts

where the enthalpy h u  pv for the incoming mass flow and the specific heat capacity
∂u
cv ∂T are introduced. When open systems are studied, it is convenient to consider the
enthalpy since it encapsulates both the internal energy and the mechanical work of the
inflowing masses.

4.6 Control Volume Model

To describe the fluid state in a perfectly mixed container, the differentials that are derived
from the ideal gas law (4.51), and the differentials that are derived from the energy
conservation equation (4.50) is considered. To completely specify the gas properties,
also the mass concentration vector is needed. An idea is to use a chemical equilibrium
program that calculate the concentration of products for a given temperature and pressure.
In the present work, it is assumed that the gas composition is frozen, thus the composition
depends only on the air/fuel ratio.

4.6.1 Differential Form of the Ideal Gas Law

The differential of the ideal gas law (4.42a) is:

V dp RT Q dm i  m TdR  m RdT (4.51)

i

where it is assumed that the size of the control volume is constant, i.e., dV 0. All
differentials, except dR are either requested or available. According to the chain rule, the
differential dR is calculated as:
∂R ∂R
dR ˆ p, T, x b  dp  dT  ˆ©x b R T dx b (4.52)
∂p ∂T

where ©x b is the gradient of the mass concentration vector x b . It is assumed that no

reaction between molecules occurs in the container, i.e., the gas composition is frozen,
thus the differential (4.52) is:
dR ˆ©x b RT dx b (4.53)

Mass Concentration Differential dx b

Since fluids with different mass concentrations can be mixed, an expression of the mass
concentration differential vector dx b is needed. A schematic view of the process, for i
fluids which are flowing into a control volume V , is shown in Figure 4.6. In the figure,
dx b i is the concentration vector, and dm i is the mass of the inflowing fluid i. This gives
an expression of the mass conservation in the perfectly mixed control volume according
to:
d ˆm x b  Q x b i dm i (4.54)
i

where the left hand side describes the mass change of species in the control volume V .
The mass change of the species in V is equal to the sum of the incoming and outgoing
4.6. Control Volume Model 47

V1 V

m 1 , T1 , x b 1
x b 1 dm 1

x b i dm i m, T, x b
Vi
m i , Ti , x b i

Figure 4.6: Conservation of mass and energy.

masses of respective species, i.e., equal to the right hand side of (4.54). The differential
d ˆm x b  can also be written:

d ˆm x b  xb Q dm i  m dx b (4.55)
i

where the chain rule is applied, and the summation of all incoming masses dm i dm i P
is introduced. Combining (4.54) and (4.55), the mass concentration differential vector
dx b for a number of species in a perfectly mixed container is:

dx b Q x̂b m xb dm i
i

(4.56)
i

where m is the total mass in the control volume. Since the gas composition in the
control volume is not affected by the outflowing masses, x̂ b i is used instead of x b i . The
concentration vector x̂ b i can be expressed:

xb i When i is into the mixer ˆdm i A 0

x̂ b i œ (4.57)
xb When i is out from the mixer ˆdm i B 0

where x b i is the mass fraction of the incoming fluid i, and x b is the mass fraction in the
control volume. The differential (4.56) expressed in air/fuel ratio λ can be combined
with the vector (4.35) to get:

dx b Q xb ˆλ̂ i m xb ˆλ dm i

(4.58)
i

where
λi When i is into the mixer ˆdm i A 0
λ̂ i œ (4.59)
λ When i is out from the mixer ˆdm i B 0

according to the flow directions.

48 Chapter 4. Thermodynamic Concepts

4.6.2 Lambda Concentration Differential d λ

The differential of the mass fraction vector can be written:
dx b
dx b dλ (4.60)
dλ
and the derivative of (4.35) with respect to lambda is:
dx b a  b ˆA~F s
(4.61)
dλ ˆˆ A~F s λ  12

where a ˆA~F s x a and b S x f are introduced for easier notation. The differential
of the mass fraction, which was derived in (4.58), is combined with (4.35), (4.60), and
(4.61) to get:
< =
1 @ a λ̂ i  b aλ  b AA a  b ˆA~F s
m
Q @
@
ˆ A~F s λ̂ i  1

Adm i
ˆ A~F s λ  1 A ˆˆ A~F s λ  12
dλ (4.62)
i @
> ?
which can be solved according to dλ:
< =
1
Q @@@ ˆˆAA~~FFsλ̂λ 1
@  A
dλ Š λ̂ i  λAAdm i (4.63)
m i > s i  1 A
?
where
λi When i is into the mixer ˆdm i A 0
λ̂ i œ (4.64)
λ When i is out from the mixer ˆdm i B 0
as presented earlier. In this case, the mass fraction differential which is a vector can
be replaced with the lambda differential which is a scalar. It can also be noted that the
stoichiometric matrix S, the mass concentration vector of air x a and fuel x f do not appear
in (4.63). The differential depends only on the stoichiometric air/fuel ratio ˆA~F s .

4.6.3 Partial Derivatives of Gas Property Functions

In the state equations, which will be presented in (4.67), the partial derivatives with
respect to lambda for the thermodynamic quantities u and R appear. These partial
derivatives can be expressed using the derivative of the mass concentration vector (4.61)
which can be rewritten:
dx b ˆ A~F s
Šx a  xb  (4.65)
dλ ˆ A~F s λ  1

using (4.35). The partial derivatives of internal energy u, and specific gas constant R of
the frozen mixture are written:
∂u
∂λ
Q u i dxd λb,i ˆ A~F s
ˆ A~F s λ  1
Šu a ˆTa   u ˆT, 덏 (4.66a)
i
∂R
∂λ
Q R i dxdλb, i ˆ A~F s
ˆ A~F s λ  1
ŠR a  Rˆ 덏 (4.66b)
i
4.6. Control Volume Model 49

where ua is the internal energy, Ta is the temperature, and Ra is the gas constant of the
ambient air. When the mass fraction of the ambient air changes, these constants need to
be updated.

4.6.4 State Equations

The differentials of the ideal law (4.51), the energy conservation equation (4.50), and the
change in air/fuel ratio can be summarized in the following state equation differentials:

∂R
V dp  m T d λ  m R dT R T dm
∂λ
∂u
mŠcv dT  dλ dE  u dm (4.67)
∂λ
1
dλ dΛ
m
where

dm Q dm i (4.68a)
i
dE Q h i dm i (4.68b)
i
< =
@ ˆ A~F s λ  1 A
dΛ Q @
@
ˆ A~F s λ̂ i  1
Š λ̂ i  λAAdm i (4.68c)
i @
>
A
?

are introduced. Differential dm is the total change of mass in the volume, dE is the
energy contribution of the incoming fluids, and dΛ is the air/fuel ratio contribution. The
air/fuel ratio, as previously defined, is:

λi When i is into the mixer ˆdm i A 0

λ̂ i œ
λ When i is out from the mixer ˆdm i B 0

according to the flow directions. In (4.50), dQ and d W̃ are assumed to be zero, i.e.,
a perfectly insulated container with no applied external work. State variables for the
control volume are chosen as pressure p, temperature T and air/fuel ratio λ. In Figure 4.7
the exchange of mass, energy, and lambda according to (4.68) are showed.

4.6.5 Variation in Ambient Absolute Humidity

To handle variation in absolute humidity, the mass fraction vector of ambient air x a
in (4.35) and (4.65) needs to be updated. The absolute humidity is calculated using the
moist air model presented in Buck (1981). The humidity affects the stoichiometric air/fuel
ratio ˆA~F s , the internal energy of air ua ˆTa , and the gas constant of air Ra which
appear in (4.35), (4.66), (4.68), and (5.26). Since these quantities are defined as constants,
the gas properties are updated instantaneous in all thermodynamic components. This
results in transients when the absolute humidity of the ambient air changes. For the gas
50 Chapter 4. Thermodynamic Concepts

V1
V3
p1
T1 dE 1 ˆT1 , dm 1 
λ1 R1 dΛ 1 ˆ λ 1 , λ 3  p3
dm 1 @ 0 T3
dE 1 ˆT1 , dm 1  dm 1 A 0 λ3
dΛ 1 ˆ λ 3 , λ 1  0

dE 2 ˆT2 , dm 2  dm 2 A 0
V2
dΛ 2 ˆ λ 2 , λ 3 

R2 dE 2 ˆT2 , dm 2 
dΛ 2 ˆ λ 3 , λ 2  0
dm 2 @ 0
p2
T2
λ2

Figure 4.7: In the figure, three control volumes, together with two flow restrictions are presented.
The mass flows through the restrictions have the directions from volumes V1 , and V2 to volume
V3 . The flow variables through the restrictions are the mass flow, the enthalpy flow, and the
lambda flow according to (4.68).
4.7. Conclusion 51

turbine application, this phenomenon is negligible because of the high throughput speed
compared to the changes in the ambient conditions. The reason is that the ratio between
the control volumes and the mass flow is small, i.e., the gas exchange is fast in the control
volumes.

4.7 Conclusion
In the chapter, fundamental thermodynamic concepts are presented that can be used to
describe a gas medium used in a combustion process. In the present study, these concepts
are implemented in the gas turbine package – GTLib used to model an industrial gas
turbine, which will be presented further in Chapter 5. The central part of the chapter is the
development of a control volume model, where an exhaust gas is described using the three
state variables: (i) pressure p, (ii) temperature T, (iii) and air/fuel ratio λ. The ambient
air, which is used in the combustion, can handle different amounts of humidity through
an adjustment of the mass fraction of water. The framework handles hydrocarbon fuels
with a specified number of individual substances. To model atmospheric air in a control
volume, the λ variable needs to be specified large, i.e., preferably infinitely large, but a
large number is sufficient in practice. When the air/fuel ratio λ is used as a state variable
instead of the mass fraction of substances the total number of state variables is decreased.
The advantage with using the presented thermodynamic concept is the ability to
integrate the combustion process into the model. During the combustion, the mass
concentrations of species in the incoming air are changed, since species of carbon diox-
ide and water are produced under the consumption of oxygen. The number of these
constructed species depends on the hydrocarbon fuel where the reaction formulas are
represented by the stoichiometric matrix. The mass fraction of species in the exhaust gas
is specified with the state variable air/fuel ratio λ. For this procedure, the assumptions
are: (i) the combustion is lean (i.e., λ A 1), (ii) the exhaust gas is frozen in composition
(i.e., no dissociation effects between the species occur), and (iii) no pure substances can
be introduced throughout the gas path. The first assumption is not a problem in a gas
turbine application, since an excess of oxygen is available. The second assumption can be
a problem if the flame temperature is too high. It is shown that for temperatures above
1500 K, the thermodynamic properties of the fluid start to change. For a temperature of
2000 K, the error in heat capacity c p is about 8 % against a calculation where dissociation
effects are considered. For the temperature of 1500 K, the error is lower than 2 %. The
third assumption states that no pure substances can be introduced into the exhaust gas
since this will destroy the air/fuel ratio. This means that, e.g., water steam cannot be
injected into the gas media throughout the gas path.
 Chapter 5

GTLib – Thermodynamic Gas Turbine Modeling

Package

The objective of the chapter is to introduce the thermodynamic gas turbine modeling
package GTLib, which is used to construct a physical based gas turbine model. The
gas turbine model constructed in GTLib can be used for: (i) performance analysis,
(ii) control issues, and (iii) as a base for further investigations when an FDI-system is
constructed. When a model based diagnosis approach is used for the FDI-system design,
it is crucial to have diagnosis statements based on good models. Roughly speaking,
diagnosis tests based on good models give better diagnosis performance, i.e., increase the
fault detectability and reduce the false alarm probabilities than when models with lower
accuracy are used. Therefore, a good idea is to have diagnosis tests in the FDI-system
based on the knowledge from the model used for performance analysis.
An important part of GTLib is the gas model which is used in the components along
the engine’s gas path to calculate the thermodynamic properties of the fluid. Using GTLib,
the two main benefits are achieved: (i) an overall reduction in the gas turbine model
equations compared to the reference gas turbine model, and (ii) the ability to achieve
a systematic procedure to automatically construct test quantities when designing the
FDI-system.

Outline of the Chapter

In Section 5.1, an overview of a typical appearance of the performance characteristics
of a gas turbine is presented. The performance characteristics for the compressor and
the turbine utilize the concept of corrected parameters which will be explained in the
section. In Section 5.2, a small example that emphasize the difference in results between
the gas description in GTLib and SITLib (developed by SIT) is shown. In the control
volume example, a transient in the ambient air composition of substances is evaluated
for the two cases. In Section 5.3, the implementation of GTLib at a high level is presented.
Here, the gas model and its implemented thermodynamic functions together with the

53
54 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

other gas turbine components are shown. In Section 5.4, an overview of the designed gas
turbine model is presented.

5.1 Gas Turbine Performance Characteristics

A simple approach to describe the performance of a gas turbine component, e.g., the
compressor, is to assume that the isentropic efficiency ηis is constant. In, e.g., Hadik
(1990), the isentropic efficiency is considered to be ηis 0.87. A more sophisticated
method to describe the performance of gas turbine components is to use look-up tables
of corrected parameters. The advantage with using corrected parameters in the look-up
tables is the ability to describe the performance in other operating conditions than
at the measured reference conditions. The corrected parameters are collected in non-
dimensional groups which have a background in dimensional analysis (Buckingham,
1914). A relation between the non-dimensional groups is presented in, e.g., Dixon
and Hall (2010); Saravanamuttoo et al. (2001); Heywood (1988); Volponi (1999), and is
reviewed here: »
mflow T01 Rγ nD
Π, ηis  fŠ ,»  (5.1)
¯ D p 01
2
T01 Rγ
η ‡is ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
m flow ‡
n‡

where n is the shaft speed, D is the impeller diameter, Π is the pressure ratio, mflow is
the mass flow of air, R is the specific gas constant, γ is the heat capacity ratio, T is the
temperature, p is the pressure, and ηis is the isentropic efficiency. The subscript in, e.g.,
T01 denotes stagnation temperature at the inlet of the component and the subscript will
be removed in the following section to get simpler notation. The function arguments are
corrected parameters and will be denoted with ˆ‡ in the sequel. For a specific gas turbine,
the impeller diameter is fixed so D in (5.1) can be neglected. Normalized quantities of
the corrected parameters can be constructed by multiplication of a non-dimensional
constant to get:
¿
‡ mflow pref Á
Á
À
T R γ
mflow,norm 100 (5.2a)
mflow,ref p Tref Rref γref
¿
‡ n Á T R γ
À ref ref ref
Á
nnorm (5.2b)
nref T R γ
ηis
η‡is,norm (5.2c)
ηis,ref
¿
CT Á R γ
C‡T ,norm
Á
À (5.2d)
C T ,ref Rref γref

where ˆref denotes the value at the reference state. In the last expression, the flow
º
m T
capacity notation is introduced where C T flow
p
. The reference parameters must be
given together with the look-up tables which are described herein.
5.1. Gas Turbine Performance Characteristics 55

5.1.1 Compressor Map

The performance of the compressor is described by look-up tables of corrected and
normalized parameters. The variables given by the map are the normalized mass flow of
air, and the isentropic efficiency according to:

‡
mflow,norm ‡ , α
f 1,η ˆΠ, nnorm (5.3a)
η‡ is,norm f 1,Γ ˆΠ, m
‡
flow,norm , α  (5.3b)

where the function in (5.1) is extended with the angle α of the inlet guide vanes (IGVs).
The IGVs are used to change the inlet angle of the mass flow to obtain as high
efficiency as possible regardless of the load. Since the machine is 2-shafted, the IGVs
depend directly on the shaft speed of the gas generator and are therefore built into
the look-up tables. Thus, they can be removed explicitly from (5.3). With IGVs, the
compressor can also cover a wider operating range since the surge line of the compressor
is moved. In Figure 5.1, the isentropic efficiency and the normalized speed lines are
plotted versus the pressure ratio and the corrected normalized mass flow. In the figure,
also the normalized speed lines are plotted versus the normalized mass flow.

Π Surge
 
η‡is,norm 1.00
Line
22 η‡is,norm 0.99
η‡is,norm 0.98
η‡is,norm 0.90
18

14

10
1.0
0.9
6 0.8
0.7
Choke 0.6 ‡
nnorm
2 Line
50 60 70 80 90 100
‡
mflow,norm

Figure 5.1: In the figure, a typical appearance of the performance characteristics of a compressor is
shown. In the compressor map, the efficiency η‡is,norm and the normalized mass flow mflow,norm
‡
‡
are viewed for different pressure ratios Π and normalized speeds nnorm . The surge line and the
choke line are also shown in the figure.
56 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

5.1.2 Turbine Map

The performance of the turbine is described in the same way as the compressor, i.e., using
of normalized variables and look-up tables. Here, the calculated variables are the turbine
flow capacity C T and the isentropic efficiency η i s according to:

‡
CT,norm ‡ »nref 
f i ,η ˆΠ, nnorm (5.4a)
Tref
n
η‡is,norm ‡ » ref 
f i ,Γ ˆΠ, nnorm (5.4b)
Tref

where Tref and nref from (5.2b) are neglected. Subscript i 2, 3 represents the compressor-
turbine and the power-turbine respectively. In Figure 5.2, an example of a turbine map
is viewed. In the figure, typical appearances of corrected and normalized isentropic
efficiency, and turbine flow capacity are plotted versus the pressure ratio.

η‡is,norm C T‡ ,norm
1 1

0.99

0.98
0.99
0.97 ‡
nnorm 170
‡
nnorm 175
‡
nnorm 180
0.96 ‡
nnorm 185
‡
nnorm 190
0.98 0.95
2.5 3 3.5 2.5 3 3.5
Π   Π  

(a) Turbine efficiency. (b) Turbine flow capacity.

Figure 5.2: In the figure, a typical appearance of the performance characteristics of a turbine is
shown. In the turbine maps, the efficiency η‡is,norm and the normalized flow capacity CT,norm
‡ are
‡
calculated for different pressure ratios Π and normalized speeds nnorm .
5.2. Variation in Ambient Absolute Humidity 57

5.2 Variation in Ambient Absolute Humidity

To achieve high model accuracy, one factor that can affect the performance estimations
is the humidity in the ambient air. This phenomenon is often neglected since it is
assumed that the change in absolute humidity does not vary significantly in an industrial
gas turbine application. In Mathioudakis and Tsalavoutas (2002), an investigation is
performed where it is shown how humidity influences the performance analysis for
an industrial gas turbine. In the case study, it is argued that a large change in absolute
humidity affect the performance analysis negative which can result in a case when a
not correct diagnosis statement is taken. When experimental data is considered, a large
deviation (on daily basis which shows a variation of about 2 per cent points) is present for
the flow capacity of the compressor. When the compensation for the absolute humidity
is considered in the evaluation, the undesirable daily variations in the flow capacity
deviation are significantly reduced.
When the absolute humidity in the ambient air changes, also the constants which
depend on the air, e.g., (i) the mass fraction vector of air x a (ii) the air/fuel ratio ˆA~F s
and (iii) and thermodynamic properties such as u a and R a need to be updated. These
constants affect, e.g., the mass fraction of exhaust gas in (4.35), and partial derivatives in
(4.66). Thus, when a compensation of the absolute humidity is desired the thermody-
namic constants must be updated to maintain the correct exhaust gas mixture given by
the air/fuel state variable.
For models constructed in GTLib, all the thermodynamic constants are updated
immediately in all gas turbine components. This results in an error during transients in
the ambient air composition which Example 5.1 should symbolize. For the gas turbine
application, this phenomenon is negligible because of the high throughput speed com-
pared to the changes in the ambient conditions. The ratio between the control volumes
and the mass flow is small, i.e., the gas exchange is fast in the control volumes.

Example 5.1
Before the gas turbine performance model is presented, a good idea is to introduce a
small simulation example that shows the main difference in gas properties between a
thermodynamic system which uses the GTLib package and a system which uses SITLib
package (Modelica Media). In Figure 5.3, such system is constructed (for the two cases)
which consists of: (i) a source S 1 , (ii) a sink S 2 , (iii) a control volume V10 , and (iv) two
pressure losses d p 1 and d p 2 . The same type of models is used in the two simulation
cases, and a step is injected in the ambient conditions such as temperature T and relative
humidity φ. In the example, it is interesting to compare how the gas properties in the
volume V10 are affected when the temperature and relative humidity have changed. To
show this phenomenon, it is assumed that the incoming volume mass is much less then
the available control volume mass in V10 , i.e., it takes long time to change all the mass
in the control volume. In Figure 5.4, the temperature T, the pressure p, the specific gas
constant R, the specific enthalpy h, the density d, and the relative humidity φ of the
control volume V10 are viewed for the two simulation cases.
In the figure, it can be seen that during transients in ambient conditions, gas proper-
58 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

T
p RH

PP V10 P
dp1 dp2
S1 S2

Figure 5.3: In the figure, a thermodynamic system with a control volume V10 , a gas source S 1 , a gas
sink S 2 , and two pressure losses d p 1 , d p 2 is presented. The input signals to the system are the
pressure p, the temperature T, and the relative humidity φ of the ambient air. It is assumed that
the pressure in the gas source is higher than in the gas sink which gives a mass flow direction
to the right in the figure.

ties such as specific gas constant, specific enthalpy, and density change instantaneously
in GTLib, except for the temperatures in subfigure 5.4(a). This is because ˆA~F s and x a
in (4.35) are changed directly according to the ambient conditions, so they are updated
simultaneously in all components in the model. When the transients have declined, the
gas properties converge to the same values for the two simulation cases.
In the example, the incoming mass is much less than the mass in the control volume
which can be seen as an extreme case. For the real gas turbine application, the throughput
speed for the gas turbine is high so the error due to transients in ambient conditions is
not a problem. The ratio between the control volumes and the mass flow is small.

5.3 GTLib Components

The developed gas turbine package GTLib can be used when a gas turbine model is
constructed. The advantage with GTLib is the ability to build up a model which can be
used for performance analysis and as a base when diagnosis test quantities used in an
FDI-system are constructed. An important part of GTLib is the gas model (Section 5.3.3)
where the thermodynamic relationships presented in Chapter 4 are implemented. These
relationships are then introduced in the control volume model presented in Section 5.3.4.
In the following subsections, the components of: the compressor, the turbine, and the
combustor are presented.

5.3.1 Global Environment Model

To simulate the gas turbine model, a global environment component (GEC) needs to
be defined. In the GEC, the mass fractions of the ambient air x a , the hydrocarbon
fuel x f , and the stoichiometric matrix (4.33) are specified. When these parameters are
specified, the stoichiometric air/fuel ratio ˆA~F s can be calculated according to (4.36).
All thermodynamic gas turbine components use the constants defined in the GEC.
The gas model in its original design can only handle fuel and air gases with a fix
concentration of substances. To imitate more realistic environment conditions, a moist
5.3. GTLib Components 59

T K p kPa
310
106.8 SITLib
GTLib
300
Amb 106.4
SITLib
290 106.0
GTLib
t t
0 1 2 0 1 2
(a) Temperature (b) Pressure

R J ~ˆkgK  h M J ~kg 
295 -0.4
294 SITLib
GTLib
293
-0.6
292
291 SITLib
GTLib
290 t -0.8 t
0 1 2 0 1 2
(c) Gas constant (d) Enthalpy

ρ kg ~m 3  RH  
1.30
SITLib 80
1.25 GTLib
60
1.20
40 Amb
1.15 t t
0 1 2 0 1 2
(e) Density (f) Relative humidity

Figure 5.4: In Figure (a), a step in temperature T and relative humidity RH of the ambient air
is introduced at time 0, 1, and 2. The experiment is performed for the two test cases where
GTLib (dashed line) and SITLib (solid line) are used in the model presented in Figure 5.3. In
the figures, the thermodynamic properties of the gas in the control volume V10 are shown. The
gas properties in all figures mismatch during the transient but the effect on the temperature is
very small. The mismatch during a transient depends on an instantaneous change in the gas
properties in GTLib, e.g., the gas constant R goes from  294 to 291 immediately. In the real gas
turbine application, this phenomenon is not a problem since the mass flow is large compared
to the size of the control volumes.
60 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

air model (Buck, 1981) is introduced when the mass fraction vector x a of the ambient air
is calculated. When the ambient air vector is updated, all constants where the ambient air
properties are used need also to be updated, e.g., in (4.35), (4.66), and (5.33). With this
procedure, the gas properties are updated immediately in all gas turbine components.
This results in an error during transients in the ambient air composition which Exam-
ple 5.1 symbolizes. For the gas turbine application, this phenomenon is negligible because
of the high throughput speed compared to the changes in the ambient conditions. The
ratio between the control volumes and the mass flow is small, i.e., the gas exchange is
fast in the control volumes.

Moist Air Model

To describe the amount of water steam in the ambient air, the moist air model presented
in Buck (1981) is adopted here. The amount of water is then used to describe the whole
x a vector and implemented in the function humidAirCalc(p 0 , T0 , RH) given the input
variables: pressure p 0 , temperature T0 , and relative humidity RH of the ambient air.
In the ambient gas model, the saturation pressure of water is described by Buck (1981).
This expression is used in the meteorological context and has high accuracy in the region
of 80 to 50 o C. The saturation pressure of water is described by:
17.502T0
p H 2 O s 6.1121 ‰1.0007  3.46 103 pŽ exp ˆ  (5.5)
240.97  T0
ˆ

where the ambient temperature T0 is expressed in Celsius and the absolute pressure p is
expressed in bar. The saturation pressure of water p H 2 O s is expressed in hectopascal.
ˆ 

The relative humidity φ (which is known) is defined:

pH2 O
φ 100 (5.6)
p H2 O s
ˆ 

Thus, it is possible to calculate the partial pressure p H 2 O of water vapor using (5.5). Here,
it is assumed that the moist air consists of dry air and water steam, so the partial pressure
of dry air p air is simply equal to the difference in atmospheric pressure p and the partial
pressure of water vapor p H 2 O , i.e., p p H 2 O  p air . This, together with ideal gas law
(4.42a) gives an expression for the mass fraction of water according to:
p H 2 O Rair
xH2 O (5.7)
p H 2 O Rair  pair R H 2 O
Since the mass fraction of substances in the dry air is known, the mass fraction of
the moist air is also determined. When the moist air medium is known, it is possible
to calculate thermal properties such as enthalpy and heat capacities as a function of
temperature and air/fuel ratio throughout the gas path. It can be noted that a change
in the absolute humidity in the ambient air affects the stoichiometric air/fuel ratio for a
given hydrocarbon fuel. How the amount of water depends on the ambient condition
such as relative humidity and temperature can be seen in Table 5.1.
It can be seen in the table that the amount of water in the ambient air increases
drastically with temperature and humidity. In the medium model, the ambient conditions
5.3. GTLib Components 61

Table 5.1: Mass of water vapour, in gram, for 1 kg moist air at datum pressure, at temperature T,
and relative humidity φ.

T ˆC o  φ 40 % φ 60 % φ 80 %
15 4.21 6.33 8.45
20 5.78 8.69 11.61
25 7.85 11.80 15.78
30 10.53 15.85 21.20
35 13.99 21.08 28.22

affect the gas properties immediately in every gas turbine component. This depends on
the fact that the ambient variables are declared as global and all medium models use
these global variables when the mass fraction of species in the ambient air is calculated.
The ambient air concentration is then used when the actual gas properties are calculated.

5.3.2 Connectors
In Modelica, information between components is shared through connection points that
are called connectors. In a connector, there are basically two types of variables which are
defined either as a flow or a non-flow variable. In a connection point, flow variables are
summed to zero, and non-flow variables are set equal.
With this approach, the flow variables are identified from (4.68) and Figure 4.7 to be:
dm, dE, and dΛ. The summation in (4.68) is performed for the number of flows in each
connector which is the same as the number of connected components. In general, each
non-flow variable is specified in the control volume and each flow variable is calculated
in the restrictions between the control volumes. The considered variables, which are
used in the connectors are summarized in Table 5.2.

Table 5.2: In the table, the connector variables used in the GTLib package are presented. These
variables are either defined as a flow or a non-flow variable. The flow variables are summed to
zero, and the non-flow variables are set equal in the connector points.

Type Variable Description

non-flow p pressure
flow dm mass flow
non-flow h enthalpy
flow dE enthalpy flow
non-flow λ air/fuel ratio
flow dΛ air/fuel ratio flow
62 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

5.3.3 Gas Model

The gas model (shown in Figure 4.5) is accessible in each component in the GTLib
package where the state of the fluid is calculated. Given the state variables: pressure
p, temperature T, and air/fuel ratio λ all the other thermodynamic properties can be
calculated, e.g., the enthalpy h and the heat capacity c p . The gas model consists of two
parts: (i) an equation based part, and (ii) a function based part. The equation based
part consists of the fundamental equations for an ideal gas and the function based part
is used for calculation of thermodynamic properties, typically described by look-up
tables which are provided by the Modelica standard library. The tabulated data is the
well known NASA polynomial coefficients summarized in McBride et al. (2002).
In Modelica, flexible models can be defined due to the object oriented nature of
the language. Therefore, the gas model is divided into: (i) a mandatory part, and
(ii) an optional part. In the mandatory part, only the most important thermodynamic
quantities are calculated given the state variables. The mandatory part consists of the
thermodynamic relations:

p ρRT (5.8a)
u hRT (5.8b)
h hˆT, λ (5.8c)
R R ˆ λ (5.8d)

where the mandatory quantities: enthalpy h, gas constant R, internal energy u, and
density ρ are calculated using the equations in (5.8). The mandatory part can be extended
with a number of predefined functions which take the gas state variables as input argu-
ments. The optional functions which can be utilized in the gas model are described in
the remaining part of the present section.

Mass Fraction Calculation in Exhaust Gas

The function lambda2mass(λ) makes a conversion between the air/fuel ratio λ and the
mass fraction vector x b of the exhaust gas. This function is central in the gas model since
it is called everywhere when the gas properties are calculated. The function evaluate
(4.35) and is reviewed here:

ˆ A~F s λx a  Sx f
x b ˆ λ
ˆ A~F s λ  1

where ˆA~F s is the stoichiometric air/fuel ratio, S is the stoichiometric matrix (e.g.,
(4.33)), x a is the mass concentration of the ambient air, and x f is the mass concentration
of the fuel. These parameters are either defined as constants or calculated in the global
environment component (Section 5.3.1). When the global environment component is
utilized it is possible to update the thermodynamic properties of the ambient air during
the simulation.
5.3. GTLib Components 63

Enthalpy
The enthalpy function hˆT, λ makes the calculation according to (4.43a) and is reviewed
here:
h ˆT, λ Q h i ˆT xb,i ˆλ (5.9)
i

where h i represents the standardized enthalpy (calculated according to the NASA poly-
nomials and given in (4.19)) for substance i for a given temperature. The input variables
to the function are the gas temperature T, and the air/fuel ratio λ of the gas. The output
of the function is the enthalpy of the gas mixture.

Entropy
The entropy function sˆ p, T, λ is calculated based on the standardized entropy s oi (4.21)
of an ideal gas for each substance i. The entropy s i of an ideal gas is then calculated
according to (4.23) and is reviewed here:

p
s i ˆT, p s oi ˆT   R i ln Š  (5.10)
po

where the standardized entropy s oi ˆT  and pressure p o is given for the datum state. The
entropy for the mixture is then calculated using a summation of the substances weighted
with the mass fraction of each substance:

sˆT, p, λ Q s i ˆT, p i xb, i ˆλ (5.11)

i

which can be simplified according to the derivation in (4.44). The data that is included
in the absolute entropy vector is calculated according to the NASA polynomials. The
input variables to the function are the pressure p, the gas temperature T, and the air/fuel
ratio λ of the exhaust gas. The output of the function is the entropy of the gas mixture.

Isentropic Temperature
For an ideal compression or expansion thermodynamic process the entropy in (5.11) is
constant (see Figure 2.1). To calculate, e.g., the isentropic temperature T2s , the entropy
function should be inverted. The function T_sˆ p, s, λ calculates the temperature T
of an isentropic compression (or expansion process) for the input variables pressure
p, entropy s, and air/fuel ratio λ. The function invert the entropy expression in (5.11)
numerically using a number of iterations until:

s 1  s 2 ˆ p 2 , T2s , λ 2  @ ε (5.12)

is small enough for the iteration temperature parameter T2s when the isentropic process
goes from state 1 to 2s.
64 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

Specific Gas Constant

The function R(λ calculates the specific gas constant for the exhaust gas. For each sub-
stance i of the gas, the specific gas constant R i (in mass basis) is tabulated in the standard
Modelica package. The gas constant is simply calculated according to R i R̃~ M i , where
M i is the mole mass of the substance i and R̃ is the universal gas constant. The specific
gas constant of the gas mixture is calculated according to (4.43b) and is reviewed here:

R ˆ λ Q R i xb,i ˆλ (5.13)

i

where x b, i is the mass fraction of substance i.

Heat Capacity with Constant Pressure

The functions c_pˆT, λ calculate the heat capacity (for a constant pressure process) of
the exhaust gas. The input variables to the function are the gas temperature T and the
air/fuel ratio λ. The heat capacity is calculated according to (4.43c) and reviewed here:

c p ˆT, λ Q c p, i ˆT xb, i ˆλ (5.14)

i

where x b, i is the mass fraction of substance i.

Heat Capacity with Constant Volume

The functions c_vˆT, λ calculate the heat capacity (for constant volume process) of
the exhaust gas. The input variables to the function are the gas temperature T and the
air/fuel ratio λ. Since the ideal gas assumption is made, the heat capacity is calculated
according to:
cv ˆT, λ c p ˆT, λ  Rˆ λ (5.15)

where R is the specific gas constant.

Isentropic Exponent
The function gammaˆT, λ calculates the isentropic exponent γ. The input variables to
the function are the gas temperature T, and the air/fuel ratio λ. The isentropic exponent
is the ratio between c p and cv which is calculated according to:

c p ˆT, λ
γ (5.16)
cv ˆT, λ

where heat capacities functions c_pˆT, λ and c_vˆT, λ are called.

5.3. GTLib Components 65

5.3.4 Control Volume

The governing state equations of the control volume component are:

dp dT ∂R dλ
V dmRT  mˆR  T  (5.17a)
dt dt ∂λ dt
dT ∂u dλ
mcv dE  udm  m (5.17b)
dt ∂λ dt
dλ
m dΛ (5.17c)
dt
where the flow quantities are defined as:

dm Q dm i (5.18a)
i
dE Q dE i Q h i dm i (5.18b)
i i
< =
Q dΛ i Q @@@ ˆˆAA~~FFsλ̂λ 1
@  A
dΛ Š λ̂ i  λAAdm i (5.18c)
i i > s i  1 A
?

and the calculations of dE i and dΛ i are performed by the pressure loss component that
connects the control volumes according to Figure 4.7. The summations in (5.18) are
performed in the control volume. The air/fuel ratio λ̂ i is:

λi When i is into the control volume ˆdm i A 0

λ̂ i œ
λ When i is out from the control volume ˆdm i B 0

The partial derivatives (4.66), according to the air/fuel ratio are calculated as:

∂u ˆ A~F s
ˆua  u (5.19a)
∂λ ˆ A~F s λ  1
∂R ˆ A~F s
ˆRa  R (5.19b)
∂λ ˆ A~F s λ  1

where the derivation of the partial derivative of the mass fraction function (4.65) is used
together with (4.35). The subscript a represents the property of the ambient air. The total
mass m in the control volume (with volume V ) is described by:

m Vρ (5.20)

where ρ is the density which is defined in the gas model in Section 5.3.3.

5.3.5 Compressor
The compressor component consists of: (i) a mechanical, and (ii) a thermodynamic part
since the energy is transformed from mechanical to thermodynamic energy through a
compression. The compressor has ports where cooling air is tapped, used to decrease
66 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

thermal stresses (caused by heat) in the first blade of the turbine. The ports used for
cooling air are placed at enthalpy ratio r h i > 0, 1, where r h i 0 represents the low
pressure side and r h i 1 represents the high pressure side. During the compression phase,
the temperature T and pressure p of the gas increase. For an isentropic compression
the entropy is constant (shown in, e.g., Figure 2.1). Thus, when the compression is ideal
it is possible to calculate the discharge temperature (or enthalpy) at the output of the
compressor. This is done according to the previous defined function T_s and can be
written:
T2s Ts ˆ p 2 , s 2 , λ 1  (5.21)
When the isentropic temperature T2s increase, also the enthalpy increase. Thus the
change in enthalpy ∆h can be written:
h ˆT2s , λ 1   hˆT1 , λ 1 
∆h (5.22)
ηis
where ηis is the isentropic efficiency defined by a look-up table (see, e.g., Figure 5.1). The
air/fuel ratio is not changed during the compression, thus is λ 1 λ 2 . The thermodynamic
power Pthermo can be written:
Pthermo
Q ˆ1  r h i mflow,r h i ∆h (5.23)
i

where r h i > 0, 1 is the enthalpy ratio, index r h i denotes the port position, and the
summation is over all ports. A mass flow mflow,r h i which leaves the compressor has a
negative sign. The mechanical power Pmech is:
Pthermo
Pmech (5.24a)
ηm
dφ
Pmech
dt i
Q
τi (5.24b)

where η m is the mechanical friction constant, φ is the shaft angle, and τ i is the applied
torque. The mass and the energy balances are written:
0 Q mflow,r hi
(5.25a)
i

0 Q Hflow,r hi

Q r h mflow,r
i hi
∆h (5.25b)
i i

The performance parameters such as isentropic efficiency ηis and incoming mass
flow mflow are calculated according to the look-up tables defined in (5.3), together with
the normalized corrected parameters in (5.2). The normalized corrected parameters are
all calculated at the low pressure side of the compressor.

5.3.6 Turbine
The turbine component is analogue with the compressor component, except that no cool-
ing air is extracted from the gas expansion. Instead, compressed air from the compressor
5.3. GTLib Components 67

is injected in the first turbine blades to protect the material from high temperature
stresses. Since the look-up tables are valid for the hot gas (before cooling air is injected)
it is necessary to mix the hot and cool gases after the characteristic calculations are made.
Since exhaust gases and cooling air have different air/fuel ratio the mixed gas has the
air/fuel ratio λmix according to:
ˆ λinj mflow,inj λe mflow,e   λinj λe ˆA~F s ˆmflow,inj  mflow,e 

λmix (5.26)
ˆ λe mflow,inj  λinj mflow,e ˆ A~F s  ˆmflow,inj  mflow,e 

where subscript inj denotes the compressed air, and subscript e denotes the hot exhaust
gas. Equation (5.26) is calculated according to (4.41). When the air/fuel ratio of the
mixture is determined the enthalpy change ∆h is:
∆h h ˆT4s , λmix   hmix ηis (5.27)
where hmix is the enthalpy of the mixture, and T4s is the isentropic temperature after the
gas expansion. The mass flow of the exhaust gas mflow,e is calculated:
º
Te
CT mflow,e (5.28)
pe
where C T (and ηis ) are calculated according to the look-up tables defined in (5.4), together
with the normalized corrected parameters in (5.2). The thermodynamic power Pthermo is:
Pthermo  mflow,mix ∆h (5.29)
where mflow,mix mflow,inj  mflow,e . The mechanical power Pmech is calculated according
to (5.24). The mass and the energy balances are written:
0 mflow,mix  mflow,drain (5.30a)
0 Hflow,mix  Hflow,drain  mflow,mix ∆h (5.30b)
where subscript drain denotes the low pressure side of the turbine.

5.3.7 Combustor
The combustion in a gas turbine is assumed to be a constant pressure process. For a
constant pressure combustion process the sum of the reactant enthalpies (air and fuel)
is equal to the sum of the product enthalpies (substances). Given the total enthalpy,
the adiabatic flame temperature T f can be calculated according to (4.25). Since it is not
necessary to determine the flame temperature T f explicitly, the enthalpy at the exhaust
port of the combustor component is calculated according to:
he h a χ a  h f ˆ1  χ a  (5.31)
where h f is enthalpy of fuel, h a is enthalpy of air, h e is enthalpy of the hot exhaust gas,
and χ a is the mass fraction of air. The mass fraction of air is:
mflow,a
χa (5.32)
mflow,a  mflow,f
68 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

with the same subscripts as previous. The air/fuel ratio λ of the exhaust gas is:

mflow,a ~mflow,f
λ (5.33)
ˆ A~F s

directly according to the definition in (4.27). A pressure loss model through the combus-
tor is defined according to Saravanamuttoo et al. (2001):

2A2 ρ 1 ∆p
PLF 2
(5.34)
mflow,a

where PLF is the pressure loss factor which is constant, ∆p is the pressure drop, A is the
maximum cross-sectional area of the chamber, and ρ 1 is the density for the incoming air.

5.3.8 Pressure Losses

Two types of pressure loss components are implemented in GTLib. These pressure losses
are: (i) a simple pressure loss component (5.35a), and (ii) a turbulent pressure loss
component (5.35b) according to:
¿
Á
Á
À
∆p
mflow mflow,ref (5.35a)
∆pref
¾
2ρ 1 ∆p
mflow A (5.35b)
ξ

here ξ is a pressure loss factor, ∆p is the pressure loss, p 1 is the pressure of the upstream
gas, ρ 1 is the density for the upstream gas, and subscript ref denotes reference values.

5.4 Gas Turbine Model

The main difference between the gas turbine model (described in the present chapter)
and the reference gas turbine model (described in Section 2.3.1) is the description of the
gas model. In the developed gas turbine model, the exhaust gas is specified through the
air/fuel ratio λ and in the reference gas turbine model the gas is specified through the
mass fraction of species. A benefit with the model constructed in GTLib is the reduction
of equations and state variables according to the air/fuel description. The reduction in
model equations reduces the demanded simulation time by about 20 per cent. The model
constructed in GTLib is not as general as the reference model constructed in SITLib since
pure substances cannot be injected throughout the gas path, only the exhaust medium
(with a air/fuel ratio λ > 1, ª ) can be injected.
The constructed gas turbine model consists of a number of standard components,
e.g., control volumes, turbines, and valves which is shown in Figure 5.5. The gas turbine
model and the reference model consist of the same type of components and have the
same type connection to the environment. Thus, it is possible to simulate both models
5.5. Conclusion 69

using the same simulation platform. The available instrumentation sensors measure:
(i) pressures, (ii) temperatures, and (iii) shaft speeds throughout the gas path. Between
the output of the compressor C1 and the output of the power turbine T0 there are no
available measured gas path parameters. Since gas path parameters between these two
points are important to supervise, it is necessary that the developed diagnosis tests can
estimate those parameters.
When the gas turbine model is designed it can be used for performance analysis
and as a base when diagnosis test quantities used in the FDI-system are constructed.
Why the gas turbine model and not the reference model can be used for this purpose
depends on the parsers which are developed. Since the Modelica Media package used in
the reference model is very general, it can also be a challenge to develop parsers for this
model. Thus, the developed parsers can only parse a subset of the Modelica language.
The components in GTLib fulfill the specification of the parsers and can therefore be
used to generate diagnosis test quantities. It will also be shown that the reference model
has unobservable state variables which is not desirable when tests based on observers are
constructed. The observability of the gas fraction in the reference model is also unclear.
More about the test selection and construction procedure will be described in Chapter 6.

5.5 Conclusion
In the chapter, a gas turbine modeling package – GTLib implemented in Modelica is
presented. The GTLib package consists of two main parts: (i) a gas model, and (ii) a gas
turbine component library. The gas model handles the calculation of thermodynamic
properties of the fluid, and the gas turbine component library consists of the gas turbine
components. Later on in the chapter, a gas turbine model is constructed from the
components in the GTLib package. In the GTLib, the air/fuel ratio concept is introduced
which reduces the number of equations and variables in the overall gas turbine model.
The constructed model can be simulated together with the existing simulation platform.
The gas turbine model can handle different changes in ambient conditions. These ambient
conditions are the pressure, the temperature, and the relative humidity.
The benefit with using GTLib package is the reduction in model equations compared
to the reference model constructed in SITLib. The accuracy of the two models is similar
except for transients in absolute humidity of the ambient air. In GTLib, the air properties
are updated simultaneously in all control volumes in the gas turbine model which gives
the behavior. A disadvantage with GTLib is the loss in generality, here the only gases that
are admitted to be used are the air and fuel. The consequence is that an injection of, e.g.,
pure oxygen somewhere in the gas path is not allowed. An advantage with GTLib is that
the reduced number of equations gives a decreased simulation time when the simulation
platform is simulated. The main propose with GTLib is that a diagnosis and supervision
system can be constructed with the GTLib gas turbine model as a base.
70 Chapter 5. GTLib – Thermodynamic Gas Turbine Modeling Package

IGN

BPV
port_fuel

BV2
V31 CC V49
dp49_5

dp3_31

flange_b
V5
BV1 byp_vlv

mechLoss2
V3

y
setCool
CoolT1

C1 T1 T0
mechLoss1
flange_a shaft1 shaft2
V2
dp1_2
y
setCool

setCool
y
CoolT0 V60
V10

Bleed_3

BV1_vlv

BV2_vlv

V75
dp0_1

Cool1
P
y Sink
setCool

dp75_8
feed drain
p1s

p3s

p8s

T2s

T3s

T75s

nC1s

nT0s
y

y
p1

p3

p8

T2

T3

T75

nC1

nT0

Figure 5.5: In the figure, a model of the 2-shafted gas turbine presented in Figure 2.4 is shown.
The model consists of components from the GTLib package, e.g., control volumes, valves,
and turbines. This model can be simulated together with the simulation platform viewed in
Figure 2.5 since both gas turbine models have the same environment connections except for the
gas sources. This model can be used for dynamic simulations exactly as for the reference model
where the accuracy between the two models is high, but with a reduced simulation time. During
start-up phases, the actuator signals BV1 and BV2 are used to control the bleed valves. The
other actuator signals are IGN and BPV, where IGN control the ignition and BPV the bypass
valve through the combustion chamber. The sensors measure pressure and temperature before
and after the compressor C1, pressure and temperature after the power turbine T0, and the shaft
speed of the gas generator and the power turbine. Between the output of the compressor and
the power-turbine there are no measured gas path parameters.
 Chapter 6

Modeling, Analysis, and Transformation of the

Diagnosis Model

The overall objective with the present work is to achieve a systematic method to construct
an FDI-system from an available model used for performance analysis. The first step in
the procedure is to design a diagnosis model where diagnosis concepts are introduced.
To simplify the test quantity construction, the diagnosis model has only signal based
connections to the environment (compared to the performance model which has physical
based connections). In the diagnosis model design, the knowledge that is built in the
performance model is transferred to the diagnosis model according to the object oriented
nature of the Modelica language.

Detection and Isolation of Faults

Before the FDI-system is built, it is relevant to make some preliminary investigations how
introduced faults, or component deteriorations, affect the fault detection and isolation
performance. For the present gas turbine application, is it possible to detect and isolate:
(i) faults in the input and output signals, and (ii) drifts in the performance parameters
with the given set of measurement sensors? By definition, when a fault f i is detectable it is
possible to distinguish between a non faulty system behavior and a faulty system behavior
affected by the induced fault f i . The detectability and isolability criteria depends on the
particular sensor configuration where more sensors increase the diagnosis performance.
The fault isolability requirements can be defined in a similar manner as for the fault
detectability. Instead of distinguishing between a non faulty mode and a faulty mode,
it should be possible to distinguish between two different faulty modes. For linear
differential algebraic systems, the isolability is formally defined in Frisk et al. (2009) as a
null-space calculation of the system matrix H ˆ p in:

H ˆ p x ˆ t   L ˆ pz ˆ t   F ˆ p f ˆ t  0 (6.1)

71
72 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

where x ˆt  > Rn x represents the unknown variables, z ˆt  > Rn z represents the known
signals, and f ˆt  > Rn f represents the unknown faults. The matrices: H ˆ p, Lˆ p, and
F ˆ p are polynomial matrices in the differential operator p d ~dt, and they have an
appropriate dimension. For an over-determined system, the number of rows in the
matrices is larger than the dimension of x. In the linear case, the isolability properties
can be determined through a null-space calculation of the system matrix H together
with the faulty mode F j which represents how fault f j affects the system. To check if f i
is isolable from f j , the null-space calculation should not be perpendicular to the faulty
mode F i which is formally written:
N HF j ˆ pF i ˆ p x 0 (6.2)
where N HF j ˆ p is the null-space of H ˆ p F j ˆ p. Thus, the fault f i is isolable from
fault f j in (6.1) if, and only if (6.2) is satisfied. To check in practice if the system may be
faulty, the linear filter:
r ˆt  γN H ˆ pLˆ pz ˆt  (6.3)
is designed. The filter in (6.3) is called a consistency based residual generator where
the design parameter γ (weight factor) is a vector of suitable dimension. The residual
generator should fulfil:
lim r ˆt  0 (6.4)
t ª
when the system is in a non-faulty mode.
For nonlinear systems, it can be difficult to get an exact characterization of the
model equation (i.e., based on consistency relations) which can be used when residual
generators are constructed. In most cases when a residual generator is constructed, the
system model together with the measurement equations are transformed into a number
of smaller subsets where only the interesting faults are present see, e.g,. Blanke et al.
(2003); Chen and Patton (1999); Nikoukhah (1994). The faults that are not present in
the subset of equations are not detectable by the residual generator, i.e., theses faults are
decoupled by the residual generator. A common method to construct residual generators
is to design estimators based on state estimates, i.e., an observer design. In this class of
residual generators, the state estimates of the observer is compared to the measurement
signals to construct a residual generator. Residual generators based on state observers
are presented in, e.g., Patton and Hou (1998); Martínez-Guerra et al. (2005); Svärd and
Nyberg (2008).
To answer questions about detectability and isolability properties of a model in a
general way, only the model structure can be considered (Blanke et al., 2003). For the gas
turbine model considered herein, an analytic characterization of the model equations
used for residual generators is not practical to find due to: (i) the size of the system,
(ii) the nonlinear behavior of the model, and (iii) the look-up tables used for performance
calculations. Thus, structural methods together with an observer design are considered
when the test quantities (residual generators) used in the FDI-system are constructed.

Outline of the Chapter

The chapter starts with an historical overview of the gas turbine monitoring field where
the estimation parameters (which are called health parameters) used in the Gas Path
6.1. Gas Turbine Monitoring 73

Analysis are presented. In Section 6.2, the diagnosis model is presented which introduces
the faults and estimation parameters used for diagnosis. The following sections deal with
the analysis of the diagnosis model. The investigations are: (i) the DAE-index analysis in
Section 6.3, (ii) the structural analysis in Section 6.4, and (iii) the observability analysis
in Section 6.5. In Section 6.7, a number of parsers used for an automatic extraction of
the test equations are presented. These parsers are used to convert the Modelica models
constructed in GTLib package into runnable Matlab code.

6.1 Gas Turbine Monitoring

In industrial gas turbines, deterioration in components throughout the gas path is com-
monly occurring and contributes to the overall performance degradation of the engine.
Diagnosis and supervision of performance degradation in the application is a widely
studied topic in the gas turbine monitoring research field, see, e.g., Aker and Saravana-
muttoo (1989); Volponi et al. (2003); Doel (2003) where the performance parameters are
estimated with different methods. When reliable performance estimations are available,
it can be easier for the service engineers to efficiently plan service and maintenance of
the gas turbine. In Aker and Saravanamuttoo (1989), it is investigated how compressor
fouling affects the performance parameters using a linear fouling model. For a medium
fouled compressor, the estimations appear to be reasonably accurate for the linear fouling
model.
In papers Diakunchak (1992); Kurz and Brun (2001); Kurz et al. (2009), several mech-
anisms which cause degradation in gas turbines are presented. The major contribution
of degradation mechanisms in an industrial gas turbine is fouling. The fouling is caused
by small particles and contaminants in the air that are caught by the compressor. These
particles increase the roughness of the rotor and stator surface. Another effect that
results in performance degradation is the tip clearances which is a common diagnosis
for aged gas turbines. Tip clearances denote an increasing gap between the rotating
blades and the stationary casing. Fouling due to increased roughness can partially be
restored by washing the compressor, while a component replacement is often needed
for tip clearances. In the paper Brekke et al. (2009), compressor fouling in two different
off-shore gas turbine applications is investigated. The fouling analysis showed that a
considerable amount of contaminants appeared at the compressor inlet, at the inlet guide
vanes, and at the first rotor stage of the compressor. In this case, the main contaminant
in the samples was sodium-based salts which indicate that the gas turbine performance
can be restored by a compressor wash.
In Diakunchak (1992), an analysis due to fouling is performed using experimental
data from an industrial gas turbine site. In the paper, it is shown that the fouling of
the compressor results in a 5 % reduction in inlet mass flow and 1.8 % reduction in
compressor efficiency from the nominal values. This gives a reduction in the output
power by about 7 % and increases the heat rate of the fuel by about 2.5 %.
74 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

6.1.1 Gas Path Analysis

One of the most famous and pioneering tool for gas turbine engine supervision and
sensor diagnosis is the gas path analysis (GPA). The GPA algorithm can be summarized
into the main steps: (i) measurement normalization, (ii) reference value generation,
(iii) estimation of performance deviation, and (iv) diagnosis decision. This algorithm was
introduced by Urban (1969) and an investigation of the method in a gas turbine engine
application was presented in Urban (1972). The GPA method is based on thermodynamic
relationships where one of the main objective is to estimate deterioration in gas path
components from a number of measured sensor signals which are denoted as measure-
ment deltas (∆). In Urban (1969), a chart that contained the most commonly occurring
gas turbine engine parameter interrelationships in a general matrix form was presented.
These equations could be used to estimate steady state and transient variations in the
performance parameters for an arbitrary gas turbine engine during most conceivable sets
of input conditions. How the chart is calculated is re-printed in Giampaolo (2009) and
the linear relationship between the measured signals and the engines health parameters
is written (Volponi, 1994):

∆Z H ∆h  e (6.5)

where ∆Z is the measured parameter deltas, ∆h is the deviation in performance, H is

the influence coefficient matrix presented in Urban (1969), and e is the measurement
noise. Elements in ∆h are typically: efficiencies ∆η, and flow capacities ∆Γ of the gas
path components such as compressors, turbines, and fans. Elements in ∆Z are typically:
spool speeds ∆N, temperatures ∆T, and pressures ∆P. The matrix H can be divided
into two parts: (i) an engine fault influence matrix H e , and (ii) a sensor fault influence
matrix H s , where the previous defined matrix H and health parameter ∆h in (6.5) are
extended with the sensor fault dependencies.
The main objective with the gas path analysis framework is to estimate the perfor-
mance deviation vector ∆h. Depending on the size of the H matrix, different approaches
are relevant to consider. When engine’s health is considered and not the sensor faults, the
system (6.5) is often over-determined, i.e., the number of measurement deltas are larger
than the number of performance parameter deltas. This results in an optimization prob-
lem where, e.g., least-square methods can be applied to solve the optimization problem.
When also faults in sensors are considered in the equation system (6.5) the ∆h vector is
extended with the faults and the matrix H is extended with the influence of the faults.
This results often in an under-determined system. For this case, the estimation procedure
results in, e.g., minimization of some predefined cost function, or a maximization of
the probability that gives the observed measurement. In Doel (1994) the commercial
gas path analysis program TEMPER is presented. In TEMPER, it is assumed that the
deviation in performance ∆h and the measurement noise e are Gaussian distributed. To
maximize this conditional probability distribution function is equivalent to minimize a
quadratic form which has a weighted least-squares solution.
6.1. Gas Turbine Monitoring 75

The Measurement Delta Vector

The measurement delta vector ∆Z describes the deviation, from a nominal baseline
in percent, for a number of known signals (or combination of known signals). These
deviations (or deltas) are then collected in the ∆Z-vector in (6.5). For example, the
measurement delta ∆Z P of the gas path parameter P is defined:

∆ZP 100
P ‡ Pnom

‡
(6.6)
Pnom
‡

where P ‡ is the corrected»value according to the ambient standard conditions (e.g., the
corrected speed n‡ n~ˆ T R 㠍 as described in (5.2b)), and Pnom ‡ is the nominal value
under the chosen reference condition. The nominal value is typically calculated using a
reference engine model together with an input signal u. The inputs can be, e.g., the mass
flow of fuel, and/or the power generated by the application. A sketch of the calculation
procedure is shown in Figure 6.1. The measurement deltas can directly be considered
to detect abnormal changes in the supervised components according to deteriorations
or sensor faults. The measurement deltas are utilized in, e.g., Ganguli and Dan (2004)
where a recursive median filtering approach is used to recognize an abnormal change in
the supervised components. When the measurement deltas are determined they can be
used in (6.5) (for the linear case) to calculate the performance deviation ∆h vector using,
e.g., some optimization method.

Measurement
ymeas
Data
P‡
Normalization 
100
Σ Pnom
‡

∆ZP
u Reference
Engine Model Pnom
‡

Figure 6.1: Measurement delta ∆Z P calculation of the gas path parameter P, where ˆ‡ denotes
the corrected version of the parameter according to, e.g., (5.2) and Volponi (1999). The corrected
nominal value Pnom‡ is typically calculated using a reference engine model together with an
input signal u.

6.1.2 Engine Health Monitoring

A common approach in the gas turbine diagnosis research field to capture performance
degradation is to introduce a number of physical based quantities named health pa-
rameters. The motives to use those extra parameters are: (i) the health parameters are
indicators for overhaul and maintenance, and (ii) the health parameters compensate for
the deterioration in components which result in more reliable estimates of unmeasured
signals used by, e.g., the controller. The health parameters were included in the perfor-
mance deviation ∆h vector in the static equation system (6.5). A natural extension of the
76 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

static model is to introduce the health parameters in a dynamic gas turbine model. As in
the static case, these health parameters are typically corrections (or deviations from a
nominal baseline) of efficiencies and flow capacities. The considered health deterioration
appears in, e.g., the compressors, the turbines, the fans, and the nozzles. The introduced
health parameters can be estimated with a number of techniques, see, e.g., Luppold et al.
(1989); Kobayashi and Simon (2003); Borguet and Léonard (2008). In the first two papers,
Kalman filters are utilized to estimate the health parameters directly. In the third paper,
a nonlinear Kalman filter is used together with a quadratic optimization procedure. The
optimization problem is studied to estimate the health parameter values which are then
used in the Kalman filter to estimate the other state variables of the system.
Since performance degradation in the gas turbine is naturally occurring, it can be
difficult to avoid false alarms when the diagnosis system does not compensate for the
deterioration in components. This depends on the fact that the errors in the performance
model get larger when the component deteriorations increase. This can result in estimated
sensor values that differ too much from the measured sensor value which may trigger a
false alarm (Kobayashi and Simon, 2003).

6.2 Gas Turbine Diagnosis Model

The objective with the diagnosis model is to specify the properties of the test quantities
which are later on used when the FDI-system is designed. Using the diagnosis model, the
ability to capture abnormal process behaviors using some extra estimation parameters is
introduced. The considered abnormal process behaviors can be divided into the three
general groups: (i) performance deterioration in components, (ii) component faults,
and (iii) sensor and actuator faults. Component faults can be, e.g., a leakage in a valve
or an increased friction in a bearing. In this chapter, the methodology to transform the
diagnosis model into smaller subsets that are called test equations is introduced. Those
test equations are then implemented and evaluated on experimental data in Chapter 7. A
graphical representation of the diagnosis model, along with the input and output signals,
is shown in Figure 6.2 and in Figure 6.3.
The gas turbine diagnosis model is a reduced and a simplified version of the reference
gas turbine performance model which is developed by SIT and shown in Figure 2.5. The
phrase reduced means that the number of equations and state variables in the model has
decreased according to the utilization of the GTLib package. The phrase simplified means
that some model simplifications have been done. In the diagnosis model, the following
components are removed: (i) the bleed valves in the compressor, (ii) the bypass valve
through the combustor, and (iii) the heat losses in components after the combustion
(but heat losses in the combustor are still considered). The bleed valves are usually used
during startup and shutdown phases (to avoid stall and surge in the compressor) and at
full load operations these valves are fully closed. The bypass valve is usually used during
partial base loads but does not affect the performance calculations significantly. Since it
is assumed that the valves are closed it is possible to remove them from the model. Since
the bleed valves are removed, the model is only valid during operational conditions, and
are not valid during startup and shutdown operating modes (in Chapter 8 where startup
6.2. Gas Turbine Diagnosis Model 77

IGN
port_fuel
m_f

n_A

P_A
1
t0
V31 CC V49
dp49_5
S
dp3_31

FuelSource rpm MW

V5
torque torque
p0
V3

mechLoss2
y
setCool
CoolT1
phi0

C1 T1 T0
mechLoss1
shaft1 shaft2
V2
dp1_2
y
setCool

setCool
y

CoolT0 V60
V10

p0
Bleed_3

V75
dp0_1

Cool1
P
y Sink
T
P setCool Constant
feed P
dp75_8
drain
p2s

p3s

p8s

t2s

t3s

t7s

nC1s

nT0s
y

y
p2

p3

p8

t2

t3

t7

nC1

nT0

Figure 6.2: The figure gives an overview of the gas turbine diagnosis model. To get the diagnosis
model, some of the components from the gas turbine model used for performance analysis are
removed. These components are the bleed valves and the combustor bypass valve. Thus, the
diagnosis model is valid only during operational conditions. The input signals to the diagnosis
model are: (i) temperature T0 , pressure p 0 , and relative humidity φ 0 of the ambient air, (ii)
the mass flow of fuel m f , and (iii) speed and power of the power-turbine T0. Since the model
has no physical based connections to the environment, it is possible to simulate the diagnosis
model with the given input signals.
78 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

and shutdown operating modes are studied, these valves are introduced again).

6.2.1 Input Signals

All industrial gas turbines are equipped with a number of actuators and instrumentation
sensors. The input signals are divided into two groups: (i) ambient air signals, and
(ii) other signals. The first group of signals is: pressure p 0 , temperature T0 , and relative
humidity φ 0 of the ambient air. These signals are used to setup the surrounding ambient
conditions, and used to calculate the concentration of substances in the ambient air using
a moist air model, i.e., how much water steam that is mixed with dry air. The other group
of signals is: mass flow of fuel m f , power generated by the application PA , and speed of
the driven component n A .

6.2.2 Output Signals

A number of measurement signals throughout the gas path is available for monitoring and
control. The considered output signals are: (i) temperature T2 and pressure p 2 before the
compressor, (ii) temperature T3 and pressure p 3 after the compressor, (iii) temperature
T7 and pressure p 8 after the power-turbine, and (iv) speed of gas generator n C1 and the
power-turbine n T0 .

6.2.3 Health Parameters

To capture degradation in efficiency and mass flow, health parameters in: (i) the com-
pressor C1, and (ii) the compressor-turbine T1 are introduced in the diagnosis model.
It is also possible to insert health parameters in the power-turbine T0, and in the inlet
and outlet duct to detect abnormal pressure drops. Depending on the configuration, the
considered health parameters are utilized in the performance equation:

ηi f i ,η ˆ. . .  ∆η i
(6.7)
Γi f i ,Γ ˆ. . .  ∆Γi

where η i is the efficiency, Γi is the flow capacity, and f i , j represents the nominal char-
acteristics. Subscript i represents the components: C1, T1, or T0. In function f i , j , the
concept of corrected and normalized parameters are utilized according to (5.2)–(5.4).
Since the component deterioration is assumed to be slow relative to the other gas turbine
dynamics, constraints according to:

∆η̇ i 0
(6.8)
∆Γ̇i 0

are added to the model for each health parameter.

6.2. Gas Turbine Diagnosis Model 79

p0 t0 φ0 mf PA nA

CC

C1 T1 T0 A

∆η C1 ∆ΓC1 ∆η T1 ∆ΓT1

T2 p2 T3 p3 n C1 T7 p8 n T0

Figure 6.3: The gas turbine with the output signals (solid), the input signals (dashed), the input
ambient signals (dotted), and health parameters (arrows) is shown in the figure. The secondary
air flows, used to cool the first turbine blades, are shown with dashed arrows.

6.2.4 Component Faults

In the diagnosis model, faults in components can be introduced in the similar manner as
for the health parameters. A component fault is, e.g., a leakage in a valve or an increased
friction in a mechanical bearing of the shaft.

6.2.5 Faults in Signals

A faulty instrumentation sensor or actuator can be defined as a signal that shows an
abnormal behavior. The abnormal behavior can be interpreted as, e.g., the sensor char-
acteristic specified by the manufacture that is no longer maintained. The abnormal
behavior can result in different kind of faulty behavior, e.g., these faulty behaviors can
appear as:

• Abrupt changes – the faulty value changes behavior immediately.

• Incipient faults – the faulty value appears, and gradually increases in amplitude.

• Intermittent faults – the faulty value appears and disappears with a time interval.

• Bias faults – the faulty value is constant, i.e., the value has a constant offset.

In most cases, the faulty behavior of the sensor or actuator is unknown. In this case,
the sensor signal can be modeled with an unknown variable f i added to the measured
quantity.
80 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

Sensor Faults
The sensor fault in the gas turbine model are introduced as:

y p1 p1  f p1
y p3 p3  f p3 (6.9)
  

where y i is the known measurement signal, and p i is the measured quantity.

Actuator Faults
The faults in the input signals of the gas turbine model are introduced as:

um f mf  fm f
u PA PA  f PA (6.10)
unA nA  fnA

where u i is the known input signal, and p i is the input quantity.

6.2.6 Differential Algebraic Equation Form

The diagnosis gas turbine model, with the added health parameters and sensor/actuator
faults, has the general form:

F ˆż, z, u  f u  0 (6.11a)
y h ˆz   f y (6.11b)
z ˆ x, ∆ T (6.11c)

where ∆ ˆ∆η C1 , ∆ΓC1 , ∆η T1 , ∆ΓT1 T represents the unknown health parameters, x

consists of the unknown variables, y consists of the known measurement signals, u
consists of the known input signals, f s consists of the unknown sensor faults, and f a
consists of the unknown actuator faults. The functions F and h together with their
arguments are vector valued functions with appropriate dimensions. Equation (6.11a) is
a general mathematical description where the algebraic and the dynamic constraints are
mixed, i.e., in a DAE-form. The expression in (6.11) is the starting point for the model
analyses performed in the remaining part of the chapter.

6.3 DAE-Index Analysis

The objective with this section is to transform the system in (6.11) to a system in state
space form:

ż 1 f ˆz 1 , u  (6.12a)
y h ˆz 1  (6.12b)
6.3. DAE-Index Analysis 81

where z 1 represents the state variables, y represents the measurement signals, and u
represents the input signals. The functions f and h together with their arguments are
vector valued functions with appropriate dimensions. A central concept in this section
is the differential index (Hairer et al., 1991; Ascher and Petzold, 1998) of the system. The
differential index of the system tells the minimum number of times the constraints must
be differentiated to obtain a state space form.

6.3.1 Differential Index Reduction

The first step in the procedure is to check and reduce the differential index of the diagnosis
model. After row and column permutations are performed, the DAE system in (6.11)
can be written in the form:
Ē 0 ˙
 z̄ f¯ˆz̄, u
0 0 (6.13)
y h ˆz̄ 
where u consists of the known actuator signals, z̄ consists of the unknown variables,
and Ē is a constant matrix which does not have full column rank (is singular). For
this system, a state space form cannot be directly obtained before an index analysis is
performed. In general, a model with a DAE index-1 (or 0) is easier to handle in practice
than a system with a higher index (Ascher and Petzold, 1998). One way of handling this
situation is to employ an index reduction technique and reduce the index of the system.
A common method is to differentiate well chosen model equations a suitable number
of times to obtain a low index DAE model. The constraints which are differentiated
are removed from the model and treated separately. A disadvantage with the model
reformulation is the numerical inabilities to satisfy the removed constraints which can
appear. To handle the numerical problems techniques with, e.g., dummy derivatives
(Mattsson and Söderlind, 1993) can be considered. To select which constraints that
should be differentiated, Pantelides’s algorithm (Pantelides, 1988) is considered. Using
Pantelides’s algorithm, also the differential index of the system can be determined. The
input to the algorithm is the structural model (shown in Section 6.4) of the system. The
index analysis shows that (6.13) is a DAE index-2 system, thus one derivation of selected
algebraic constraints is necessary. When the equations, chosen by Pantelides’s algorithm,
are differentiated the diagnosis model (6.13) is written in the form:
E1 0 ż 1 f 1 ˆz 1 , z 2 , u 
   (6.14a)
0 0 ż 2 g 1 ˆz 1 , z 2 , u 
0
g̃ 0 ˆz 1 , z 2 , u  (6.14b)
∂g
where z 1 > Rn 1 , z 2 > Rm 1 , rankˆE 1  n 1 , and rankˆ ∂z 21  m 1 (at least locally). The
algebraic constraints g̃ 0 , which are differentiated, are saved in (6.14b).

High Index Property of the Diagnosis Model

The Mechanical Rotational library, which is a part of the Modelica standard package,
share torque τ and angle φ (5.24b) in connection points between rotational components.
82 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

The index analysis shows that the high differential index property of the gas turbine
model is a consequence of the angles dependencies in the connection points (angles
are equal in a connection point). Pantelides’s algorithm suggests to differentiate the
equations in the form: φ 1 φ 2 , φ 2 φ 3 , . . . to obtain: φ̇ 1 φ̇ 2 , φ̇ 2 φ̇ 3 , . . .. When the
system is replaced with these new differentiated equations, the differential index of the
system (6.14a) is 1. The equations, before they are differentiated, are saved in the g̃ 0 part.
To get the quadratic E 1 matrix, e.g., QR-factorization or SVD-decomposition can be
considered. Here, it is enough to replace the dynamic constraints: φ̇ 1 φ̇ 2 , . . . , φ̇ n 1 φ̇ n
with the algebraic constraints: ω 1 ω 2 , . . . , ω n 1 ω n to get a quadratic E 1 matrix.

Semi-Explicit Index-1 DAE Form

The system in (6.14a) can be written in semi-explicit index-1 DAE form:

E 1 ż 1 f 1 ˆz 1 , z 2 , u  (6.15a)
0 g 1 ˆz 1 , z 2 , u  (6.15b)
y h 1 ˆz 1 , z 2  (6.15c)

where the E 1 -matrix is invertible, and the Jacobian ∂g 1 ~∂z 2 is non-singular. This means
that the algebraic constraint g 1 is invertible (at least locally), and can be solved for z 2
(Hairer et al., 1991; Ascher and Petzold, 1998). For the semi-explicit index-1 DAE system it
is possible to distinguish between the differential variables z 1 , and the algebraic variables
z 2 . This semi-explicit system can directly be transformed into a state space form through
an inversion of the E 1 matrix and an elimination of the algebraic variable z 2 using the
constraint g 1 .

State Space Form

A state space form of the semi-explicit index-1 DAE system (6.15) is:

ż 1 f 2 ˆz 1 , u  (6.16a)
y h 2 ˆz 1 , u  (6.16b)

where the algebraic variable z 2 has been eliminated. This is possible since the Jacobian
matrix ∂g 1 ~∂z 2 of (6.15b) is, at least locally, non-singular. An observability analysis
performed in Section 6.5 shows that the system (6.16) has unobservable modes. Thus, an
observer cannot be constructed directly from the system (6.16) and further analysis of
(6.15) is necessary to perform.

6.4 Structural Analysis

The objective with the structural analysis is to select relevant parts, and investigate fault
detectability of the model. For nonlinear DAE systems, it can be difficult to get an
exact characterization of the model equation which can be used when diagnosis tests
are constructed. To have a structural analysis with higher accuracy, it is a good idea
to specify as much detail as possible about the system. For the gas turbine engine, the
6.4. Structural Analysis 83

gas flow has a given direction and the combustor is always turned on. This knowledge
is utilized when the structural analysis is performed and gives a structural model that
gets sparser. The structural model is a coarse model description where only the variable
dependencies in each equation are considered. The analytic model is described by a
matrix where each element has a true (1), or a false (0) value. How the variables affect
the analytical equations are not considered in the analysis. Whether they affect the
expression through, e.g., exponential, logarithmic, or using a look-up table, they still
get the same variable dependencies in the structural model. When a specific variable is
included in a specific equation, the matrix element belonging to the certain variable is
(1). All these variable dependencies of the equations are collected in the structural model
where the rows represent the equations and the columns represent the variables. When
the model is in this form, the opportunity to develop fast algorithms for model analysis
and especially for diagnosis purposes is available. The disadvantage is that only the best
case solution is obtained.

6.4.1 Dulmage-Mendelsohn Decomposition

A method that can be useful when relevant subsets of model equations are chosen is the
Dulmage-Mendelsohn (DM) decomposition, presented in Dulmage and Mendelsohn
(1958). The DM decomposition is an equivalent description of a bi-partite graph, which
specify the calculation chain of a system. The decomposition works on the structural
model, and rearranges rows and columns to obtain the structure shown in Figure 6.4.
In the figure, the structural model is divided into the three separated parts: (i) an
under-determined part M  , (ii) an exactly-determined part M 0 , and (iii) an over-
determined part M  . These parts are described as:

(i) In the under-determined part, the number of equations in M  is less than the
number of variables in X  . When the underlying analytical model can be simulated,
this part never appears. The gas turbine diagnosis model used here does not involve
this part and is therefore not considered in the analysis.

(ii) In the exactly-determined part, the number of equations in M 0 is equal to the

number of variables in X 0 . The exactly-determined part is divided into a number
of components a i . The components that consist of more than one equation are said
to be strongly connected which means that the equations contain loops or cycles.
For diagnosis purposes, the exactly-determined part of the model provides no
extra information. Thus, this part of the model can be removed without loosing
any redundancy.

(iii) In the over-determined part, the number of equations in M  is more than the
number of variables in X  . This indicates that redundancy is available and the
degree of redundancy depends on the number of available measurement sensors.
When diagnosis tests are constructed, different subsets of the M  can be chosen.
In Krysander et al. (2008), a class of these subsets is denoted Minimal structural
over-determined sets (MSO sets). These subsets are easy to determine using the
84 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

X X0 X

a0 M

a1

a2
equations

 M0

f1 a n 1
f2
an
f3

f4
f5
aª M

f6

variables

Figure 6.4: The figure shows a Dulmage-Mendelsohn decomposition of a general structural model
where the rows represent the equations and the columns represent the variables. The light grey
shadowed area consists of zero matrix elements. The matrix element in the darker grey area
can either be zero or one. The elements that are left span the boxes a 0 . . . aª . Injected faults
f 1 , f 2 , f 3 are not detectable, but faults f 4 , f 5 , f 6 can possibly be detectable.
6.4. Structural Analysis 85

proposed algorithm. The MSO sets are the smallest over-determined subsets with
redundancy which can be obtained.

For diagnosis purposes, it is only necessary to consider the over-determined part

M  because faults that appear in any equations in the exactly determined part are not
detectable because of the lack of redundancy. In Figure 6.4, the faults f 1 , f 2 , and f 3 are
not detectable, and the faults f 4 , f 5 , and f 6 are in an ideal case detectable. However,
measurement noise can, e.g., disturb the test quantities that can result in a fault which is
not detectable in a practical application.

6.4.2 Investigation of Actuator Fault and Health Parameter Isolation

In Krysander and Frisk (2008), a method is presented where the over-determined M 
part is divided into a number of smaller subsets of equations, i.e., equivalence classes.
The advantage with using these subsets is that if any of the equations in the subset is
removed, the remaining equations in the subset get exactly determined, thus the degree
of redundancy is equal to one. In practice, when a diagnosis test (or residual generator)
is constructed it is necessary to have at least one redundant equation. In an equivalence
class only one redundant equation is available. When this equation is removed the faults
in the remaining subset are no longer detectable. Thus, it is not possible to construct two
tests that discriminate between the two (or more) faults in the same equivalence class, and
therefore, they are not isolable from each other. A decomposition of the over-determined
part in Figure 6.4 is shown in Figure 6.5 where the equivalence classes are determined.
To analyze the isolation performance of the actuator faults and the health parameters
in the gas turbine model (6.11) the structural model is studied. The sensor faults are not
considered since they appear automatically in the over-determined part and appear also
in separate equivalence classes. In the analysis, the derivative of the health parameters
is not considered and the deterioration in the components is treated as fault. This is
written:

ηi f i ,η ˆ. . .  ∆η i (6.17a)
Γi f i ,Γ ˆ. . .  ∆Γi (6.17b)
ũ i ui  fi (6.17c)

where f i , j represents the look-up tables, u i represents the input signals, and ∆η i , ∆Γi , and
f i are all considered to be zero for the healthy gas turbine. In Figure 6.6, the structural
model of the gas turbine diagnosis model using the Dulmage-Mendelsohn decomposition
is shown. The structural model is divided into the exactly-determined M 0 , and the
over-determined M  part. Since only the over-determined part is considered in the
FDI-system the first 390 equations can be removed. For example, variables that appear
in the exactly-determined part (and can be removed) is, e.g., φ i which is the angle of the
shaft shown in (5.24b). The shaft angles are also state variables. Since all the considered
faults and health parameters appear in the over-determined part they are detectable.
In Figure 6.7, the over-determined part from Figure 6.6 is divided into a number of
equivalence classes. It is shown that the faults in the input signals of power and speed
86 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

X 1 ... X n1 X n X 0

M 1  1

 

M n1 f4  1

f5
M n  1
f6

M 0 

Figure 6.5: In the figure, the decomposition of the over-determined M  part from Figure 6.4 into
equivalence classes is shown. The equivalence classes are the sets of pair ˆ M i , X i  where
i 1 . . . n. The equivalence classes have one more equation than the number of variables which
results in that the faults f 5 and f 6 may be detectable but not isolable from each other. Since f 4
isn’t in the same equivalence class as f 5 and f 6 , the fault f 4 may be isolable from fault f 5 and f 6 .
6.4. Structural Analysis 87

of the generator are not isolable. The other fault and health parameters may be isolable
depending of, e.g., how the tests are constructed or the noise level in the measurement
sensors.

6.4.3 DAE-index 1 Conservation in the Over-Determined M  Part

In the diagnosis tests, it is only the over-determined part that is of interest since it is only
in this part of the model redundancy is available. Therefore, it is of great importance if the
over-determined part of the model is a DAE-index 1 system when the index reduced gas
turbine model (6.16a) is a DAE-index 1 system. This is true and is shown in Theorem 6.4.1.

Theorem 6.4.1 The over-determined M  part of the DAE-index 1 system in (6.15) is also
a DAE-index 1 system when the measurement equations are removed.
Proof The index-1 DAE system:
E ẋ 1 f ˆx 1 , x 2 , u 
0 g ˆx 1 , x 2 , u  (6.18)
y h ˆx 1 , x 2 
has an invertible and constant matrix E, and a non-singular (at least locally) Jacobian
matrix ∂g ~∂x 2 , see e.g., Ascher and Petzold (1998). This can be written structurally:
ẋ 1 x2 x1
e1    
(6.19)
e2  0  

e3  0  

where e 1 represents the set of dynamic equations, e 2 represents the set of algebraic
equations, and e 3 represents the set of measurement equations. Since the the matrix E and
the Jacobian matrix ∂g ~∂x 2 are non-singular, a matching between the pair ˜˜e 1 , ˜ẋ 1 
and ˜˜e 2 , ˜x 2  exists, which results in a matching between the pair ˜˜e 1 , e 2 , ˜ẋ 1 , x 2 
that also exists, independently of the third equation set e 3 .
Next step is to make a Dulmage-Mendelsohn decomposition of the system (where ẋ 1
and x 1 are lumped together) in (6.18). The rows and columns in the Dulmage-Mendelsohn
composition can be permuted to obtain the specific structure:
ẋ 11 x 21 ẋ 12 x 22
e 11     

e 12  0   
(6.20)
e 21  0 0  

e 22  0 0 0 

e3  0 0 0 

where the variables x 1 and x 2 are split into two parts:

x1 ˜x 11 , x 12 
x2 ˜x 21 , x 22 
88 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

X0 X
0

200 M0
φ1

φn
400
∆ΓC1
∆η C1
∆ΓT1
equations

600 ∆η T1

M
800
fm f

f PA
1000
fnA

0 200 400 600 800 1000

variables

Figure 6.6: In the figure, the structural model of the gas turbine diagnosis model using the
Dulmage-Mendelsohn decomposition is shown. The equations where the actuator faults
and health parameters are injected are shown with arrows in the figure. The structural model
is divided into two parts, where the first part is the exactly-determined M 0 part and consists
of the first 390 equations. The second part is the over-determined M  part and consists of
the remaining equations. It is not necessary to considered the M 0 part for diagnosis purposes
since the absence of redundancy and this part can be removed in the diagnosis tests. Since
all the considered faults and health parameters appear in the over-determined part they are
detectable. Variables that appear in the exactly-determined part is, e.g., φ i which is the angle
of the shaft from (5.24b).
 6.4. Structural Analysis 89

X
400 ∆η C1
∆ΓC1
∆η T1
∆ΓT1
f PA
fnA

600
equations

M
800

fm f

1000

400 600 800 1000

variables

Figure 6.7: In the figure, the over-determined part of the structural model in Figure 6.6 is divided
into a number of equivalence classes. The equations where the actuator faults and health
parameters are injected are shown with arrows in the figure. Since faults f PA and f n A appear
in the same equivalence class they are not isolable. All the other fault modes may be isolable
since they appear in different equivalence classes. Since many (eight) sensors are used, the
equivalence classes get small. For a smaller number of sensors, the equivalence classes get
larger.
90 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

The equation set ˜e 21 , e 22 , e 3  in (6.20) is a redundant set while the equation set ˜e 11 , e 12 
has no redundancy. Since the original system (6.19) has a matching in ˜˜ e 1 , e 2 , ˜ẋ 1 , x 2 
and matching of variables and equations is not affected by row and column permutations,
also ˜˜e 11 , e 12 , e 21 , e 22 , ˜ẋ 11 , x 21 , ẋ 12 , x 22  has a matching. Since the third quadrant has
only zeros, a matching in the forth quadrant must exist. Therefore, it is possible to find a
matching for the pair ˜˜e 21 , e 22 , ˜ẋ 12 , x 22  and the sub-system is an index-1 DAE.
j

Remark 6.4.2 A matching between the pair ˜˜e i , ˜x i  means that it can be possible to
calculated the variable x i using the equation set e i . If a matching does not exist, the un-
known variables x i cannot be calculated using the equation set e i . In a redundant equation
set, unknown variables can be matched in several ways.

6.4.4 Diagnosis Test Equation

The diagnosis test equations are chosen using the Dulmage-Mendelsohn decomposition
of the index reduced diagnosis model (6.15). The over-determined M  part of the system
(6.15) is:

E ẋ 1 f˜‰x 1 , x 2 , uŽ (6.21a)
0 g̃ ˆx 1 , x 2 , u (6.21b)
y h̃ ‰x 1 , x 2 Ž (6.21c)

where the algebraic constraints can be written, at least locally, x 2 G ˆx 1 , u since

the system is in an index-1 DAE form. The state space form is obtained through an
elimination of the algebraic variable x 2 and an inversion of the matrix E:

ẋ 1 f ˆx 1 , u 
(6.22)
y h ˆx 1 

where x 1 represents the state variables, u represents the input signals, and y represents
the measurement signals. Before the observers are constructed in the Section 6.6, the
observability of the model is necessary to check.

6.5 Observability Analysis

If observer based techniques should be considered when the diagnosis tests are designed,
it is necessary that information about the state of the system is available in the mea-
surements, i.e., the system should be observable. It is strictly speaking not necessary
that the system is required to be observable. An observability requirement could be
replaced by a detectability requirement, i.e., any non-observable modes are stable, but
here observability will be analyzed.
An observability analysis of the index reduced diagnosis model (6.16) shows that
unobservable modes are present. To determine if a nonlinear system is observable is
6.5. Observability Analysis 91

in general a difficult question (Hermann and Krener, 1977). A number of different

observability criteria is available in, e.g., Nijmeijer and Fossen (1999). A starting-point is
to check if the nonlinear system is locally observable through a linearization in a suitable
operating point ˆx 0 , u 0  of the system in (6.22). The system matrices are:
df df dh
A V ,B V ,C V . (6.23)
dx ˆ x 0 ,u 0  du ˆ x 0 ,u 0  dx ˆ x 0 ,u 0 

The linearized system, near the point ˆx 0 , u 0 , is written:

ẋ Aˆx  x 0   Bˆu  u 0  (6.24a)
y C ˆx  x 0   y 0 (6.24b)
For the linear system in (6.24), the observability can be checked according to the observ-
ability matrix OˆA, C  in Theorem 6.5.1.

Theorem 6.5.1 A pair A > Rn n , and C > Rm n is observable if and only if the observ-
ability matrix
< C =
@ A
@ A
@ CA A
OˆA, C  @ A
@  A
@ A
@CA p1 A
> ?
has full rank n, for a p B n.
Proof See, e.g., Kailath (1980). j

6.5.1 Structural Observability

To handle numerical problems, the structural observability presented in Shields and
Pearson (1976) is investigated. Since the structure of the model is used, the method can
only provides a necessary condition for the observability. Since the model is physical
based, it can be assumed that structural observability implies observability. The structural
observability criteria can be stated:

Theorem 6.5.2 A pair A > Rn n , and C > Rm n is structural observable if and only if the
generalized observability matrix
<I A 0 ... 0 0 =A
@
@0 I A ... 0 0 AA
@
@ A
@     A
@ A
@0
@
0 0 ... I AAA
Os ˆA, C  @0 0 0 ... 0 C AA
@
@
@0 0 0 ... C 0 AA
@ A
@     A
@ A
@0
@ C 0 ... 0 0 AA
@C 0 0 ... 0 0 A?
>

with dimension n 2  nˆm  1  n 2 has structural rank n 2 .

92 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

Proof See Shields and Pearson (1976). j

The structural rank of a matrix can easily be checked using graph theoretical algorithms
for matching in bi-partite graphs, e.g., the Dulmage-Mendelsohn decomposition. In
Matlab, the sprank command can be used.

6.5.2 Removing of Unobservable Modes

For the gas turbine diagnosis model, the linearization of the system in (6.16) has un-
observable modes, which can be checked using the structural observability method.
Thus, if an observer should be constructed, these unobservable modes need to be re-
moved. As a consequence, when the exactly-determined part M 0 is removed, a number
of unobservable modes also disappear, which is stated in Theorem 6.5.3.

Theorem 6.5.3 Consider the semi-explicit index-1 DAE system (6.15). A state variable
that appears in the exactly-determined M 0 part of the structural model is unobservable.

Proof According to Theorem 6.4.1, the semi-explicit index-1 DAE system can be written
structurally:
ẋ 11 x 21 ẋ 12 x 22 x 11 x 12
e 11       

e 12  0     
(6.25)
e 21  0 0   0 

e 22  0 0 0  0 

e3  0 0 0  0 

where x 1 ˜x 11 , x 12 , and x 2 ˜x 21 , x 22 . In (6.25) a Dulmage-Mendelsohn decom-

position of the DAE system (where ẋ 1 and x 1 are lumped) is performed together with
an extra rearrangement of the rows and the columns to obtain the specific structure.
According to Theorem 6.4.1, a matching of the pair ˜˜e 11 , e 12 , ˜ẋ 11 , x 21 , and the pair
˜˜ e 21 , e 22 , ˜ẋ 12 , x 22  exists. The system in (6.25) can be written analytically:

ẋ 11 f 1 ˆx 21 , ẋ 12 , x 22 , x 11 , x 12  (6.26a)
x 21 g 1 ˆẋ 12 , x 22 , x 11 , x 12  (6.26b)
ẋ 12 f 2 ˆx 22 , x 12  (6.26c)
x 22 g 1 ˆx 12  (6.26d)
y h 1 ˆx 22 , x 12  (6.26e)

Substitution of the algebraic constraints (6.26b) and (6.26d) into dynamic equations
(6.26a) and (6.26c) gives:

ẋ 11 f˜1 ˆx 11 , x 12  (6.27a)
ẋ 12 f˜2 ˆx 12  (6.27b)
y h 2 ˆx 12  (6.27c)
6.5. Observability Analysis 93

System in (6.27) is linearized to give the matrices A and C:

x 11 x 12
x 11 x 12
 
A Œ ‘, and C ‰ 0  Ž
0 

which gives:
CAk ‰0  Ž

for k C 0. The observability matrix O in Theorem 6.5.1 cannot have full rank, so differen-
tiated states that appear in the exactly-determined part are not observable. j

When the M  part is considered, unobservable modes which appear in the exactly-
determined part are removed according to Theorem 6.5.3. It is verified that the lineariza-
tion of (6.22) is structurally observable, which indicates that observer based techniques
can be considered when the test quantities are constructed. The unobservable modes
which are removed are the angle variables φ i . When these state variables are removed,
the algebraic constraints g̃ 0 in (6.14b) can also be removed since they consist of only
angle dependencies. The angle variables were also shown in Figure 6.6.

6.5.3 Number of Health Parameters in the Model

An important question, when an diagnosis test is constructed, is how many health
parameters that can be introduced in the model. The number of health parameters affect
the observability of the model and the maximum number is equal to the number of
unique measurement sensors. This necessary condition is summarized in Lemma 6.5.4.
Where in the model the health parameters appear may also affect the observability.

Lemma 6.5.4 A pair A > Rn n , and C > Rm n is not observable if the number of health
parameters is larger then the number of measured states n c .
Proof Let x > Rn x , h > Rn h , and the number of measured states n c . The structural rank
of a matrix can never be smaller if nonzero elements are added to the element positions
of the matrix. This gives the opportunity to investigate the structural rank of the last
column in the matrix below:
<   =
@ A
@ ... 0 AA
@ 
@
@ ...  AAA
Õs @ A
@
@
...  C AA
@ ... 0 AA
@ 
@ A
>   ?

Assume that all element positions in the matrix pair (A, C), except for the health param-
eter states, are occupied according to:
x h
x h
 
A Œ ‘, and C ‰ 0Ž
0 0
94 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

where A > Rn n , C > Rn c n , and n n x  n h . If the structural rank of Õs should be n 2 ,

the last n columns need to have a structural rank n. Further, the last n columns can be
written:
< =
@  A
A @ A
 @0 0A (6.28)
C @ A
@
> 0 A?
where the zeros are omitted, and has the structural rank n x  n c . The pair ˆA, C  is not
observable if
nx  nc @ n (6.29)
which can be written:
nc @ nh
where n x n  n h is used. This gives the matrix pair ˆA, C  is not observable if n c @ n h .
j

Remark 6.5.5 The same arguments, as in Lemma 6.5.4, can be considered to verify that
two health parameters that appear in the same equivalence class in Figure 6.5 are not ob-
servable.

6.6 Filter Design

A common approach in the gas turbine diagnosis literature to estimate health deteri-
oration is to use tests based on observers. The observers are either linear (Volponi,
2003; Luppold et al., 1989) or nonlinear (Borguet and Léonard, 2008; Dewallef et al.,
2006; Rausch et al., 2007) Kalman filters. In these papers, the application is for in-flight
monitoring purposes but the focus in the present work is on industrial gas turbines.
For stationary gas turbines, the calculating capacity is assumed to be larger then for
an in-flight monitoring case. Thus, more sophisticated filters can be considered for
industrial gas turbines, and the main objective here is to design a nonlinear filter in the
form:

x̂˙ f ˆx̂, u  K ˆ y  ŷ (6.30a)

ŷ h ˆx̂  (6.30b)

where f and h are functions, u represents the input signals, y represents the measurement
signals, and K is the observer gain which can be considered as a design parameter. The
estimation error, or the residual, r y  ŷ is amplified through the K matrix. Thus, if
the observer gain K is large, the observer relies more on the measurements than on the
model equations, and if the observer gain K is small the observer relies more on the
model equations than on the measurements. This means that a large K results in state
estimates which is more sensitive to measurement noise than if a small K is used. On
the other hand, a small K corresponds to estimation errors where model uncertainties
get more significant. Thus, when the observer is designed, it is possible to tune between
a fast and noisy and a slow and filtered estimator.
6.6. Filter Design 95

A special type of a nonlinear Kalman filter are the so-called Constant Gain Extended
Kalman Filter (CGEKF), presented in Safonov and Athans (1978), where also the robust-
ness and stability of the filter concept are investigated. The CGEKF is a simplification of
the Extended Kalman Filter (EKF) Kailath et al. (2000). In the CGEKF, the observer gain
is calculated in advance for a given operating point and can be calculated offline which
is not the case for an EKF. Thus, the computational burden is smaller for a CGEKF than
for an EKF since the Jacobian of the system (6.30) is not calculated in each time step. In
the gas turbine application papers Kobayashi et al. (2005); Sugiyama (2000) the CGEKF
methodology is considered. In these papers it is demonstrated that the CGEKF can
be used in a wide operating range for an in-flight aircraft engine diagnosis application.
These studies indicate that it is not necessary to calculate a new Kalman gain for each
operational point during a flight program.
In the present gas turbine application, the CGEKF methodology is investigated since
the observer has an attractive form which is suitable for real time application for a good
choice of the operating point. Usually works a CGEKF well for a suitable choice of the
operating point where the observer gain is calculated. The advantage is that the observer
gain can be determined off-line which is attractive for an estimator that is used in a real
time application.
The starting point of the observer construction methodology is the gas turbine
diagnosis DAE model defined in (6.21). How the equations, used in the filters, are selected
is shown in Section 6.4. Before a summary of the filter design procedure is presented,
the equations of the linear/nonlinear Kalman observers are shown and important steps
such as observability and observer tuning are discussed.

6.6.1 Kalman Filter

A linear dynamic system, expressed in state space form, can be written:

ẋ Ax  Bu  Gw
(6.31)
y Cx  Du  v

where A > Rn x n x , B > Rn x n u , C > Rn y n x , and D > Rn y n u are system matrices, G >
Rn x n w is a gain matrix of the process noise, u is the input signal, w is white process
noise, and v is white measurement noise. The noise signals satisfy:

E w ˆt w ˆsT  Qδ ˆt  s, E v ˆt v ˆsT  Rδ ˆt  s,

(6.32)
E w ˆt  E w ˆs  0, E v ˆt  E v ˆs 0,

for the dummy scalar variables t, s and the Dirac function δ. The Kalman filter equations
(Kailath et al., 2000) of system in (6.31), are:

x̂˙ Ax̂  Bu  K ˆ y  C x̂  Du

(6.33)
ŷ C x̂  Du  v

The Kalman gain K is:

K PC T R1 (6.34)
96 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

where P is the solution to the Riccati equation

Ṗ AP  PAT  GQG T  PC T R 1 CP  Q (6.35)

for the given noise matrices Q and R, and for a given initial probability condition
T
P ˆ0 E
‰x ˆ0  x̂ ˆ0ˆx ˆ0  x̂ ˆ0Ž .

6.6.2 Stationary Kalman Filters

The stationary Kalman filter is obtained by setting Ṗ 0 in (6.35).

6.6.3 Nonlinear Kalman Filters

A nonlinear dynamic system can be written in the form:

ẋ f ˆx, u, w 
y g ˆx, u  v
(6.36)
w  ˆ0, Q 
v  ˆ0, R

where the linearization

∂f ∂g ∂f
A V , C V ,G V
∂x x̂ ,u ∂x x̂ ,u ∂w x̂ ,u

is applied when the Kalman gain K is calculated according to the Riccati equation in
(6.34). Then it is possible to construct the two nonlinear Kalman filters EKF and CGEKF.
These filters can be written:

1. Extended Kalman Filter (EKF):

x̂˙ f ˆx̂, u  K ˆx̂, uˆ y  h ˆx̂, u

(6.37)
ŷ g ˆx̂, u
where the Kalman gain K ˆx̂, u is updated in every time step according to (6.34).

2. Constant Gain Extended Kalman Filter (CGEKF):

x̂˙ f ˆx̂, u  K ˆ y  h ˆx̂, u

(6.38)
ŷ g ˆx̂, u

where the Kalman gain K is calculated for the stationary Kalman filter in one
operational point according to (6.34).
6.7. Parsers for an Automatic Extraction of Sub Systems 97

6.7 Parsers for an Automatic Extraction of Sub Systems

Since the number of variables and equations in the gas turbine model is large, a systematic
method to analyze and select relevant equations automatically is attractive to have. The
number of equations and variables is more than 1000 as shown in the structural model in
Figure 6.6, which is too large for manual processing. It should also be easy to investigate
various types of diagnosis setups such as models with different number of sensors and
health parameters. Those analyses are difficult to perform in the Dymola environment,
thus a parser for model exportation is necessary to have.
To analyze model properties, the Matlab environment is preferably used since it
has available routines for, e.g., structural methods, model linearization, and Kalman
gain calculation. Structural methods are, e.g., the Dulmage-Mendelsohn decomposition
(dmperm.m), and the equivalent class construction. Another motive to use Matlab is
that filtering and other signal manipulations are easier to make than in the Dymola
environment. The final aim with the parsers is to transform the flat Modelica gas turbine
model to a number of test quantities based on Kalman filters. These test quantities have
the form shown in (6.30) where f and g are Matlab m-functions.
In this section, parsers used for: (i) DAE model extraction, (ii) structural model
construction, (iii) DAE-index reduction, and (iv) state space form construction are
presented. The name parser is used since most of the time there are text strings that are
manipulated. For the equation manipulation, the MuPAD symbolic engine is called from
the implemented parsers. The MuPAD symbolic engine is a part of the Symbolic Math
Toolbox contained in Matlab. In this work, Matlab version 7.11 and MuPAD version 5.5
are used.

6.7.1 Automatic Extraction of the DAE Model

To fulfill the requirement, a parser that can export the diagnosis model implemented in
Modelica to the Matlab environment is developed. The parser is relatively simple and
can only handle a restricted functionality but the functionality is sufficient to be used
together with models constructed in the GTLib package. After the model is exported,
the MuPAD symbolic engine is invoked to handle the symbolic transformation of the
equations and variables according to the methods described in this section. A brief
description of the DAE extraction parser is summarized in:
1. Before any symbolic transformation of the model is done in Dymola, it is possible to
save the flat Modelica model. The flat Dymola model is a text file that is called .mof,
and contains information, such as equations, variables, parameters, constants,
functions, etc. All this information is needed when the model is simulated outside
the Dymola environment. In the .mof-file, all the object oriented hierarchies are
removed, so the file consists of pure equations and variables. Hence, the .mof-file
is the input to the DAE extraction parser.
2. All functions that are called from the Dymola model appear in the beginning of the
.mof-file. These functions are for example the NASA polynomials that describe
the gas properties in a specific gas, and the functions that are implemented in
98 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

GTLib which are mentioned in subsection 5.3.3. All of these functions are removed
from the .mof-file and converted to Matlab functions for future investigations.

3. Next step is to find, save, and remove variables, parameters and constants from
the file. The values of the parameters, the constants, and the initial values of the
variables are saved. The variables appear with different declarations such as Real,
Integer, and Boolean which are handled separately. These variables can also appear
as the type input and output which also have to be handled separately.

4. Equations that consist of if -statements have to be treated with care, since the global
equation system changes with the if -condition. For example, a model with six
if -statements (that doesn’t has any elseif part) gives 26 64 different equation
systems. This can be difficult to handle in practice, so here it is necessary to state the
condition of the if -statements. Fortunately, most of the if -statements in the model
come from the flow direction of the gas. Thus, it is assumed that the direction is
known, which transforms respective if -statement to only one equation.

5. Now the .mof-file consists of only pure equations, and the final step is to interpret
these equations to a format that is familiar by the MuPAD symbolic engine. The
earlier saved variables, parameters, and constants have also to be converted to a
format that the MuPAD symbolic engine can understand.

The output of the DAE extraction parser is a differential algebraic equation system with
attached variables, and given input, and output signals. The output DAE system can be
treated by the MuPAD symbolic engine available in the Matlab environment.

6.7.2 Structural Model Parser

The structural model of the DAE model can easily be acquired through a variable find
approach by the MuPAD symbolic engine. The acquired DAE system in the previous
step consists of a number of functions. These functions can be look up tables, algorithms,
and mathematical functions. In the structural model, it is not necessary to consider
the internal variables in each function. Therefore, it is sufficient to only consider the
input and the output variables for each function. The input and the output signals in the
analytical model, as can be seen in Figure 6.2, are not considered in the structural model
since they are known quantities.
The output of the structural model parser is the structural model of the DAE model
that is given as an input argument to the parser.

6.7.3 Index Reduction Parser

The structural model together with the DAE model are inputs to the index reduction
parser. Actual rows in the equation system that are needed to be differentiated are deter-
mined with Pantelides algorithm (Pantelides, 1988). These rows are differentiated, and
the underlying equations are saved to get proper initial conditions to the differentiated
equation system.
6.7. Parsers for an Automatic Extraction of Sub Systems 99

The output from the Index Reduction Parser is an equation system that has a dif-
ferential algebraic index equal to 1. The system is also transformed into a form that the
E-matrix in (6.15) has full rank according to the variable elimination procedure described
in Section 6.3.1.

6.7.4 State Space Form Construction Parser

The intention with this parser is to easily construct a set of runnable Matlab functions that
can be used to simulate the index reduced equation system defined previously. For the
parser, it is assumed that the input equation system is an index-1 DAE (which has no re-
dundancy), and can be written in a semi-explicit form. The overall Dulmage-Mendelsohn
decomposition of the system is shown in Figure 6.8, where the set of equations and vari-

ẋ 1 x2 x1
e1 b1

e2 b2

Figure 6.8: Dulmage-Mendelsohn decomposition of a system in semi-explicit form where ẋ 1 and

x 2 are considered as unknown variables. In each step, when the system is solved, x 1 is the
previous state of the system and hence it can be considered as an input signal.

ables ˜e 1 , ẋ 1  is the differential part, and ˜e 2 , x 2  is the algebraic part of the system. This
decomposition is done because in each time step for the solver, x 1 is the previous state of
the system and hence it can be seen as an input signal to the differential part b 1 .
For the algebraic part b 2 in Figure 6.8, the Dulmage-Mendelsohn is once again
utilized, and gives the results shown in Figure 6.9. To solve the overall system, variable
x̃ n is first calculated. This result is utilized in the next equation set ẽ n 1 and x̃ n 1 is
then calculated. This solving process goes on until the last variable x̃ 1 is calculated.
If the set ẽ i consists of two or more equations, it is said that c i is a strongly connected
component. This means that the variables contained in c i cannot be substituted without
transformation, i.e., an inversion of the component is necessary. For linear systems, this
can be done easily since Gaussian elimination can be applied to obtain a substitution
chain, similar to the structural model. For a nonlinear strongly connected component, a
numerical solver can preferably be used since an analytic solution can be difficult to find.
Another interpretation of the behavior of strongly connected components is to consider
a bi-partite graph, which consists of loops or cycles, see, e.g., Blanke et al. (2003). A
100 Chapter 6. Modeling, Analysis, and Transformation of the Diagnosis Model

x̃ 1 x̃ 2 ... x̃ n 1 x̃ n
ẽ 1 c1

ẽ 2 c2

 

ẽ n 1 c n 1

ẽ n cn

Figure 6.9: In the figure, the Dulmage-Mendelsohn decomposition of the pure algebraic compo-
nent b 2 from Figure 6.8 is shown. The arrows in the figure illustrate the solution path.

description of the state space form construction parser is summarized in the pseudo
algorithm below:

1. First, the equation system is divided in two parts as illustrated in Figure 6.8. The
b 1 and b 2 components are located through a DM decomposition, where the sets
of unknown variables are ẋ 1 and x 2 . The variable x 1 is considered as known since
the previous state of the system is known.

2. The algebraic part b 2 is treated separately, and the structural model is considered
when the substitution chain is determined as Figure 6.9 illustrates.

3. For each component c i , the parser determines if the component consists of only
one equation. If this is true, the MuPAD symbolic engine is called and tries to find
an analytic solution, i.e., a pure substitution.

4. If the MuPAD symbolic engine failed to find an analytic solution, or if the com-
ponent is strongly connected, a nonlinear Matlab solver is incorporated in the
generated file.

5. After the variables in x 2 are calculated, they are inserted in the differential part
of the system, together with the previous state x 1 . Finally, the state of the system
in the next time step can be calculated with a numerical ODE-solver after the
component b 1 in Figure 6.8 is inverted.

The outputs from the parser are the two .m-files f and G that can be interpreted as:

x 2,n 1 ‡ 
Gˆx 1,n 1 , x 2,n (6.39a)
1
ẋ 1 fˆx 1,n 1 , x 2,n 1  (6.39b)
6.8. Conclusion 101

where index n represents the actual time step, and ˆ‡ represents an initial guess of
the iteration variable (i.e., the previous value of the iteration variable). Input argument
x 2,0 in (6.39a) is needed by the nonlinear solver to get an appropriate starting point.
Expression (6.39b) can then be called by an ordinary differential equation solver in
Matlab.

6.8 Conclusion
In the chapter, a gas turbine diagnosis model is presented that can be used when a
FDI-system is designed. In the diagnosis model, faults and estimation parameters used
for diagnosis purposes are introduced. The faults are introduced in the actuator and
sensor signals, and the estimation parameters are called health parameters which should
compensate for gradual deviation in performance.
The diagnosis model is analyzed for: (i) DAE-index, (ii) observability, and (iii) struc-
tural properties. The test equations which are used in the FDI-system are chosen using
structural methods. Here, the whole over-determined part is chosen. The observability
analysis shows that unobservable state variables are removed using the selected set of
equations. The split into equivalence classes shows that faults in the input signals of
power and speed of the applied generator is not isolable. All the other fault modes and
health parameters may be detectable and isolable. To make the analyses, a number of
parsers is developed. These parsers take the Modelica gas turbine diagnosis model and
convert it into runnable Matlab code.
 Part III

Application of Methodology
 Chapter 7

Performance Estimation in Industrial Gas

Turbine Engines

The performance of an industrial gas turbine degrade gradually due to certain factors
such as environment air pollution, fuel content, and ageing to mention some of the
degradation factors. Degradation due to compressor fouling can partially be restored
by an online/offline compressor wash. Therefore, it is important to supervise the degra-
dation to efficiently plan service and maintenance. The gas turbine fleet consists of a
lot of individuals, with different kinds of properties and configurations, that have to
be monitored by the service engineers. Therefore, it is desirable that it should be easy
to construct and evaluate test quantities and collect them in the FDI-system. When
the diagnosis test quantities are constructed from a component based model, it should
be easy to, e.g., replace components or change parameter values depending of the su-
pervised engine. The main objective of this chapter is to apply the methodology for
diagnosis test quantity design presented in Chapter 6. The diagnosis tests are constructed
in an automatic manner directly from the gas turbine diagnosis model. The diagnosis
tests can then be used for performance estimation, or as smaller part in a more general
FDI-system.
The chapter starts with an introductory evaluation where three basic methods to
detect compressor fouling is presented. These methods involve only the measurement
signals and not the gas turbine model. In Section 7.3, the well known Measurement Delta
method is investigated using the experimental data where the focus is to monitor trends
in the deltas. The deltas is generally the difference between a simulated and a measured
gas path quantity. In Section 7.4, the nonlinear Kalman filter is designed and generated
for the SGT-700 gas turbine model. The filter is then investigated in Section 7.5 where
three case studies are presented. In the first case study, simulated data from the reference
platform is evaluated for different operational points and different atmospheric weather
conditions. In the remaining case studies, experimental data is evaluated. In evaluation 2,
the focus is on state estimation for control, where the flame temperature is supervised

105
106 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

and compared for two filters. In evaluation 3, the supervision is presented according to
detection of a fouled compressor.

7.1 Background
Estimation of unmeasured gas turbine performance parameters can be used in, at least,
two contexts: (i) monitoring, and (ii) control. In monitoring applications, a typical
objective is to study deteriorations in gas turbine components, e.g., compressors and
turbines, which affect the estimated performance parameters. Therefore, in monitoring
applications it is of great importance to study the deviation from a nominal baseline
to decide whether the engine indicates a faulty behavior or not using the gas path
measurements. Monitoring of gas turbines, and especially aircraft engines, is a widely
studied topic in the gas turbine diagnosis literature (Volponi, 2014; Luppold et al., 1989;
Kobayashi and Simon, 2005; Doel, 2003). For control applications, model based feedback
control strategies are studied in Watts et al. (1992); Härefors (1997). In these papers, the
control strategies are based on linear models which are derived from a thermodynamic
engine model. In the controller design, the component deteriorations are not considered
which can result in an unnecessary overall reduction in the gas turbine performance.
An important signal, used in the controller, is the combustion flame temperature. Since
this signal is not measured, the signal needs to be estimated, e.g., by using a model and
thereby may the temperature be affected by the component deterioration.
In papers Diakunchak (1992); Kurz et al. (2009); Kurz and Brun (2001); Brekke et al.
(2009), several mechanisms that cause degradation in gas turbines are presented. The
major contribution of degradation in industrial gas turbines is fouling which is caused by
small particles and contaminants in the air. These particles increase the roughness of the
rotor and stator surface. Another degradation effect is tip clearances which is a common
diagnosis for older gas turbines. Tip clearances denotes an increasing gap between
the rotating blades and the stationary casing. Fouling due to increased roughness can
partially be restored by washing the compressor, while a component replacement is
often needed for tip clearances. In the last paper, performed by Brekke et al. (2009),
deterioration effects due to compressor fouling are investigated in an offshore industrial
gas turbine application.
A common solution, about how to estimate deviation in performance from a nominal
baseline is to introduce health parameters (Luppold et al., 1989; Kobayashi and Simon,
2003; Borguet and Léonard, 2008). In the two first papers, Kalman filters are used to
estimate the introduced health parameters. The degradation in performance is natural,
so if the model does not compensate for this degradation it can be difficult to avoid
sensor false alarms. In Kobayashi et al. (2005), a nonlinear Kalman filter is demonstrated
that can be used in a wide operating range for an in-flight aircraft engine diagnosis
application, providing an engine model with high accuracy.
One factor that can affect the model accuracy, and especially the performance of the
health parameters, is the absolute humidity in the atmospheric air. The humidity effects
for an in-flight application are studied in, e.g., Bird and Grabe (1991) where methods
based on parameter correction are considered. In Mathioudakis and Tsalavoutas (2002),
7.1. Background 107

a study is performed where humidity effects are investigated in an industrial gas turbine
application. An analysis is presented of how the variation in ambient conditions affects
the health parameter estimations. It is shown that a compensation for the ambient
conditions reduces the undesirable daily variations in the estimated health parameters.
In other gas turbine papers, the variation in absolute humidity is often neglected since
it increase the complexity of the model and does not vary significantly much at many
industrial gas turbine sites.
A fundamental effect of the change in absolute humidity is the change of molecules
in the ambient air media. A change in the concentration of molecules in the media
affects the thermodynamic gas properties such as enthalpy, heat capacity, and entropy.
The thermodynamic properties influence the estimated performance. In the GTLib
framework, as described in Section 4.6.5, the concentration of the molecules in the gas
path media is changed immediately according to the change in ambient conditions.
The change in ambient conditions can be encapsulated by the developed estimators
to reduce undesirable daily variations in the estimated health parameters mentioned
in Mathioudakis and Tsalavoutas (2002). Therefore, an evaluation is performed in
Section 7.5.1 where the effect of the absolute air humidity is investigated using simulated
data.

7.1.1 Experiment Setup

For the case studies, measurement data are collected from a gas turbine site in the Middle
East. Because of the difficult environmental conditions at the site, the compressor is
washed frequently. At the site, the air consists of a high grade of air pollutions such as:
sand, salt, oil, and other contaminations which affect the engine’s performance. The
gas turbine SGT-700 at the current site is a 1-spool and a 2-shafted engine where the
application is the so-called mechanical drive. A notable aspect with this specific gas
turbine model is the absence of an instrumentation sensor between the gas generator and
the output of the power turbine. The lack of these types of sensors makes the diagnosis
and monitoring procedure more difficult, since no measurement signals are available in
the gas path between the output of the compressor and the output of the power turbine.
Similar gas turbines, launched by other manufactures, have thermocouples between the
gas generator and the power turbine. Having ideal thermocouples in that cross-section
should reduce the uncertainty of the gas path parameters in the gas generator. In the
SGT-700, the thermocouples are instead installed in the exhaust gas at the output of the
the power turbine.

Gas Turbine Site

A schematic view of the gas turbine site is shown in Figure 7.1 where the gas turbine and
its surrounding components are presented. The dashed arrows in the figure represent
physical connections, i.e., mechanical, thermodynamic, and electrical connections. Solid
arrows represent signals to and from the controller.
The input signals to the controller are the measurement of the atmospheric air,
the gas path instrumentation sensors, the power generated by the application, and the
108 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

mfuel Starter
Fuel System
Motor

Actuators
Controller

Gas Turbine Application

Ambient

Sensors

Figure 7.1: In the figure, a schematic view of the gas turbine experimental platform is shown. The
dashed double arrows represent a physical connection, while solid arrows in the figure represent
ordinary signals. In, e.g., Modelica, physical based connections are represented by equations
and ordinary signals have only one flow direction.

desired rotational speed of the power turbine. The atmospheric air sensors measure the
pressure p 0 , the temperature T0 , and the relative humidity φ 0 of the ambient air. The
instrumentation sensors measure pressures, temperatures, and shaft speeds along the
engine’s gas path. In Section 6.2, the exact positions of these sensors are described. Finally,
the generated power by the application is not measured. Instead, the signal is estimated
using internal sensors in the driven application component. Hence, the power estimation
procedure does not utilize any of the gas path measurements, so a sensor fault in some of
the gas path measurements does not affect the reliability of the power estimation. Thus,
the power signal is here considered as a reliable non faulty measurement signal.

As shown in Figure 7.1, a measurement signal of the mass flow of fuel is available.
This measurement signal does not belong to the standard site equipment since a mass
flow meter is expensive for the customer to install. Thus, the mass flow meter is only
installed on request by the customer. The controller use this signal (multiplied with the
heat value of the fuel) to, e.g., calculate the flame temperature. When the mass flow meter
is absent, the mass flow can be calculated using a model of the fuel system together with
the measurement of fuel pressure and fuel valve positions. At sites where a mass flow
meter is available, more information about the operation is available. This information
can be included into the diagnosis system where, e.g., the estimated mass flow signal can
be compared with the measured signal by the mass flow meter.
7.1. Background 109

One Year of Experimental Data

The considered sequence of available experimental data came from one year of operation.
During the operation, the gas turbine is started and stopped a number of times. These
startup and shutdown sequences are removed from the measurement sequence when
the test quantities are evaluated since the diagnosis model is not valid during startup and
shutdown. The original data have a sample time of 5 min. To reduce the number of data
points (and decrease the evaluation time), the data are re-sampled with a sample time of
60 min. During the operation, the compressor is washed five times (middle of November,
end of December, end of March, end of June, and middle of September). The time points
of the compressor washes are shown in Figure 7.4 and Figure 7.9. In Figure 7.4, the washes
are marked with arrows.

Preliminary Investigation of the Engine’s Nominal Health Status

A preliminary study is performed where the focus is to determine the health status of
the clean engine, i.e., calculate the health parameters directly after a compressor wash.
With the methodology described in Section 6.6 a constant gain extended Kalman filter
(CGEKF) is designed. The filter is tuned in that way so the response in the health param-
eters is fast. Then an approximate value of the health parameters after an compressor
wash is determined. The value of the health parameters are shown in Table. 7.1.

Table 7.1: Health parameter values for a clean machine.

Health Parameter Clean Machine [%]

∆η C1 -3.0
∆ΓC1 -1.0
∆η T1 +3.0
∆ΓT1 +10.0

At the present site, the engine is aged and due to the difficult environmental conditions
the leakage and tip clearance in the compressor-turbine T1 is probably larger than for the
reference machine which affect the performance negative. An increased tip clearance
results in a lower turbine efficiency and a higher mass flow (Kofskey and Nusbaum, 1968).
As Table 7.1 indicates, the value that differs most from the nominal value is for the flow
capacity C T of the compressor-turbine T1 (which is proportional to the mass flow). To
have all health parameters in the same range, the reference value C T ,re f in (5.2d) for the
compressor-turbine T1 is increases with 6 %, which decrease the health parameter ∆ΓT1
with 6 percent points. This value is also suggested by SIT for the present engine.
Since the exact positions of the sensors along the engine’s gas path are not known and
the actual absolute pressure can differ from the nominal model. Thus, a basic idea is to
add a constant bias to the measurement signals to compensate for individual properties
in the gas turbine fleet and to get health parameters in the same interval for a newly
manufactured and cleaned machine.
The bias terms are calculated once for the cleaned gas turbine. In the experimental
data sequence that is investigated, the bias terms are calculated directly after the first
110 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

compressor wash is performed. The bias term b i of measurement sensor i is calculated:

nom me as
k yi,j  y
bi Q i,j
k
j 1

where k is the number of samples that corresponds to two days of operation, ymeas is
the measured quantity, and y nom is the reference value calculated from the nominal
model similar to the description in Figure 7.3. The values of the bias terms, in percent of
the nominal value, are shown in Table. 7.2. A positive value in the table means that the
measured signal is lower than nominal value.

Table 7.2: Bias in measurement sensors.

yi y p1 y p3 y p8 y T2 y T3 y T7 y n C1 y n T0
b i % 1 5 1 0 0  1 0 0

The biases are added to the measurements according to:

yi y i ,meas  b i

where b i is the constant sensor bias calculated once. In the present case, the bias compen-
sation is more or less negligible for all sensors except for the discharge pressure sensor
p 3 after the compressor. For the sensor measuring p 3 , the bias compensation is about
5 % from the nominal reference value (after the parameter C T ,re f is increased). The large
difference between the measurement and the nominal value for the discharge pressure
does not depend on the actual sensor calibration since the discharge pressure is measured
with three sensors where all the sensors show values in the same interval. An idea is
that the difference in pressure came from leakages in the compressor turbine due to tip
clearances.

7.2 Introductory Methods To Detect Compressor Fouling

The objective with this section is to investigate three simple methods that can be used
to detect compressor fouling. The three first methods (a)–(c) are based on pure mea-
surement signals, which means it is simply to make an investigation of fouling with
these methods. The two first methods (a)–(b) are based on estimation of mass flow
through the compressor using different sensors. The fourth method (d) is based on the
performance model, and is more complex to evaluate and is presented here only for
comparison. Estimation using the performance model will be presented in Section 7.3.
All four methods are summarized in Table 7.3.
In all four methods, the objective is to find a so-called baseline. The baseline is
determined for a number of time samples of a cleaned compressor. It is necessary to have
the samples in a number of various operational points to get a proper baseline. Time
samples that are collected after the nominal condition of the compressor are assumed
7.2. Introductory Methods To Detect Compressor Fouling 111

Table 7.3: In the table, four methods to detect compressor fouling are presented. It is shown if the
method relies on the measurements and/or the physical model.

Measure- Physical Mass flow

ments Model Estimation
(a) Bell-Mouth Based Estimation x x
(b) Pressure Ratio Based Mass x x
Flow Estimation
(c) Power vs. Mass Flow of Fuel x
(d) Model Based Mass x x x
Flow Estimation

to appear below the baseline. When the distance between the samples and the baseline
is sufficiently large, it is time to wash the compressor. The baselines in the Figure 7.2
are only sketched by hand and the purpose with the introduced baseline is not to get
the perfect position of the baseline through minimization of some criteria. Instead, the
baseline should symbolize where a possible position could be in a simplified manner.
The measurement sequence that is used in the investigation of compressor fouling
detection is collected between two compressor washes. The length of the sequence is
about three months, where the first 40 samples start in January, and the last 40 samples
end in March. These samples are marked in Figure 7.2 together with all the considered
points. The first samples (clean compressor) should be used to span the baseline, and the
last samples (fouled compressor) should be used to detect the fouled compressor. In all
Figures 7.2(a)–(d), normalized quantities from (5.2) are used.

7.2.1 Bell-Mouth Based Estimation

In Diakunchak (1992), three methods to detect deterioration due to compressor foul-
ing are presented. One of these methods descends from an estimation of the mass flow
through the compressor, and a common approach to calculate the mass flow is to measure
the static pressure drop over the inlet duct of the compressor, i.e., over the bell-mouth.
To get a reliable estimation of the absolute mass flow amplitude, the bell-mouth measure-
ments need to be calibrated, together with the inlet pressure and the inlet temperature.
For monitoring purposes, it is enough to consider the relative changes in mass flow, and
according to Scott (1986), bell-mouth based measurement is a good technique to detect
compressor fouling. It is assumed that the mass flow is proportional to the square root
of the pressure drop over the bell-mouth. Thereby a decreased pressure drop over the
bell-mouth, for a given normalized rotational speed can be interpreted as an increasing
deterioration due to compressor fouling. To find when it is time to wash the compres-
sor, a baseline of a cleaned compressor need to be constructed. This can be done by
running the gas generator for a number of different rotational speeds, and then plot the
actual baseline for these speeds. Points that appear below this baseline indicate a fouled
112 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

compressor, and when these points have dipped to a certain value, it is time to wash the
compressor.
In Figure 7.2(a), the pressure loss measurement over the bell-mouth is plotted versus
the normalized speed of the gas generator. In the figure, it is shown that it can be difficult
to find a proper baseline, since the cloud of points are lumped, scattered, and cover nearly
the same area. Thus for this specific sequence the bell-mouth measurements are no good
to determine compressor fouling.

7.2.2 Pressure Ratio Based Mass Flow Estimation

For the case with a constant compressor pressure ratio together with fixed ambient
conditions, the compressor need to rotate faster to compensate for the degradation in
efficiency due to fouling (Kurz et al., 2009). At the same time, the compressor will also
consume more power and fuel. Therefore it is interesting to study the pressure ratio
versus the normalized rotational speed of the compressor. In Figure 7.2(b), the same
measurement sequence is plotted as for the bell-mouth measurement, but the compressor
ratio is instead considered. As Figure 7.2(b) shows, it is easier to find a baseline in the
pressure ratio plot than in the previous plot based on bell-mouth measurements. All
the points that should represent a fouled compressor appear in the lower interval in
Figure 7.2(b), which is desirable. It is undesirable that points that appear in the upper
interval are not representative for a clean compressor. Finally, a considered baseline is
sketched in the figure together with all available points.

7.2.3 Power versus Mass Flow of Fuel

In Figure 7.2(c), the power generated by the application is plotted versus the mass flow of
fuel. This test case can be intuitive when it is assumed that the overall efficiency of the
engine and the heat value of the fuel are nearly constant for all operation points. For this
test case, the points that symbolize a fouled compressor appear in the lowermost layer
seen from the baseline.

7.2.4 Performance Model Based Mass Flow Estimation

According to Meher-Homji (1987), a mass flow meter is preferred against a mass flow
estimation based on bell-mouth measurement, as an indicator of fouling. For the case
where the mass flow through the compressor is not measurable, model based techniques
can be utilized to calculate the actual mass flow. These techniques utilize the thermo-
dynamic heat and mass balances. Here, the performance model constructed in GTLib
package is simulated with the actual input signals shown in, e.g., Figure 6.3, to determine
the mass flow of air. The result of the study is presented in Figure 7.2(d), and shows good
result where all samples before the compressor wash appear in the lowermost layer seen
from the baseline. The data points which represent fouled and cleaned conditions are
also separated.
7.2. Introductory Methods To Detect Compressor Fouling 113

d p BM kPa Π 
10 13.5
baseline

13

12.5
Whole sequence (cleaned fouled)
40 samples after a wash (cleaned)
40 samples before a wash (fouled) ‡
nnorm
8 ‡
12
0.93 0.94 0.95 0.96 nnorm 0.93 0.94 0.95 0.96
(a) Bell-mouth pressure loss versus normalized speed of (b) Pressure ratio versus normalized speed of the
the compressor. compressor.

P ~Pre f  mflow,norm 
0.80 0.90
baseline baseline

0.75

0.85

0.70

‡
mfuel,norm ‡
nnorm
0.95 0.80
0.80 0.85 0.90 0.93 0.94 0.95 0.96
(c) Generated power by the application versus the (d) Normalized mass flow of air through the compressor
normalized mass flow of fuel. versus the normalized speed.

Figure 7.2: In the figure, four techniques to detect a fouled compressor are presented. The con-
sidered measurement sequences are collected between the two compressor washes that are
performed at the beginning of January and at the end of March. All presented signals are
based on measurements except the signal of the normalized air flow rate mflow,norm through
the compressor. The mass flow of air is calculated by the gas turbine model, which is based on
the mass and heat balances. The normalized quantities are defined in (5.2). The objective with
the study is to find a baseline that can be determined for a clean machine (green points) for a
number of different operational points. In the figures, the baseline is only sketched by hand to
illustrate the principle. Before the compressor wash, the measurement points should be below
the baseline (red points). According to this principle, (d) gives the best result and (a) gives the
worst result.
114 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

7.3 Measurement Delta

In Urban (1972); Urban and Volponi (1992); Simon et al. (2008), the so-called Measure-
ment deltas are introduced. A brief introduction of these measurement deltas is available
in the gas path analysis Section 6.1.1. The measurement deltas are an important part of the
gas path analysis (GPA). The deltas are assumed to capture the deviation from a nominal
baseline for a given number of known signals (or combination of known signals). For
monitoring purposes, the deltas can be used to supervise trends or detect abrupt changes
that can indicate a unhealthy behavior of the engine. Since the reference model, which is
considered herein, consists of a number of corrected parameters and performance maps
the delta calculation in Figure 6.1 can be replaced with the diagram in Figure 7.3.

y i ,meas


u Reference  100
Engine Model Σ y i ,nom
∆y i
y i ,nom

Figure 7.3: Measurement delta ∆y i calculation of the measured quantity i.

The measurement deltas are written:

y i ,meas  y i ,nom
∆y i 100 (7.1)
y i ,nom

where y i ,meas represents the measured value, and y i ,nom represents the nominal value of
the measured quantity i. The results of the delta calculation (of one year of experimental
data) for the measured discharge temperature ∆T3 of the compressor and the exhaust
gas temperature ∆T7 after power turbine are shown in Figure 7.4 and Figure 7.5.
It is possible to construct deltas for other types of quantities, e.g., mass flows and
efficiencies. It can be difficult to obtain these quantities since relevant parts of the
performance model need to be chosen. Here, only deltas of the measured quantities
are constructed. The deltas which give the best visible results are for the measured
temperatures. It is possible to separate trends due to the compressor washes in the deltas
which Figure 7.4 and Figure 7.5 indicate. These trends can be monitored by an FDI-system
to decide when a compressor wash is necessary to perform. The best visible result is
shown for the ∆T7 measurement. A disadvantage with the deltas is that the ambient
conditions, e.g., the atmospheric air temperature shown in Figure B.1 may impact the
estimations. Compare, e.g., the absolute value of the deltas for the compressor washes
in the end of December and in the end of June. This complicate the usage of a static
threshold to detect compressor fouling.
It is desirable to have a static threshold Lw to detect when it is time to wash the
compressor due to fouling. The compressor should be washed regularly when a certain
degree of fouling has occurred. As seen in Figure 7.4 and Figure 7.5 a static threshold Lw
 7.3. Measurement Delta 115

Lw

Lw

t
L0

L0

Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[%]

[%]
Nov
3.0

1.5

3.0

1.5

(a) ∆T3 for the first six-month period. (b) ∆T3 for the second six-month period.

Figure 7.4: In the figure, the delta calculation ∆T3 together with the given static thresholds are
shown. The upper static threshold is used to detect compressor fouling, while the lower static
threshold should indicate the performance level of a cleaned compressor.
 116 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

Lw

Tw

t
T0
L0

Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[%]

[%]
Nov
6

(a) ∆T7 for the first six-month period. (b) ∆T7 for the second six-month period.

Figure 7.5: In the figure, the delta calculation ∆T7 together with the given static thresholds are
shown. The upper static threshold is used to detect compressor fouling, while the lower static
threshold should indicate the performance level of a cleaned compressor.
7.4. Constant Gain Extended Kalman Filter 117

used to trigger a compressor wash is not suitable since the delta levels are different for the
summer and winter half year. Also the relative change of the deltas between the washes
are different, e.g., the performance restored by the wash in the middle of November is
larger than for the wash in the end of June.
The plot results can be misleading since the status of fouling before each wash is not
exactly known. Thus, how fouled the compressor is can be different before each wash is
performed. On the other hand, after a compressor wash the degree of fouling should be
the same independently of the winter or the summer period since it is assumed that the
same washing procedure is utilized. The lower threshold L 0 in the figures suggests that the
estimations are not independent of the ambient air condition. Especially, slow changes
in increasing atmospheric air temperature can be difficult to be discriminated from a
cleaned compressor, and vice versa. This can clearly be seen at the end of February in the
figures where the deltas have increased abnormally much, which can be misunderstood
as a fouled compressor.

7.4 Constant Gain Extended Kalman Filter

The Kalman gain is calculated in the operating point where: (i) the ambient air conditions
are at datum state, (ii) the power PA generated by the generator is 21 MW, (iii) the speed of
the power turbine is given by the generator at 50 Hz ( 6500 r pm), and (iv) the mass flow
of fuel (with a lower heating value of about 50 MJ/kg) is calculated using the simulation
platform for the above input signals. In the present work, a mechanical drive application
is studied which results in a power turbine which can have a variable speed. A sketch
of the CGEKF together with the gas turbine and its environment components is shown
in Figure 7.6. The output signals from the observer component are: (i) the estimated
health parameters ĥ, (ii) the estimated combustion temperature T̂ f , (iii) the estimated
measurement signals ŷ, and (iv) an estimation of all the other states variables x̂.

7.4.1 Observability
In the gas turbine diagnosis model, eight unique measurement positions along the en-
gine’s gas path are available. According to Lemma 6.5.4, the maximum number of health
parameters that can be considered in the model is eight. If more health parameters are
used, the observability of the system is not reached. Since the requirement in Lemma 6.5.4
is necessary but not sufficient, the number of health parameters that are observable could
be smaller than the maximum number specified in the Lemma. Thus, an observability
analysis of the gas turbine diagnosis model is necessary to perform.
To check the observability of the diagnosis model, with different health parameter
configurations, the linearized model is studied. According to the numerical problems
that can appear for larger models when observability is checked with the criteria in
Theorem 6.5.1, the structural observability criteria in Theorem 6.5.2 (Shields and Pearson,
1976) is used. The structural observability analysis shows that the health parameter
configuration, described in (6.7), is structurally observable in the linearization point of
the chosen CGEKF observer. If another health parameter, i.e., for the efficiency and mass
118 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

mfuel Starter
Fuel System
Motor

Actuators
Controller

Gas Turbine Application

Ambient

Sensors
DETECT
FOULING

y
ŷ
u CGEKF ĥ
T̂ f x̂

Figure 7.6: Schematic view of the gas turbine experiment platform, where the Constant Gain
Kalman filter is introduced to estimate the gas turbine health h and the combustion temperature
T f . The health estimation is used to detect fouling and the combustion temperature is used by
the controller to supervise the turbine inlet temperature (TIT). The estimation of ŷ can be used
to, e.g., detect an abrupt sensor or actuator fault.
7.4. Constant Gain Extended Kalman Filter 119

flow through the power turbine is added, or the two fault in the input signals PA and n A ,
the linearized model is no longer structurally observable which induce unobservability.
For the case where the two health parameters for the inlet and outlet duct are added,
the structural observability is not affected. A health parameter in the inlet duct can
for example be interpreted as a pressure drop in the air filter due to, e.g., fouling. This
indicates that it can be possible to supervise the health parameter. Finally, the structural
observability of the linearized diagnosis model is strongly connected to the number of
health parameters, and where in the model these parameters are introduced.

7.4.2 Observer Tuning

The covariance matrices Q and R affect the estimation of the state vector x̂ in the filter.
When the uncertainties of the measured signals are independent, the R matrix is diagonal.
The diagonal elements of the matrix represent the variance of respective measurement
signal. The standard deviation are calculated for a number of samples directly after a
compressor wash for a given stationary point. It appears that the standard deviation
lies in the interval 1–3 % of the sensor reference value in the operating point. This gives
matrix elements in R as:
2
Ri,i Šr i y re f , i 

where i 1, . . . , 8, and r i represents the uncertainty of sensor i. The model uncertainty

matrix Q is determined in a similar manner according to:

2
Q j, j C q Šq i x re f ,i 

where j 1, . . . , n, q i represents the uncertainty, and n is the number of state variables.

Since the reference values of the health parameters are zero, the x re f ,i for these parameters
have the value according to the characteristics in the nominal case. All q i :s are given the
same uncertainty of 1 % except the q i :s of the health parameters which are given much
smaller values since the component deterioration is slow. When the constants r i and q i
are specified, the relation between Q and R can be adjusted using the scalar parameter
C q . With this scaling factor it is easy to make a compromise between the noise sensitivity
and the response of the estimator. The most important parameters are the q i :s which
represent the model uncertainty of the health parameters. When the health parameters
are varying too slowly, the component deterioration appears in the estimation errors
y  ŷ which is not desirably and implies that the q i :s of the health parameters are tuned
incorrectly. Thus, the q i :s of the health parameters are increased.
The method to tune the filter is not an optimal procedure but is a simple enough to
construct Kalman filters without using the exact values of the uncertainty matrices, since
they are unknown. A stop criterion when the tuning is complete is when the residuals
(r i y i  ŷ i ), roughly speaking, have no visible trends and the values are centered around
the x-axis.
120 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

7.4.3 Filter Design Summary

The filter design procedure consists of four important steps. These steps are: (i) the
index reduction, (ii) the over-determined M  part, (iii) the state space form, and (iv) the
filter construction. A summary of the design procedure is:
1. Index reduction:
(a) Start with the model given in (6.13). Since the measurement signal y is
a known signal vector, the system is over-determined. Thus, remove the
measurement equation y h̄ ˆx  to get a system that is exactly determined.
(b) Acquire the structural model of the system.
(c) Check, and reduce the DAE-index of the system (if necessary). For this step,
Pantelides algorithm (Pantelides, 1988) is invoked. The input to the algorithm
is the structural model. The output of the algorithm is the equations that
need to be differentiated to receive a smaller index problem.
(d) Take back the removed measurement equations from the first step.
2. Find the over-determined M  part:
(a) Once again, acquire the structural model of the system. This time, the
measurement equations are also included.
(b) Find an over-determined part of the structural model. For this step, the
Dulmage-Mendelsohn decomposition (Dulmage and Mendelsohn, 1958) is
performed where the whole over-determined M  part is chosen.
(c) The actual test equations are now selected.
3. The state space form:
(a) Write the semi-explicit DAE-index 1 system in an ordinary state space form
through an inversion of the algebraic constraints and a symbolic transfor-
mation of the E matrix in (6.21a). Some of the algebraic constraints are
nonlinear so they are solved with a nonlinear numerical solver.
(b) Linearize the system in a suitable operating point and calculate the matrices
A and C.
(c) Check the structural observability of the linearized system.
4. The filter construction:
(a) Specify the measurement uncertainty matrix R, and the model uncertainty
matrix Q, for the Kalman filter where the elements represent the health
parameters are given small values.
(b) Calculate the Kalman gain K, through solving the Riccati equation (6.35),
for the given uncertainty matrices Q and R.
(c) Implement the developed Kalman filter.
7.5. Case Studies 121

7.5 Case Studies

In this section, the design of an observer based CGEKF estimator is evaluated using
(i) data generated from the simulation platform (Evaluation 1), and (ii) experimental
data collected from a gas turbine mechanical drive site (Evaluation 2–3). The objectives
of the simulation study are to: (i) see if the injected performance deteriorations can be
estimated by the designed CGEKF estimator, and (ii) investigate how different ambient
conditions affect the estimations. The objectives of the experimental case studies are to:
(i) estimate the actual health state of the gas turbine over time, (ii) see if the CGEKF
gives reliable state estimates, and (iii) check if the CGEKF based concept is suitable to
use when the time for a compressor wash is determined. In the evaluation, four health
parameters are introduced in the diagnosis model according to Figure 6.3. These health
parameters represent: (i) efficiencies, and (ii) flow capacities of the compressor C1 and
compressor-turbine T1.

7.5.1 Evaluation 1: Atmospheric Weather Condition Dependence

The objectives in this evaluation are to: (i) assure that the CGEKF converges to reasonable
state values during a sweep in different operating points simultaneously deteriorations in
performance are introduced, and (ii) investigate the difference between an observer that
compensates for changes in absolute humidity in the incoming air and one observer that
does not compensate for changes in absolute humidity in the incoming air. Incoming
gases affect the performance of the gas turbine model; therefore it is interesting to
investigate how large the effects in performance are when the air humidity is changed.
In Mathioudakis and Tsalavoutas (2002), a study is presented where the variation in
humidity has a negative effect on the health parameter estimation. The observer that does
not compensate for change in absolute humidity is developed for the datum atmospheric
conditions such as: p 0 1.013 bar, T0 25 Co , and φ 0 60 %. These environment values
give an amount of water in the incoming air according to the moist air model presented
in Buck (1981). The numerical values for different ambient conditions are shown in
Table 5.1.
Input data for the two filters are collected from the reference platform shown in
Figure 2.5. In this evaluation study, the ambient temperature is varied according to the
interval 15–35 Co and the relative humidity is varied according to the interval 40–80 %
when the pressure is at the datum state. At the same time, the power generated by the
external application is varied in the interval 16–26 MW. Because of the varied power,
the speed of the gas generator GG is also varied. In this evaluation study, an electric
generator with fixed frequency is used. Thus the power turbine has a constant speed, but
different generator frequencies are investigated in other studies that have been performed,
and with similar results of the estimated health parameters. The objective is to estimate
deviation in performance and therefore a deterioration in the considered performance
parameters is injected. These injected deteriorations, in percent from respective baseline
reference value are presented in Table 7.4.
The results from the simulation study is presented in Figure 7.7, where the estimated
health parameters are shown. In the Figures 7.7(b)–(e) the estimated degradations in the
122 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

Table 7.4: Injected deterioration in performance equations.

Health Parameter Injected Deterioration [%]

∆η C1 -3.4
∆ΓC1 -2.5
∆η T1 +2.4
∆ΓT1 +4.6

gas path components are shown. Figure 7.7(a) shows the generated power and the amount
of water in 1 kg atmospheric air. The highest and lowest values denote the two extreme
cases from Table 5.1. For these two extremes, the observer that does not compensate for
absolute humidity has a variation in the health estimation of about 1–2 percentage points
for all cases except for the efficiency of the compressor turbine.
The observer that compensates for atmospheric weather conditions follows the in-
jected deterioration in the reference platform nearly perfect even when the ambient
conditions are changed, except during transients. One explanation of this phenomenon
is that the gas properties, according to the change in atmospheric air, are updated simul-
taneously in the whole observer. It can also be seen in the figure that different operating
points do not affect the estimations so much, which is desirable. In Figure 7.7, the power
turbine has a fixed speed, but similar simulation studies are performed where also the
power turbine speed is varied. The results of these studies are similar with the outcome in
Figure 7.7. The conclusion is that the observer that compensates for the different ambient
conditions gives better performance estimations and converges to the actual state of the
system also in the case when the speed of the power turbine is varied.

7.5.2 Evaluation 2: State Estimation for Control

An important signal for the controller during operation is the flame temperature T f .
Since the temperature is not measured, the signal has to be estimated, e.g., by an observer.
An incorrect estimate of this signal can result in a too high turbine inlet temperature
(TIT) which can lead to unnecessary heat stresses in the turbine blades. When the
compressor deteriorates due to, e.g., fouling the compressor discharge and turbine inlet
temperatures increase. It is desirable to run the machine with so high TIT as possible.
Thus, if the component deteriorations are known it is easier to maintain the optimal TIT.
Therefore, it is interesting to see how the fouling affects the flame temperature T f using
the experimental data together with the model.
An investigation is performed where the flame temperature T f is compared for two
CGEKFs. The first CGEKF does not compensate for the deterioration, and the second
CGEKF does compensate for the deterioration using the health parameters. The two
filters have the same Q and R matrices except for the rows and columns which represent
the health parameters. The result of the investigation is shown in Figure 7.8, where the
7.5. Case Studies 123

[g] [MW]
30
25
20
10 20

0 15
(a) Water steam – solid , Generated power – dashed

[%]
-2
-3
-4
(b) Health parameter: ∆ηC1

[%]
-2
-3
-4
(c) Health parameter: ∆ΓC1

[%]
4
3

2
(d) Health parameter: ∆ηT1
[%]
6
5
4

(e) Health parameter: ∆ΓT1

Figure 7.7: In the figure, performance estimation of data generated in simulation platform shown
in Figure 2.5, with injected degradations according to Table 7.4 (not viewed here), are shown.
Two CGEKF observers are evaluated, i.e., an observer that compensates (solid lines) and an
observer that does not compensate (dashed lines) for the variation in ambient conditions are
shown in (b)–(e). The power of the external application is varied together with changes in the
absolute humidity of the incoming air are viewed in (a).
124 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

difference between the turbine inlet temperatures for the two CGEKFs is defined as:

T f ,4  T f ,0
∆T f (7.2)
T f ,4

where T f ,4 is the flame temperature for the filter that does compensate for the deteri-
oration, and T f ,0 is the flame temperature for the filter that does not compensate for
the deterioration. In Figure 7.8 it is shown that the difference in temperature of the two
filters will be up to 1.5 %. The figure shows that the CGEKF with four health parameters
estimates a higher temperature than the CGEKF with no extra health parameters. The
difference between the two estimates varies also with the component deterioration. For a
clean machine, the temperature estimates are almost the same but start to deviate accord-
ing to the increased deterioration. Thus, if T f ,0 is used in the controller, then a variable
temperature threshold which depends on the actual deterioration in the engine is needed
to maintain reliable TIT. Instead, if the temperature T f ,4 is used, then the deterioration
is compensated using the health parameters which results in a more reliable temperature
estimation. The temperature T f ,4 assumes to be more reliable since the estimation errors
(residuals) are smaller for the CGEKF that compensates for the deterioration than for
the CGEKF that does not compensate for the deterioration as Figure 7.8(b)-(d) indicates.
In total, there are eight residuals but only three of them are shown in the figure.

7.5.3 Evaluation 3: State Estimation for Monitoring

To maintain good availability and reliability of the operation it is important to supervise
the engine’s performance. The performance depends of the efficiencies and mass flows
in the gas path components. These parameters are not measured, thus they need to be
estimated by, e.g., an observer. In the present application, the deviation in efficiency and
mass flow from a baseline is estimated by the CGEKF. For the compressor, the estimation
results are shown in Figure 7.9. The data are collected during one year of operation and
the compressor is washed five times: (i) mid November, (ii) end of December, (iii) end
of March, (iv) end of June, and (v) mid September.
To determine when it is time to wash the compressor, a good idea is to study the
health parameters ∆η C1 and ∆ΓC1 . These parameters vary in the interval length of two
percentage points. The interval is marked with the upper and lower thresholds L 0 and
Lw to match the estimated deviation, where the lower threshold Lw can be used to
trigger a compressor fouling alarm in the FDI-system. In Scott (1986), a compressor wash
is suggested when the compressor efficiency has deteriorated 3 %. At SIT, a common
procedure is to wash the compressor before it has deteriorated 3 %. Thus, a rule of thumb
is to wash the compressor when the performance has deteriorated 2–3 %. Since the true
deterioration is unknown, it is not possible to know how fouled the compressor was
when the washes were performed. However, the time distances between the washes are
different, i.e., the second wash is performed after 45 days and the third wash is performed
after three months. Before the third wash (end of March), the performance has not
deteriorated as much as in the other cases according to the estimations. A possible
explanation for this can be that the distance to the previous wash is too large although
7.5. Case Studies 125

%

0 t
Dec Feb Apr Jun Aug Oct
(a) Flame temperature difference ∆T f

15 K

-15 t
(b) Residual r T3 y T3  ŷ T3
kPa 
20

-20 t
(c) Residual r p 3 y p3  ŷ p 3
K
20

-20 t
Dec Feb Apr Jun Aug Oct
(d) Residual r T7 y T7  ŷ T7

Figure 7.8: Flame temperature difference ∆T f (black line) and residuals r i between the CGEKFs
which does compensate (red line) and does not compensate (blue line) for the performance
degradation in the gas turbine.
126 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

%
2

0
L0

-2
Lw
-4 t
(a) Health parameter ∆ηC1

%
2

-2

-4
t
(b) Health parameter ∆ΓC1

%
2

-2 t
(c) Health parameter ∆ηT1

%
2

-2 t
Dec Feb Apr Jun Aug Oct
(d) Health parameter ∆ΓT1

Figure 7.9: Estimation of the health parameter during a time interval of one year. The degradation
in efficiency of C1 can be bounded to an interval of about 2 percent points independent of the
atmospheric weather conditions during the year. Thus, it is possible to use a threshold Lw to
detect when it is time to wash the compressor. The compressor is washed five times: (i) mid
November, (ii) end of December, (iii) end of March, (iv) end of June, and (v) mid September.
When the compressor wash is performed, the efficiency goes up to L 0 . A degradation in mass
flow is shown but it is not so easy to have a static threshold in that case.
7.5. Case Studies 127

the performance has not dropped enough. Therefore, a simple explanation is that the
compressor is washed too early since the time distance from previous compressor wash
is too large.
During the winter and summer half year, the ambient conditions such as temperature
(Figure B.1) and relative humidity are different. As the ∆η C1 parameter indicates (except
for the wash at end of March), it is possible to use the same threshold during the winter
and the summer half year to determine if it is time to wash the compressor. The second
health parameter ∆ΓC1 is more sensitive to the ambient conditions since the estimation
level is different for the winter and summer half year. For the ∆ΓC1 parameter, it is not so
obvious to use a static threshold for the compressor fouling detection procedure, but the
parameter can anyway be important to supervise and useful in the decision procedure.
For the compressor-turbine T1, the monitored health parameters ∆η T1 and ∆ΓT1 are
shown in Figure 7.9(c)–(d). Since these parameters are not varying so much from the
nominal baseline and show no visible correlation according to the compressor washes it
is assumed that the turbine is healthy. The efficiency of the turbine T1 tends to be little
higher on the summer half year than on the winter half year.

7.5.4 Discussion of the Results in Evaluation 2–3

One of the objective with the experimental case studies is to see if the CGEKF method
gives reliable estimates of the state variables. Since experimental data are considered
in Evaluation 2–3, correct values of the state variables are not available. Measurement
noise and model uncertainties are always present, therefore the choice of the Q and R
matrices affects the state estimates. Since these uncertainties are unknown, the matrices
are considered as tuning parameters to obtain reliable state estimates for the present case
study. The state estimates considered here is considered to be reliable because: (i) the
residuals are centered around the x-axis, (ii) no visible trends are seen in the residuals
according to the level of compressor fouling, (iii) trends in the residuals disappear when
the health parameters are introduced (and the residuals get smaller compared with a
case when no health parameters are used see, e.g., Figure 7.8), and (iv) the trend in the
estimated performance degradation of the compressor is slow and decreases with time.

7.5.5 Nonlinear versus Linear Estimator

An important question is how accurate is the designed nonlinear CGEKF compared
to a linear Kalman filter (KF) during different operational conditions (according to
the performance chart in Figure 2.3) given the same covariance matrices? To answer
that question, a parameter study is performed where a linear Kalman filter (6.33) is
designed for the reference ambient conditions at the stationary operation point 50 Hz
(n A  6500 r pm) and 21 MW. The two filters have four health parameters each, and
evaluated in a number of stationary points at three ambient conditions. The data for
the evaluation are generated using the simulation platform in Figure 2.5. The difference
between the health parameters at different operating points for the two filters are shown
as maps in Figures 7.10–7.12 for the ambient conditions. The evaluation is performed
at nominal performance conditions (i.e, no component deteriorations are considered)
128 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

which results in estimations of the health parameters performed by the CGEKF that are
zero in all stationary points. Thus, the maps show the estimation error of the KF for the
health parameters in percent of the nominal value. In Figure 7.10, the operating range
of the gas turbine for the experimental data is shown as grey area in the figures. The
difference between the max and min of, e.g., the ∆η C1 -parameter is 0.2 percent points
and the ∆ΓC1 -parameter is 0.8 percent points. The performance deterioration for these
two health parameters for a fouled compressor is about 2–3 percent points which was
presented in Figure 7.9. When also the ambient conditions change the difference gets
larger (see Figures 7.11–7.12).
The evaluation shows that the nonlinear CGEKF is more accurate (the health pa-
rameters are zero for all the stationary points) than the linear Kalman filter, but the
difference is not so noticeable when the ambient conditions and the turbine speed are
nearly constant. When the whole operating area as shown in Figure 2.3 should be covered,
the CGEKF is a better choice than the KF. The linear filter is much less time consuming,
than the nonlinear filter (seconds instead of minutes). For the industrial gas turbine
application, the increase in time for the nonlinear filter is not a problem since application
is not time critical. The experimental data is collected for one year which is a long period
of time compared to the simulation time.

7.6 Summary of the Performance Estimation Techniques

In this section, a summary of the considered performance estimation techniques used
for supervision is presented.

7.6.1 Bell-Mouth Based Estimation

The first standardized method to detect compressor fouling is based on the bell-mouth
pressure drop measurement shown in Figure 7.2(a). The method is simple since the
model assumption of the mass flow through the compressor is not so sophisticated.
The bell-mouth measurement gives an estimation of the relative mass flow through
the compressor. For the available measurement sequence the method indicates poor
performance. It is hard to distinguish when a compressor wash is needed, since a baseline
for a clean compressor lies in the middle of all points in the diagram. It is desirable that
all points in the diagram lie below the baseline and the distance increases with time. A
better baseline estimation is received when a physical based model is used to estimate
the mass flow as Figure 7.2(d) indicates.

7.6.2 Measurement Delta

The second method is based on the so-called measurement deltas. These measurement
deltas are more or less, a comparison between an estimated and a measured gas path
parameter. A first step is to construct these deltas directly for the measured quantities,
i.e., for the pressures, the temperatures, and the shaft speeds along the engine’s gas
path. The benefits with this method are: (i) it is simple to construct deltas because the
7.6. Summary of the Performance Estimation Techniques 129

PA MW  PA MW 

0. 0.25 1
26 1 eg i a l 0.2 26

-3
a r ent
0.0 on 0.15
5 0.5
dat erim

0
-1
-1.5
24 0 0.1 24
p

-0 0.05
Ex

.0
5
22 22

-2.5
0
0
20 20
1

-2
-0.

-0.5
Linearization
18 point 18

0
5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(a) Health parameter: ∆η C1 [%] (b) Health parameter: ∆ΓC1 [%]

PA MW  PA MW 

0.2 0.5
26 26
0
0
-1.4

-1
-0.6
-0.4
-0.2

-0.5

24 24
0
-1.2

0
22 22
-2
-0.8 -1

-1.5

20 20

18 0 18 0
0

5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(c) Health parameter: ∆η T1 [%] (d) Health parameter: ∆ΓT1 [%]

Figure 7.10: In the figure, the difference in health parameter estimation (in percent of the nominal
value) for the nonlinear CGEKF and the linear KF at the ambient conditions T 25X C and
φ 60 is shown. For stationary conditions, the CGEKF estimate the correct value of the health
parameters (which is zero), thus the error for the KF is seen in the figures. The linearization
point for the KF and the region where experimental data are collected are shown in the figure.
130 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

PA MW  PA MW 
0 0.1

0
26 26 1.5

0.5
0
-0.1

-1
-0.5
24 -0.1 24 1
-0.2
-0.2
-0.3

-2
22 22
0.5
-0.3

-1.5
20 -0 20
.4

18 18

0
5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(a) Health parameter: ∆η C1 [%] (b) Health parameter: ∆ΓC1 [%]

PA MW  PA MW 

26 26 0
-1.4

-2.5
-0.4

-0.5
-0.2

24 24
0
-1.2

22 22
-0. -0.8 1

-1.5 -2
-

20 0 20
6

18 18 0
0

0.2
-1

5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(c) Health parameter: ∆η T1 [%] (d) Health parameter: ∆ΓT1 [%]

Figure 7.11: In the figure, the difference in health parameter estimation (in percent of the nominal
value) for the nonlinear CGEKF and the linear KF at the ambient conditions T 15X C and
φ 40 is shown.
7.6. Summary of the Performance Estimation Techniques 131

PA MW  PA MW 
0.85
26 0.8 26
-0.5
0.

0.75

-2
-2.5
7

24 24

-1

-3.5 -4
22 22
0.6

5 0.75
20 0.6 20

-1.5
0.7 0.75 -3

-1
18 0.85 18
0.8
5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(a) Health parameter: ∆η C1 [%] (b) Health parameter: ∆ΓC1 [%]

PA MW  PA MW 
0

26 0.2 26
1
-0.8

-0.4
-0.6

-0.2

-0.5

24 24
0
-1.4

22 22 0.5
-1.5
-1.2

20 20
.2
-0

-1
-1

18 18 0.5
0

5000 5500 6000 6500 7000 5000 5500 6000 6500 7000
n A r pm n A r pm
(c) Health parameter: ∆η T1 [%] (d) Health parameter: ∆ΓT1 [%]

Figure 7.12: In the figure, the difference in health parameter estimation (in percent of the nominal
value) for the nonlinear CGEKF and the linear KF at the ambient conditions T 35X C and
φ 80 is shown. The grey are are not valid for the given ambient conditions.
132 Chapter 7. Performance Estimation in Industrial Gas Turbine Engines

performance model is already available, and (ii) trends due to performance degradation,
e.g., compressor fouling, can be detected. The performance model is not augmented with
extra parameter or state variables. Therefore, deterioration in components is not modeled
explicitly which results in residuals that are dependent on, e.g., compressor fouling. Thus,
these residuals may not be suitable for sensor and actuator fault diagnosis since they also
capture the performance deterioration. It can also be difficult to associate a number of
deltas to a physical fault in the gas turbine. A more sophisticated approach is to use the
measurement deltas in the GPA methodology (6.5). Together with the influence matrix
H the deviation in performance can be calculated.
The disadvantage with the method is the dependency of the ambient weather con-
dition. The deltas appear to detect trends in performance degradation but also in the
ambient atmospheric condition. Therefore it can be difficult to know exactly when a
service of the gas turbine is necessary to perform, without further normalization of the
data.

7.6.3 Constant Gain Extended Kalman Filters Constructed Using the

Proposed Equation Selection Procedure
The last method is based on state estimations using a Kalman filter. For this method only
relevant parts of the original performance model are considered in the filter which leads
to a reduced model in the sense of equation and variables. In an observer, the filtering
of signals is encapsulated by the feedback term, so an external filter is not necessary.
Benefits with this method are: (i) the modeling is more physical based since health
parameters can be injected in the performance equations, (ii) the deviation in, e.g., the
efficiency of the compressor is explicitly estimated, (iii) the deviation in efficiency seems
to be less dependent of the atmospheric conditions than the measurement delta method,
(iv) fault diagnosis can be introduced, and (v) when the framework to generate observers
is developed it is possible to generate diagnosis test in a relatively easy manner. Using the
designed filter, other state variables, e.g., the combustion flame temperature is estimated.
The flame temperature can be utilized in the controller to calculate the turbine inlet
temperature. This parameter is important to supervise since too high temperatures in
the first blades of the turbine should be avoided.

7.7 Conclusion
The topic addressed in the chapter is an evaluation of the designed diagnosis test quan-
tities implemented as constant gain extended Kalman filters (CGEKF). The filter is
augmented with health parameters representing efficiency and flow capacity in the com-
pressor and compressor-turbine. The objective with the health parameters is to capture
deviation in performance. The CGEKF estimator is evaluated using: (i) data generated
from the simulation platform, and (ii) experimental data collected from a gas turbine
mechanical drive site. The objectives of the simulation study are to: (i) see if the injected
performance deteriorations can be estimated by the designed CGEKF estimator, and
(ii) investigate how different ambient conditions affect the estimations. The objectives
7.7. Conclusion 133

of the experimental case studies are to: (i) estimate the actual health state of the gas
turbine over time, (ii) see if the CGEKF gives reliable state estimates, and (iii) check if
the CGEKF based concept is suitable to use when the time for a compressor wash is
determined.
For the case study which utilized simulated data, the injected performance deteriora-
tions can by estimated by the CGEKF. In the most extreme cases, a change in the ambient
condition can be misinterpreted as a performance deterioration in the gas turbine. This
can result in, e.g., a compressor wash which is performed to early. The experimental
case study is performed using data from a mechanical drive site during a time period of
one year. The focus in the study is to evaluate the test quantities used in the FDI-system.
The evaluation is performed through: (i) supervision of the performance deterioration,
and (ii) estimation of the flame temperature T f which is an important parameter for
the controller. The flame temperature is estimated at the same time as the test quantity
compensates for the deteriorations in the compressor C1 and turbine T1. Using the
estimated deterioration, a decision of a fouled compressor should be taken. The investi-
gation indicates that it is possible to use the FDI-system to detect a fouled compressor
since the deterioration of the compressor decreases about 2 % between the compressor
washes. Since the interval and offset of the deterioration are the same independently of
the winter or summer period, it is possible to have static thresholds for fouling detection
in the FDI-system. Finally, the case study shows that the model design and the proposed
systematic test construction procedure can be considered when an FDI-system for an
industrial gas turbine is constructed.
 Chapter 8

Investigation of Fault Diagnosis in the Startup

and Shutdown Operating Modes

During the startup and shutdown sequences of the gas turbine, loads and stresses in
components increase. The gain in stress and the fact that the engine goes through many
operating points can result in component failures which are triggered. These component
failures are, e.g., leakages in valves or increased friction in the mechanical parts. Thus, a
good idea is to use the startup and shutdown sequences to detect and isolate faults in
gas turbine components using filters. Since the engine is running at different operating
modes, the potential of a nonlinear Kalman filter can fully be utilized. The advantage
with using filters for fault estimation in nonlinear systems is that parameters can be
added easily in the model. These parameters represent the fault which can be estimated
by the filter.

8.1 Background
The startup and shutdown dynamics of the gas turbine are much faster compared to
the performance deterioration according to, e.g., fouling. In Chapter 7 the performance
deteriorations are estimated using health parameters. Here, it can be assumed that
the health parameters are constant during the startup and shutdown sequences and
they can therefore be removed from the diagnosis model. When the health parameters
are removed, it is possible to introduce parameters representing fault in components.
Another aspect is that different equations are valid during different running phases, i.e.,
the model consists of hybrid parts which are specified using if-statements. Hybrid
models are not directly handled by the developed framework for the diagnosis test
generation. Instead, the user needs to specify, e.g., if the combustion is activated, or if
the load is applied. Thus, the sequence for the startup to shutdown running mission is
divided into the running phases: (1) apply starter motor, (2) ignite combustor, (3) ramp

135
136 Chapter 8. Investigation of Fault Diagnosis in the Startup and Shutdown Operating Modes

up engine, (4) apply load, (5) maintain operating speed, and (6) ramp down engine.
These running phases are shown in Figure 8.2 with a subset of gas path parameters. As
the figure indicates, the running phases are accomplished sequential for the startup to
shutdown running mission.

8.2 Test Construction Procedure

The test construction procedure is as follows: (i) design a diagnosis model, (ii) generate
an observer for each running phase, and (iii) collect these observers in a test quantity.

8.2.1 Fault Modeling

For the fault modeling, parameters which represent the faulty behavior in the startup and
shutdown operating modes are introduced in the diagnosis model. The parameter which
represents the faulty behavior is later on estimated when the test quantity is constructed
according to the methodology described in Chapter 6. In this section, an overview
of the modeled faults and in which equations they are introduced are presented. The
faults are modeled in the same way as the health parameters were modeled in Chapter 7.
Here, the considered fault modes are: (i) leakage in the bleed valve 3 of the compressor,
(ii) increased friction in the bearing of the gas generator, and (iii) increased friction
in the bearing of the power turbine which all are introduced in the diagnosis model.
The faults are added to the diagnosis model using a component based approach, i.e., a
component which represents the fault is added using the graphical interface. The fault
component is then recognized automatically by the developed parsers.

Component leakage
A leakage is modeled using a fault parameter which represents the leakage mass flow of
the component. This unknown mass flow (which is included in the fault component)
is added to the connection point in the gas path where the leakage fault component is
placed (i.e., where the leakage should be modeled). In the connection point, the leakage
mass flow is written:

mflow,leak fleak
(8.1)
Hflow,leak mflow,leak h

where h is the enthalpy and Hflow,leak is the enthalpy flow in the connection point. The
constraint:
f˙leak 0 (8.2)

is added to the fault model. Later on, when the constant gain extended Kalman filter
(CGEKF) is designed, it is assumed that this equation has a high level of uncertainty.
Thus, the response for the fault parameter in the CGEKF gets fast.
8.2. Test Construction Procedure 137

Increased friction
A fault in the mechanical bearings is described using a variable which adjust the resistance
when the shaft rotates. The increased resistance is modeled as a change in the frictional
constant kmech and is written:
P f r i c f f r i c α ˆn (8.3)
where f f r i c represents the change in friction, P f r i c is the generated power, and α ˆn
is a function of the shaft speed n. In the same way as for the component leakage, the
constraint:
f˙fric 0 (8.4)
is added to the fault model.

8.2.2 Diagnosis Model

The diagnosis gas turbine model, with the added faults, has the general differential
algebraic equation (DAE) form:

F ˆż, z, u 0 (8.5a)
y h ˆz  (8.5b)
z ˆ x, f T (8.5c)

where f represents the unknown faults, x consists of the unknown variables, y consists
of the known measurement signals, u consists of the known input signals. The functions
F and h together with their arguments are vector valued functions with appropriate
dimensions. The modeled faults are here:
BV 3 sha f t sha f t 0
f ˆ fleak , f f ric 1 , f f ric 

which represent: (i) a leakage in the high pressure bleed valve of the compressor, (ii) an
increased friction in the gas generator, and (iii) an increased friction in the power turbine.

8.2.3 Test Quantity

During the startup to shutdown running mission, different equations of the diagnosis
model are valid according to the switching between the six running phases as shown in
Figure 8.2. Therefore, the test quantity for the startup and shutdown sequence consists
of four constant CGEKFs which utilize different part of the diagnosis model and have
different feedback gains. These four filters are used for the running phases (see index
in Figure 8.2): (i) apply starter (1), and ramp down engine (6), (ii) ignite burner (2),
(iii) ramp up engine (3), and (iv) apply load (4), and operating speed (5). The filters are
generated using the test quantity generation procedure described in Section 7.4. The Q
and R matrices are tuned in that way so the response of the state variables representing
the faults are fast. The filters are generated from the same diagnosis model (except for case
(5) when health parameters are considered). Between the running phases, the estimated
state variables are saved and used as initial conditions to the next phase. The running
138 Chapter 8. Investigation of Fault Diagnosis in the Startup and Shutdown Operating Modes

phase when the engine has reached the operating speed condition (5), the test quantity
developed in Chapter 7 (which includes health parameters) can be used. The value of the
health parameters from the previous operating speed condition can be used as initial
values for the health parameters. The other filters can be adjusted with the previous value
of the health parameters, i.e., normalize the performance. The method for sensor and
actuator fault diagnosis, which is introduced in Chapter 9 can also be combined with
the diagnosis of the component faults in the startup and shutdown phases. An overview
of which faults that can be diagnosed during the various running phases is shown in
Figure 8.1.

normalize
performance

health parameters
component sensor faults component
faults actuator faults faults

Startup Operating
Shutdown
Condition

save/restore
health

Figure 8.1: An overview of the possible type of diagnosable faults during the various running
phases. When the engine starts the shutdown phase, the performance is normalized with the
values of the health parameters.

8.3 Simulation Study

The designed test quantity is evaluated using simulated data gathered from the simulation
platform shown in Figure 2.5. The component faults are injected in the simulation
platform during the ramp up engine phase (3). The faults are described as: (i) an
outgoing mass flow in the bleed valve 3 position of 3 % of the nominal mass flow through
the compressor, and (ii) an increased friction of 250 % of the nominal friction coefficients
in the shafts of the gas generator and the power turbine. Since the nominal values of
the friction coefficients are small, the values of the injected faults get large in percent.
Smaller values of the faults in the friction are possible to diagnose. The problem is rather
that large faults are difficult to investigate since the simulation platform cannot handle
too large faults. An injected fault which is too large results in: (i) many state events, (ii) a
stiff system, and (iii) a simulation that crash. During the simulation, it is assumed that
8.3. Simulation Study 139

the performance is at nominal reference conditions and no noise is added to the output
signals.

8.3.1 Fault Free Sequence

For the fault free sequence, the objective is to compare some of the estimated variables
with the corresponding variables in simulated data. Important variables (which are
not measured) to supervise are, e.g., the mass flow of air through the compressor, and
the combustion flame temperature T f . These two variables are shown in Figure 8.2
together with the rotational speed of the gas generator and the power turbine (which are
measured). The startup and shutdown sequence is divided into the six phases: (1) apply
starter, (2) ignite burner, (3) ramp up engine, (4) apply load, (5) operating speed, and
(6) ramp down engine (release load) which is marked in the figure.
The result of the fault free investigation is a test quantity which produces estimations
which agree with the simulated data. The residuals and fault parameters are zero except
during some time in the beginning of each phase. This phenomenon is seen in, e.g., the
fault parameters in Figure 8.4 or Figure 8.3 in the beginning of phase 3 where it takes
some time before the parameters go back to zero.

8.3.2 Component Faults – Leakage in the Compressor and Increased

Friction in Mechanical Bearings
For the simulation cases where component faults are introduced, the objective is to
estimate the faults correctly with the diagnosis test. How fast the test responds to the
fault depends on how the covariance matrices Q and R are chosen when the CGEKF is
designed. The tradeoff is between how much the estimations should rely on the model
and on the measurements. Thus, the elements in Q which belong to the fault parameters
are associated with a large value, i.e., the fault estimations rely more on the measurements
than on the model. When the fault is estimated correctly by the test quantity, the fault
is decoupled and the test quantity can be used to supervise the same variables as for
the fault free case. The considered faults are: (i) a leakage in the bleed valve 3 with an
amplitude of 3 % of the nominal mass flow through the compressor, (ii) an increased
friction coefficient in the bearing of the gas generator with an amplitude of 250 % of the
nominal friction coefficients, and (iii) an increased friction coefficient in the bearing of
the power turbine with an amplitude of 250 % of the nominal friction coefficients. For
each fault mode, a measurement sequence is generated where the actual fault is injected.
The fault is injected when the engine ramp up, i.e., during section 3 in Figure 8.3.
All three fault cases can be diagnosed by the developed test quantity. The supervised
variables (e.g., mass flows and temperatures) agree with the corresponding variables in
the simulated data which is not the case if no fault estimation parameters are used in
the CGEKFs. In Figure 8.3 and Figure 8.4, the fault case of the leakage in the compressor
and the increased friction in the gas generator are shown. As shown in the figures, it
seems that the fault parameters which are sensitive to the injected faults can estimate the
fault correctly and be used to trigger an alarm. The fault parameters which should not
be sensitive to the injected faults are not reacting.
140 Chapter 8. Investigation of Fault Diagnosis in the Startup and Shutdown Operating Modes

kg ~s
80
60
40
20
0 t
(a) Mass flow of air trough compressor C1.

K
1400

1000

600

200 t
(b) Flame temperature T f .

r pm

8000

4000

0
t
1 2 3 4 5 6

(c) Speed of gas generator (blue), and speed of power turbine (red).

Figure 8.2: Typical appearances of a startup and shutdown sequence. The sequence consists of the
phases: (1) apply starter, (2) ignite burner, (3) ramp up engine, (4) apply load, (5) operating
speed, and (6) ramp down engine (release load).
8.3. Simulation Study 141

4 %

-2

-4 t
(a) Estimation parameter of leakage in bleed valve 3 of compressor (red), and injected fault (black).

%
200

-200
t
(b) Estimation parameter of increased resistance in the bearing of gas generator (red), and injected fault (black).

%
200

-200
t
(c) Estimation parameter of increased resistance in the bearing of power turbine (red), and injected fault (black).

r pm

8000

4000

0
t
1 2 3 4 5

(d) Speed of gas generator (blue), and speed of power turbine (red).

Figure 8.3: Fault case – leakage in bleed valve 3 of compressor C1.

142 Chapter 8. Investigation of Fault Diagnosis in the Startup and Shutdown Operating Modes

4 %

-2

-4 t
(a) Estimation parameter of leakage in bleed valve 3 of compressor (red), and injected fault (black).

%
200

-200
t
(b) Estimation parameter of increased resistance in the bearing of gas generator (red), and injected fault (black).

%
200

-200
t
(c) Estimation parameter of increased resistance in the bearing of power turbine (red), and injected fault (black).

r pm

8000

4000

0
t
1 2 3 4 5

(d) Speed of gas generator (blue), and speed of power turbine (red).

Figure 8.4: Fault case – increased resistance in ball bearing of gas generator.
8.4. Conclusion 143

8.4 Conclusion
In the chapter, a pilot study is presented where the objective is to investigate and develop
a diagnosis test quantity used for detection and isolation of component faults in the
startup and shutdown sequences of the gas turbine application. Since the engine goes
through many operating modes in this case, the potential of the nonlinear Kalman filter
can fully be utilized. The startup and shutdown sequences can be relevant to study for
diagnosis purposes since the engine goes through many operating modes which can
trigger faults which are not seen during, e.g., stationary conditions. The focus in the study
is to demonstrate the ability to introduce parameters in the Kalman filters which are
used to estimate fault in components. In this study, three different faults are investigated:
(i) a leakage in the bleed valve 3 of the compressor, (ii) an increased rotational friction
coefficient in the gas generator, and (iii) an increased rotational friction coefficient in
the power turbine. These three faults are possible to detect and isolate using data from
the simulation platform. Here, the three faults are introduced in the same Kalman filter,
but other fault isolation configurations and fault modes are also possible to obtain.
 Chapter 9

Diagnosis and Fault Tolerant Supervision of

Industrial Gas Turbines

Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines‡

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson

Vehicular Systems, Department of Electrical Engineering.
Linköping University, SE-581 83 Linköping, Sweden.

Abstract
Monitoring of industrial gas turbines is important since it gives valuable information of
the process health, and component failures. Sensor and actuator faults affect the engine’s
operation point and complicate the determination of compressor wash intervals. Thus, it
is important to have a fault detection and isolation (FDI) system which can detect and
isolate these faults at an early stage. In the paper, an approach to construct an FDI-system
which utilizes multiple constant gain extended Kalman filters is suggested. Each filter
decouples one of the considered faults, and in case of a fault, it is only one of the filters
that are consistent with the input and output signals. The filter which decouple the
fault is used to estimate the performance, thus a fault tolerant system for supervision
and control is obtained. To evaluate the FDI-system, two case studies are performed
based on: (i) simulated data, and (ii) experimental data from a gas turbine site. The
investigation shows that all important faults are detectable and isolable, at the same time
the performance is supervised. The estimation of performance can be used for control
or to determine when it is time to wash the compressor.

‡ This chapter is an edited version of the paper Diagnosis and Fault Tolerant Supervision of Industrial Gas
Turbines which is submitted for journal publication.

145
146 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

9.1 Introduction
In industrial gas turbines, deterioration in components throughout the gas path is
commonly occurring and contributes to the overall performance deterioration of the
engine. To maintain reliable engine operations it is crucial to detect and isolate faults
in sensor and actuator signals throughout the gas path. It is important to discriminate
between an instrumentation fault and a component deterioration to avoid unnecessary
operation stops. Instrumentation faults which are left undetected can result in wrong
input signals to the controller. An important signal, used in the controller, is the com-
bustion flame temperature which is not measured by any sensors but is estimated in the
controller using the instrumentation signals together with a simplified engine model.
The flame temperature estimate is affected by: (i) the component deteriorations, and
(ii) the faults in sensors and actuators. Thus, both these aspects need to be considered in
the fault detection and isolation (FDI) system and in the controller to maintain reliable
engine operations. An undetected fault can also result in a misclassification of a fouled
compressor which increases the environment impact, the fuel consumption, and a com-
ponent failure in the worst case scenario. When a single instrumentation fault is present
in the engine, it is preferably that the operation can continue in a robust manner without
interruption until the next overhaul, which results in a more efficient planning of service
and maintenance. The focus herein is on the FDI-part, but the state variables estimated
by the FDI-system can with great advantage also be used in the controller.
One of the major contributions which cause performance deterioration in industrial
gas turbines is fouling (Diakunchak, 1992; Kurz et al., 2009). Fouling is caused by small
particles and contaminants in the air caught by the compressor blades which reduces
the efficiency and mass flow of the compressor. A common approach in the gas turbine
diagnosis literature, to estimate component deterioration is to use filters combined with
health parameters (Luppold et al., 1989; Volponi et al., 2003). The filters are often Kalman
based, and the health parameters are correction factors for, e.g., efficiencies and flow
capacities in the components. In Kobayashi et al. (2005), a constant gain extended
Kalman filter (CGEKF) is investigated where a nonlinear model can be combined with
the offline calculation of the Kalman gain matrix. It is a good idea to use the health
parameters to decide the level of compressor fouling, i.e., to determine when it is time to
wash the compressor which was investigated in Larsson et al. (2014a).
In Merrill et al. (1988); Kobayashi and Simon (2005), a fault isolation strategy based
on hypothesis testing is presented where a hypothesis is specified for each fault case. The
tests consist of linear filters and when a fault is present in the system, only one hypothesis
is valid. Therefore, it is possible to isolate the faults. When the test compensates for the
component deterioration, the residuals become less dependent on the deterioration level
which results in a possible usage of smaller threshold for fault detection. In Merrill et al.
(1988), no compensation is made for the component deterioration, and in Kobayashi
and Simon (2005) the component deterioration is compensated offline for the sensor
fault hypotheses. In Kobayashi and Simon (2003), the compensation of deterioration in
components is performed online for all fault hypothesis using only health parameters.
Since the health parameters try to explain the faulty sensor or actuator value, the detection
and isolation performance may be weaker than in Kobayashi and Simon (2005). Since
9.1. Introduction 147

noise is present in the measurement sequences, this results in test quantities which
may not alarm although they are sensitive to the considered fault. Especially slowly
drifting sensor and actuator bias faults can be difficult to detect with the proposed
method. To get better detection performance, an idea is to estimate the fault instead
of removing one sensor equation in each test quantity. This leads to the case where the
correct fault hypothesis always estimates the fault, and the estimated fault can be used
for fault detection in the FDI-system. The disadvantage is that all tests are sensitive to
all considered faults, thus ideally it is not possible to isolate the fault. In practice, it can
be assumed that the test quantity which estimates the fault also decouple the fault in
the test quantity when the drifting bias is slower than the dynamic of the estimation
parameter in the test, i.e., slowly drifting faults are easier to decouple than abrupt faults
in that sense.
To gain the overall diagnosis performance, i.e., increase the fault detectability and
reduce the false alarm probability, it is desirably to have an FDI-system which consists of
diagnosis tests based on models with high accuracy. When a model, e.g., consists of not
negligible biases, a high frequency rate of false alarms can be generated. Thus, a reduction
in false alarms can reduce the ability to detect a fault since the thresholds can be lowered.
In the developed FDI-system, the tests quantities are based on nonlinear models which
are presented in Larsson et al. (2014a). These nonlinear models are used for performance
analysis by Siemens Industrial Turbomachinery AB (SIT). In Larsson et al. (2014a), a
framework to automatically generate CGEKF based estimators from the physical based
nonlinear gas turbine model implemented in Modelica (Modelica Association, 2007)
is also presented. An automatic and systematic design method is valuable since the
model is a large differential algebraic equation (DAE) model with nonlinear parts and
the design of the filters might be a complex task without an automatic method. For the
work presented herein, experimental data and a simulation platform are provided by SIT.
Using the simulation platform, data for the FDI-system can be generated under various
operational conditions.

9.1.1 Problem Statement

The objective of this work is to design and evaluate an FDI-system for an industrial gas
turbine based on hypothesis testing, while so many health parameters as possible are
utilized in the diagnosis test quantities. The FDI-system should diagnose single sensor
or actuator faults at the same time the supervision of performance is maintained, i.e.,
an FDI-system used for fault tolerant supervision. The focus here is on detection and
isolation of slowly drifting bias faults. When a fault is detected and then isolated, the
input signals used in, e.g., the controller should be based on estimates where the fault
and the component deteriorations are decoupled. For each fault mode, a test quantity
has to be designed carefully as a part of the FDI-system.
The evaluation is performed using: (i) data generated from the simulation platform,
and (ii) experimental data collected from a gas turbine site. The health and tuning
parameters of the test quantity should be adjusted to work together with the experimental
data. For the two evaluation cases, a single sensor or actuator fault is injected in the
data sequences for each fault case. The objectives with the two evaluation studies are to:
148 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

(i) examine each test quantity if it gives reasonable values, (ii) check if the considered
faults are detectable and isolable, (iii) false alarms should be avoided, (iv) investigate
how large faults that are possible to detect and isolate with the designed FDI-system,
(v) check if it is possible to isolate the fault before the compressor wash alarm is triggered,
(vi) inspect the health parameter estimates for each fault hypothesis and compare the
results with the fault free estimates, and (vii) use the estimated health parameters to
decide if it is time to wash the compressor or not.

9.1.2 Outline and Contributions

The paper starts with the design procedure of the physical based estimators (Sec. 9.2),
and continues with the diagnosis test quantity construction (Sec. 9.3). In Sec. 9.4, the
description of the fault isolation method which is used in the FDI-system is presented.
The next step is to evaluate the FDI-system using data gathered from the simulation
platform (Sec. 9.5), and continues with an investigation of an experimental case study
(Sec. 9.6). In the two evaluations, faults are injected into the measurement signals.
The first contribution of the paper is a systematic design procedure of an FDI-system,
for an industrial gas turbine, which can be used for fault tolerant supervision and control
when a single sensor or actuator fault is present. The diagnosis test quantities in the
FDI-system are based on nonlinear differential algebraic equation (DAE) models. The
corresponding models at SIT are used for performance analysis. It is shown that the
proposed method gives good fault detection performance, also for slowly drifting sensor
and actuator bias faults. To increase the fault isolation performance, an upper and lower
limit of the health parameters is introduced.
The FDI-system is applied to an evaluation of: (i) simulated data, and (ii) experi-
mental data from a gas turbine site. For the two case studies, all considered fault modes
are detectable and the most of them are isolable. When a unique diagnosis is determined
by the FDI-system, the performance can be supervised using the test quantity which
are based on the most probably fault hypothesis. The deviation in performance of the
compressor can then be used to determine if it is time to wash the compressor, also when
a single sensor or actuator fault is present.

9.2 Gas Turbine Diagnosis Modeling

The objective in this section is to describe the gas turbine models that form the basis
of the developed FDI-system. Given these models, diagnosis test quantities Ti can
be generated which are the fundamental part of the FDI-system. The systematic and
automatic test design procedure is presented in Larsson et al. (2014a). In the design, the
nonlinear gas turbine model is transformed to a test quantity which is implemented as
a constant gain extended Kalman filter (CGEKF). The filter is then used to supervise
the performance due to, e.g., compressor fouling and control purposes. Here, the focus
is on fault isolation, thus the FDI-system consists of more than one test quantity. The
FDI-system can supervise the performance at the same time as a single fault is present.
9.2. Gas Turbine Diagnosis Modeling 149

For each estimator in the FDI-system, the main design choices are to decide: (i) the
number, and (ii) the position of the health parameters, at the same time one fault is
decoupled in the diagnosis test. The investigation shows, e.g., that it is difficult to decouple
a fault in the speed sensor of the gas generator at the same time the deterioration in
mass flow through the compressor is supervised. Therefore, the focus in this section is
to decide which health parameters that can be modeled and which faults that can be
decoupled in each test to have an estimator that can be running using the experimental
data without instabilities.
For each test quantity Ti , a physically based diagnosis model D i is constructed.
The diagnosis model is derived from a model which is similar, and validated against a
reference model (Idebrant and Näs, 2003) used by SIT for performance analysis. The two
model concepts used for diagnosis and performance analysis are first presented in Larsson
et al. (2010) and further improved in Larsson et al. (2014a). In the diagnosis model D i ,
concepts related to component degradation and fault diagnosis are introduced which
are not considered in the model used for performance analysis. Thus, the number of:
(i) component degradation parameters (health parameters), and (ii) sensor or actuator
faults, can be specified where each configuration results in a separate diagnosis model.
The nominal performance model of the gas turbine has the DAE form:

F ˆẋ, x, u 0 (9.1a)
y h ˆx  (9.1b)

where x represents the unknown variables, u represents the input signals, and y represents
the measurement signals.

9.2.1 Measurement Signals

All signals used in the FDI-system are based on measurements, but here the signals are
divided into the two groups: (i) actuator signals (input signals), and (ii) sensor signals
(output signals). In Fig. 9.1, a schematic view of the gas turbine and its measurement
signals are shown.

Input (Actuator) Signals

The available input signals are divided into two groups. The first group consists of: (i) the
pressure p 0 , (ii) the temperature t 0 , and (iii) the relative humidity φ 0 of the ambient air.
These signals are assumed to be non-faulty and are used to determine the gas properties
of the ambient air. The second group of input signals is: (i) the mass flow of fuel m f ,
and (ii) the power generated by the application PA . The signals in the second group can
be faulty and should be diagnosed by the FDI-system.

Output (Sensor) Signals

The available output signals are the measured quantities of: (i) the temperature t 2 and
pressure p 1 before compressor, (ii) the discharge temperature t 3 and pressure p 3 after
compressor, (iii) the temperature t 7 and pressure p 8 after power-turbine, and (iv) the
150 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

p0 t0 φ0 mf PA

CC

C1 T1 T0 A

∆η C1 ∆ΓC1 ∆η T1 ∆ΓT1

p1 t2 p3 t3 n C1 p8 t7 n T0

Figure 9.1: The gas turbine with the output signals (solid), the input signals (dashed), the ambient
input signals (dotted), and the health parameters (∆ i ). The secondary air flows, used to cool
the first turbine blades, are shown with dashed arrows in the figure.

speed of gas generator n C1 and the power-turbine n T0 . The subscript notation in the
measured quantity describes at which cross-sectional area the sensor is located. A low
number represents the air entrance and a high index number represents the position
where the exhaust gas leaves the gas turbine. A challenge with this specific product model
is the absence of an instrumentation sensor between the two turbines. The lack of these
types of sensors makes the diagnosis and monitoring procedure more difficult to perform.
Having ideal thermocouples in that cross-section should reduce the uncertainty of the
gas path parameters in the gas generator.

9.2.2 Health Parameters

To capture unmeasured performance deterioration in the gas turbine components, a
common approach in the field of gas turbine health management and control is to
introduce a number of physical based quantities named health parameters (Kobayashi
and Simon, 2005; Simon and Simon, 2005; Borguet and Léonard, 2008). These parameters
are introduced in the equations where efficiencies and flow capacities are calculated. The
parameters are then estimated by, e.g., a Kalman filter. Three main advantages using
health parameters appear to be: (i) physical based model concept, (ii) the ability to
supervise component deteriorations, e.g., due to fouling, and (iii) remove trends in the
residuals due to component deteriorations. When the health parameters are used, two
main disadvantages appear and make the diagnosis procedure more difficult: (iv) the
observability of the system is affected, and (v) a faulty sensor value may be, in some
sense, captured by the health parameters. The maximum number of health parameters,
which can be considered in the model due to the observability criteria, is restricted by
the number of sensors. Thus, the maximum number of health parameters is equal to the
9.2. Gas Turbine Diagnosis Modeling 151

number of sensors. In practice, this maximum number of health parameters is much

smaller and depends on, e.g., where in the model the health parameters and sensors are
located. The disadvantage (v) can result in a test quantity which is sensitive to a specific
fault, but does not trigger since the health parameters compensate for the faulty sensor
value. Thus, the fault signature is not seen in the residuals since model uncertainties and
measurement noises are always present. To reduce this problem and increase the fault
isolation ability, an idea is introduce constraints of the health parameters to restrict the
allowed interval of deterioration.
In total, there are four health parameters (see Fig. 9.1) to estimate how the compo-
nents deteriorate. These parameters are added additive in the performance characteristic
equations, and for the compressor C1 these parameters are introduced to estimate deteri-
oration in: (i) the isentropic efficiency ∆η C1 , and (ii) the mass flow of air ∆ΓC1 . In the
compressor-turbine T1, two health parameters are introduced to estimate deterioration
in: (iii) isentropic efficiency ∆η T1 , and (iv) turbine flow capacity ∆ΓT1 .
In the DAE form, the model which taking into account the component deterioration
is written:

F ˆẋ, x, ∆, u 0 (9.2a)
˙
∆ 0 (9.2b)
y h ˆx  (9.2c)

where ∆ ˆ ∆η C1 , ∆ΓC1 , ∆η T1 , ∆ΓT1 T , and x denotes the unknown variables. Since

the component deterioration is slow relative to other gas turbine dynamics, the derivative
is considered to be zeros. Thus, the constraints in Eq. (9.2b) are added to the model.

9.2.3 Sensor and Actuator Faults

Slow drifting bias faults can be difficult to detect according to the methods presented
in Kobayashi and Simon (2003); Larsson et al. (2011). This depends on a fault signature
in the residuals which cannot be discriminated from the fault free case, i.e., no tests
trigger. To increase the fault detectability, the fault f i is modeled as an unknown additive
parameter that can be estimated in the test quantity. When the estimated fault differs too
much from zero, the fault may be detected by the FDI-system. This gives the diagnosis
models for the sensor faults:

Fk ˆẋ, x, u 0 (9.3a)
y k,i h k,i ˆx   f k (9.3b)
f˙k 0 (9.3c)

where k 1, . . . , 8, and i 1, . . . , 8. For the actuator faults the models are:

Fk ˆẋ, x, u k 8  f k  0 (9.4a)
y k,i h k ,i ˆx  (9.4b)
f˙k 0 (9.4c)
152 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

where k 9, 10, and i 1, . . . , 8. The derivative of the faults are set equal to zero in
Eqs. (9.3)–(9.4) but the dynamic of the faults can be adjusted when the CGEKFs are
designed. The faults, signals, and measured quantities are summarized in Tab. 9.1 and
coincide with Fig. 9.1. Since each fault is estimated correctly by one specific test quantity,
thus the test quantity which estimates the fault is not sensitive to the injected fault, i.e.,
the fault is decoupled in the test quantity by definition.

Table 9.1: Sensor and actuator Faults.

Signal Quantity Fault Mode

y1 , y2 , y3 p1 , p3 , p8 f1 , f2 , f3 F p1 , F p3 , F p8
y4 , y5 , y6 t2 , t3 , t7 f4 , f5 , f6 Ft 2 , Ft 3 , Ft 7
y7 , y8 n C1 , n T0 f7 , f8 Fn C1 , Fn C1
u1 , u2 m f , PA f 9 , f 10 F m f , F PA

9.2.4 Constraints on Performance Deviation

To isolate a fault, it is not enough to consider the estimated faults f i for the decision since
more than one estimated fault f i can deviate from zero for a given fault mode. For the
fault isolation procedure also the residuals must be considered in the decision process.
As mention earlier, slow drifting bias faults can be difficult to detect in the residuals.
Similar arguments are presented in Kobayashi and Simon (2005), where performed
investigations show that the residuals are small even when a sensor fault is present, since
the health parameters try to explain the fault by a shifting in value. Therefore, the number
of health parameter (and where in the model they appear) affects the fault signature in
the residuals. Thus, it is a balance between: (i) the ability to see a sensor fault in the
residuals, and (ii) the ability to not see the component deterioration in the residuals.
In Larsson et al. (2013), the number of health parameters were reduced to make the
sensor fault isolation procedure possible when experimental data was utilized. Since no
component degradation in, e.g., the compressor-turbine T1 was considered, false alarms
can be generated when the compressor-turbine component deteriorates. In Kobayashi
and Simon (2005), the problem is solved using an extra residual where the weighted sum
of squared health parameters is compared with a threshold. Thus, using the threshold
an alarm can be triggered when the deviation from the healthy base line is too large.
Instead of using the weighted sum, herein a constant upper and lower limit of the health
parameters are introduced which is presented in Tab. 9.2. This is a compromise between
the usage of many health parameters and the fault isolation performance. Since the
compressor should be washed when the performance has been reduced about 2  3 %
Scott (1986) a lower limit of 4.5 % is no problem. It is not common that the performance
of the compressor increases, therefore a value in the range of 0.5  1.0 % is reasonable.
The performance of the turbine is more unsure, therefore a band interval of 2  3 % is
introduced for those health parameters.
Since these limits are used only for fault isolation and not for fault detection other
9.2. Gas Turbine Diagnosis Modeling 153

Table 9.2: Health parameter upper and lower limits.

Health Parameter Lower Limit Upper Limit

∆η C1 -4.5 % 0.5 %
∆ΓC1 -4.5 % 1.0 %
∆η T1 -3.0 % 3.0 %
∆ΓT1 -2.0 % 2.0 %

types of constraints are possible to consider. At least, the upper limit for the health
parameters of the compressor could be more dependent how fouled the compressor is
when the fault detection alarm is triggered.

9.2.5 Differential Algebraic Equation Form

The fault models in Eqs. (9.3)–(9.4) are combined with the deterioration description (9.2)
to get models in the form similar to Eq. (9.1). From these models D i , the test quantities
used in the FDI-system are generated. Depending of which fault that is decoupled,
different health parameters are used to estimate the component deteriorations since the
health parameters influence the observability and the ability to give reliable estimates
when experimental data is evaluated. To decouple a fault in the sensor measuring the
speed n C1 of the gas generator is especially difficult and gives poor estimations of the
state variables if all the four health parameters are considered. When simulated data
is considered, it is easier to get reliable estimation of the state variable. The focus in
the work is to get reliable estimations of the state variables when experimental data is
considered, therefore one health parameter is removed for the fault mode Fn C1 , F p 3 , Ft 3 ,
and Ft 7 . These fault modes are also most difficult to isolate from each other if no limits
of the health parameters are considered in the diagnosis test quantity.
Fault Free Model: The diagnosis model D 0 in the non-faulty case has the DAE form:

F0 ˆż, z, u 0 (9.5a)
y h ˆz  (9.5b)
z ˆ x, ∆0 T (9.5c)

where ∆ 0 ˆ∆η C1 , ∆ΓC1 , ∆η T1 , ∆ΓT1 T , and x consists of the unknown variables.

Sensor/Actuator Fault Model: The diagnosis model D k , in case of a fault, has the
DAE form:

Fk ˆż, z, u 0 (9.6a)
y h ˆz  (9.6b)
z ˆ x, ∆ k , f k T (9.6c)
154 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

where index k 1, 2, . . . , 10, ∆ k consists of the health parameters, and x consists of the
unknown variables. Depending of the fault model, the vector ∆ k has health parameter
elements shown in Tab. 9.3.

Table 9.3: Health parameters utilized in diagnosis model D k .

D0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10
FNF F p1 F p3 F p8 Ft 2 Ft 3 Ft 7 Fn C1 Fn T0 Fm f F PA
∆η C1 x x x x x x x x x x x
∆ΓC1 x x x x x x x x x x
∆η T1 x x x x x x x x
∆ΓT1 x x x x x x x x x x x

9.2.6 State Space Form

The purpose of the diagnosis model is to introduce faults and health parameters in an
easy manner. Since the diagnosis model is equation based, where algebraic and dynamic
constraints are mixed, some equation transformations are necessary to perform to get it
into a state space form. The first step is to transform the set of equations in (9.5)–(9.6)
automatically to a form which can be used in a diagnosis test quantity based on state
space observers. To get the test equation, two main steps are performed: (i) check and
reduce the DAE-index of the system where Pantelides algorithm (Pantelides, 1988) is
utilized, and (ii) find the over-determined part and remove the exactly determined part
using the Dulmage-Mendelsohn decomposition (Dulmage and Mendelsohn, 1958). For
these two steps, the structural model (Blanke et al., 2003) of the diagnosis model is used.
For a more detailed overview of each step in the design procedure, see Larsson et al.
(2014a) since these transformation steps are not trivial. After the transformations, the
test equations of diagnosis model D k in a state space form are:

ẋ 1 f k ˆx 1 , u  (9.7a)
y h k ˆx 1  (9.7b)

where x 1 represents the dynamic variables, u represents the known input signals, and
y represents the known measurements. The functions f k and h k have appropriate di-
mension. The index k is the same as used in (9.5)–(9.6). In the vector x 1 , the health
parameters and the faults are included.

9.3 Test Quantity Design

The main objective in this section is to present how the test quantities, used in the FDI-
system, are designed. Each test quantity Ti consists of a nonlinear Kalman filter which is
generated from the diagnosis models D i in Eqs. (9.5)–(9.6). The outputs from the test
9.3. Test Quantity Design 155

quantities, except for the non-faulty case, are: (i) a logical value L i which represents
f
if the test has triggered or not, and (ii) a logical value L i which represents if a fault is
detected or not, and (iii) an estimation fˆi of the fault f i . When none of the considered
fault modes are present in the system the fault estimates fˆi should be small, and L i , L i
f

should not triggered. Thus, when the fault estimate in any of the tests is large, a fault
is present in the system and a fault detection alarm should be triggered. To detect the
fault, the fault signal fˆi are filtered using the CUSUM algorithm (Page, 1954; Gustafsson,
2000) and compared with a static threshold. When any of the fault estimates exceed its
threshold, a fault detection alarm is triggered.
j
To determine if the test quantity Ti should alarm or not, the output residuals r i from
the filters are first filtered using the CUSUM algorithm. Then a logical OR operation
j
is applied to the CUSUM tests Ti to decide if the test quantity is triggered or not. This
j
means that if any of the CUSUM tests Ti are triggered, the test quantity Ti will also
trigger an alarm. The test quantity Ti is then used in the fault isolation procedure. A
schematic view of the test quantity is shown in Fig. 9.2.

r 1i Ti1 A J i1
y
CGEKF i

Li
u
   OR
CUSUM

r 8i Ti8 A J i8
f
fˆi Li
f
A Ji
fˆi

Figure 9.2: Test quantity Ti of diagnosis model D i .

9.3.1 Constant Gain Extended Kalman Filter

To estimate state variables and component deterioration in the gas turbine, a common
approach in the gas turbine diagnosis literature is to use filters. The filters are often
Kalman based and are linear in Luppold et al. (1989); Litt (2007); Simon and Armstrong
(2013); Simani and Patton (2008) and nonlinear in Kobayashi et al. (2005); Naderi et al.
(2011); Dewallef et al. (2006). A special type of nonlinear Kalman filters is the constant
gain extended Kalman filter (CGEKF) (Safonov and Athans, 1978) which was investigated
in a gas turbine application in papers Kobayashi et al. (2005); Sugiyama (2000) with
success. Here, the test quantity of each diagnosis model D i is written in the form:
x̂˙ i f i ˆx̂ i , u  K i ˆ y  ŷ i  (9.8a)
ŷ i h i ˆx̂ i  (9.8b)
156 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

where the functions f i and h i can be nonlinear, u represents the input signals, y represents
the measurement signals, and K i is the feedback gain which can be considered as a design
parameter. For the CGEKF the constant matrix K i is calculated in advance for a given
stationary point of (9.7) using the noise covariance matrices Q i and R i . The covariance
matrix R i is diagonal with elements belonging to the standard deviation of measurement
noise for each sensor which is presented in Tab. 9.4. The covariance matrix Q i is also
diagonal where the elements representing the model uncertainty of each state variable.
Especially the states of the health parameters have been given small uncertainty variances,
i.e., the feedback term for these states have a small impact on the estimations.

Table 9.4: Standard deviation σ of the measurement noise.

Signal Quantity σ
y1 , y2 , y3 p1 , p3 , p8 0.5 %
y4 , y5 , y6 t2 , t3 , t7 0.5 %
y7 , y8 n C1 , n T0 0.2 %
u1 , u2 m f , PA 0.2 %

9.3.2 CUSUM Filtering

For change detection in signals, a suitable choice is to use the CUSUM algorithm (Page,
j
1954; Gustafsson, 2000). The CUSUM algorithm, together with a static threshold J i are
used to decide if the residual r i and fault estimate fˆi indicates an abnormal behavior
j

j
according to the sensor or actuator faults. The CUSUM test quantity Ti for each residual
is computed
j j j j j
Ti ˆt  maxˆ0, Ti ˆt  1  Sr i ˆt S  ν i , Ti ˆ0 0 (9.9)
j
where t is time, i > 0, . . . , 10, j > 1, . . . , 8, ν i is a tuning parameter to ensure that
j j
Ti ˆt  @ 0 in the fault free case, and Sr i ˆt S is the absolute value of residual in Eq. (9.10).
j
The threshold J i J is more or less equal for all CUSUM tests since the normalization of
j
the residuals in (9.10) was performed. The test quantity triggers an alarm when Ti ˆt 
j
exceeds the specified threshold J. The design parameter ν i is tuned so the test does not
give any unnecessary alarms in the fault free case, i.e., no false alarms. A good rule of
j j
thumb is that ν i has the same order of magnitude as the residual r i in the fault free case,
i.e., here the parameter is approximately 1. The CUSUM filtering of the fault estimate f i
j
is analogue with test quantity Ti filtering procedure.
In an ideal case, when a slow varying fault f i is present in the sensor or actuator
signals all test quantities except Ti should trigger. In practice, this is not always true since,
e.g., measurement noise, fault size, and model uncertainty are always present together
with the tuning procedure of the CGEKFs and the CUSUM tests. All these things affect
the detectability and reaction rate of each test quantity.
9.4. Fault Isolation Method 157

Residuals
The output from each Kalman filter is an estimation of the state variables x̂ i and the
sensor values ŷ i . The estimated sensor values from each filter are used to construct
j
residuals r i in the form:
j j j j
r i ˆ ŷ i  y i ~σ i (9.10)
j j
where i > 1, . . . , 10, j > 1, . . . , 8, and σ i is the standard deviation of r i in the fault
free case, i.e., the same noise as presented in Tab. 9.4. Since the measurement sequence
is available offline, the residuals can be normalized with the standard deviation of the
residuals in the fault free case. With this normalization, the signals are in the same
interval.

Faults
The outputs from the CGEKFs are, e.g., estimations of the faults f 1 , . . . , f 10 . When no
faults of the considered fault modes are present in the system, the fault estimates should
be low. When any of the fault estimates are large, the system is no longer non-faulty and
a fault detection alarm should be triggered. To determine if any of the fault estimates are
large, the CUSUM algorithm is used together with a static threshold. Since a specific
fault can affect fault estimates in many tests, the fault detection signal can only be used
for fault detection and not for isolation.

9.4 Fault Isolation Method

The main objective in this section is to present how the isolation procedure of single
sensor and actuator faults is performed using the developed FDI-system shown in Fig. 9.3.
The inputs to the FDI-system are known quantities such as sensor and actuator signals.
The outputs are the estimated states x̂, a suggestion if the compressor should be washed
W, and a set of diagnoses D which are consistent with the observations. The FDI-system
is based on hypothesis testing and there is one hypothesis for each sensor and actuator
failure. Such an approach is standard in general diagnosis methodology (Blanke et al.,
2003) and similar approach was also utilized in Larsson et al. (2010); Kobayashi and
Simon (2005). Thus, let H i , i 1, . . . , 10 denotes the hypothesis that the sensor y i or
actuator u j is faulty and let H 0 denote the null-hypothesis that no sensor or actuator is
faulty. For each fault hypothesis H i the test quantity Ti is designed given the diagnosis
model D i where an additional state variable f i is introduced in the measurement sensor
y i or in the actuator signal u j . When hypothesis H i is true, all filters except the filter
in Ti cannot estimate the faulty sensor or actuator value correctly and the filters have
ideally a large error in the residuals. When hypothesis H i is true, the test Ti is used
to estimate the correct state variables which can be used for supervision and control
purposes. Since the fault is already estimated, this estimate is used to compensate for the
faulty input signal to the test T0 . Thus, when a fault is diagnosed it is assumed that the
fault is decoupled in the test quantity T0 .
To reject the null-hypothesis H 0 , some of the fault estimates fˆ1 , . . . , fˆ10 should indi-
cate a major deviation from zero. When the null-hypothesis is rejected, the system can
158 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

u, y

T1 T2 ... T10
f f f
fˆ1 L 1 L 1 fˆ2 L 2 L 2 fˆ10 L 10 L 10

T/F D FAULT
DETECT ISOLATE
COMP.

∆
WASH T0

W D x̂

Figure 9.3: FDI-system.

no longer be non-faulty. If the hypothesis H 0 is not rejected, the system is assumed to be

non-faulty and the input signals to the test quantity T0 are not modified and the test can
directly be used to estimate the performance.

The Kalman filters are used to estimate the outputs ŷ and combine these outputs with
the real measurement signals y to get the residuals. The residuals are then filtered using
j
CUSUM and compared with a given threshold value J i , which results in a logical value
L i . The logical values L i are then used in the fault isolation logical component ISOLATE
(Fig. 9.3) where the diagnoses consistent with the hypotheses are determined. In the
component FAULT COMP., the measurement and actuator signals are compensated
for the fault in the signals when a unique diagnosis D is stated. The test T0 is then
used to estimate the state variables which can be used to, e.g., supervise the turbine
inlet temperature TIT and the degree of compressor fouling. The component WASH
determines if the deterioration level of compressor performance is sufficient to perform
a compressor wash. After the wash, the health parameters for the compressor are reset.
In the component DETECT, it is determined if the H 0 hypothesis should be rejected
or not. When the H 0 hypothesis is not rejected, the ISOLATE component is not active
since it is assumed that the system is non-faulty.
9.4. Fault Isolation Method 159

9.4.1 No Fault Hypothesis H 0

When the gas turbine system is non-faulty, no tests in the FDI-system should trigger. To
reduce the number of unnecessary false alarms, and increase the fault detectability the
fault estimates: fˆ1 , . . . , fˆ10 are used to determine if the H 0 hypothesis should be rejected
or not. To trigger a fault detection alarm it is enough if any of the CUSUM tests of fˆi
exceeds its threshold. If hypothesis H 0 is not rejected, it is assumed that the system is
non-faulty.

9.4.2 Fault Hypothesis H i

When one or more of the CUSUM tests of the fault estimate fˆi trigger, the H 0 hypothesis
is no longer valid. For each test Ti that has alarmed, the fault hypothesis H i can be
removed from the possible set of valid hypotheses. When only one test T j is left in the
not alarmed test set the only valid hypothesis is H j . This results in the single sensor
or actuator diagnosis D j . When a single diagnosis D j is obtained, the estimated fault
f j is used to compensate for the faulty signal in the test quantity T0 . Depending on
the actual diagnosis, and the current component deterioration, the accuracy of the
estimated fault varies since different numbers of health parameters are used in the test
quantities T1 , . . . , T10 . For example, mass flow degradation in the compressor C1 can
affect the estimation of fault f 7 (which will be shown in Fig. 9.10). To compensate for
this phenomenon, the detection threshold for fault f 7 should be adjusted little higher.
Since all input and output signals are used in each test quantity, the test is in theory
sensitive to all faults which mean that ideally a fault cannot be decoupled and an isolation
procedure is not possible. It takes some time for the fault parameter fˆi to estimate the
fault f i , and when the fault has reached the correct value it is assumed that the fault
is decoupled in the test quantity Ti and is not sensitive to the fault f i . Since it takes
some time for the fault estimate to reach the fault level, a correct diagnosis cannot be
guaranteed during this period of time.

9.4.3 Component Failure and Foreign Object Damage

For the case when all tests Ti alarm, no hypothesis is no longer valid since a single
fault assumption is considered. A probably scenario is that a sudden fault with large
amplitude has occurred and it takes time for the fault estimate to reach the fault level.
Another explanation is that a rapid change in component performance occurs, i.e., a
component failure (CF) or some kind of foreign object damage (FOD). Since upper
and lower limits (Tab. 9.2) of the health parameters are introduced, a large shift in the
performance parameters is visible in the residuals. The CF and FOD symptoms can only
be detected since no fault models for these kinds of symptoms are considered, thus it is
not possible to isolate these fault events. In Tab. 9.5, the test alarm scheme is summarized.
160 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

Table 9.5: FDI scheme.

Fault Number of Description Fault

Detected Alarms Li Estimated
0 0BnB9 NF, H 0 valid 0
1 1BnB8 Multiple H i :s valid 0
1 n 9 Unique H i valid 1
0/1 n 10 CF, FOD, MF 0

9.5 Simulation Results

The main objective in this section is to investigate how the proposed FDI-system responds
to the various fault scenarios when: (i) performance deterioration is introduced in the
components, (ii) the speed and power of the power turbine are varied, and (iii) noise
in the measurement signals is introduced. Since data from the simulation platform
presented in Larsson et al. (2014a) is considered, the case study can be performed under
controlled forms and the correct values of the parameters are also known for comparison.
The secondary objective with the case study is to present and compare how the introduced
limits of health parameters influence the reaction ability of the tests.

9.5.1 Simulation Setup

The input signals to the simulation platform are: (i) demand power, (ii) frequency
of the driven component, (iii) ambient condition such as pressure p 0 , temperature t 0 ,
and relative humidity φ 0 of the ambient air. In the following simulation plots, only the
demand power and frequency of the driven component are varied. The other inputs
are kept constant for simplicity. Further studies, which have been performed and not
shown here, indicate that it is possible to also varying the ambient conditions with similar
results. The change in frequency of the driven component is a direct function of the
speed of the power turbine. Therefore, the speed is shown here and not the frequency. In
the simulation platform, it is also possible to introduce slow performance deterioration
in components. Here, deteriorations in compressor C1, and turbine T1 are considered.
The injected performance deterioration is presented in Tab. 9.6 were also the slope of
the deterioration is shown. Given the simulation setup, the controller calculates the
opening angles for the valves in the fuel system. Given the angles, the mass flow of fuel
is calculated by the model and the mass flow is considered to be an input signal to the
FDI-system. After the data is collected, Gaussian noise with zero mean and standard
deviation according to Tab. 9.4 is added to the signals. The simulated data which is used
as inputs to the FDI-system is shown in Fig. 9.4. A higher demand power increases the
fuel consumption u 1 , the compressor discharge pressure p 3 and temperature t 3 , exhaust
temperature t 7 , and speed of the gas generator n C1 .
9.5. Simulation Results 161

1.4
16
1.0 14

(a) y 1 bar (b) y 2 bar

1.1 310

300

1.0 290
(c) y 3 bar (d) y 4 K

720 850
800
680
750
640 700
(e) y 5 K (f) y 6 K

8
9.6
7
9.4
6
9.2
5
(g) y 7 krpm (h) y 8 krpm

1.6 28
1.4 24
1.2 20
1.0 16
30 90 150 30 90 150
(i) u 1 kg/s (j) u 2 MW

Figure 9.4: Simulated measurement signals corresponding to a variation in applied power simul-
taneously the performance in C1 and T1 deteriorates. The measurements are gathered using
the simulation platform where Gaussian noises are added to the signals according to Tab. 9.4.
162 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

Table 9.6: Performance Deteriorations and Control input Signals.

Performance Injected Duration

Parameter Deterioration [%] [days]
η C1 3.0 130
ΓC1 2.5 130
η T1 1.0 130
ΓT1 1.0 130
Demand Power Step [MW] Duration [days]
PA -2 27
Demand Speed Step [rpm] Duration [days]
n T0  1500  50

9.5.2 Fault Modes

For the given scenario of input signals and component deteriorations, faults with ampli-
tudes given in Tab. 9.7 are investigated in the simulation study. Two different kind of bias

Table 9.7: Fault modes and amplitude of injected sensor or actuator faults.

Fault Mode Signal Quantity Amplitude

FNF - - -
F p1 , F p8 y1 , y3 p2 , p8  4%
F p3 y2 p3 4 %, 5 %

Ft 2 y4 t2  3%

Ft 3 y5 t3 4 %, 3 %

Ft 7 y6 t7 3 %, 5 %

Fn C1 , Fn T0 y7 , y8 n C1 , n T0  5%

F m f , F PA u1 , u2 m f , PA 5 %, 10 %

fault signatures are investigated which propagate: (i) abrupt (step fault), and (ii) slowly
drifting bias (ramp fault) to a given fault amplitude. The fault amplitudes in the table are
chosen in a way so the faults can be detected and isolated by the FDI-system. For some
faults, the amplitude is not symmetric, and it also depends where in the sequence the
fault is injected since also component deteriorations are present in the data. The faults in
the input signals are large because they are difficult to isolate from each other, but it is
possible to isolate those faults from the other faulty modes even with smaller amplitude.

9.5.3 Fault Free Mode: FN F

In the fault free case, the FDI-system uses the input-, and output signals shown in
Fig. 9.4. No false alarms are generated by the FDI-system, and the estimated performance
9.5. Simulation Results 163

deterioration (blue line) is shown in Fig. 9.5 together with the injected deterioration (red
line). As the figure indicates, it is no problem to estimate the component deteriorations
in the fault free mode although the applied power and speed of the power turbine are
varied.

0
0
-2
-2
-4
(a) T0 : ∆η C1 % (b) T0 : ∆ΓC1 %

0 0

-1
-2
30 90 150 30 90 150
(c) T0 : ∆η T1 % (d) T0 : ∆ΓT1 %

Figure 9.5: Estimated (blue line) and injected (red line) performance deterioration corresponding
to the measurement signals in Fig. 9.4. The estimation is performed using the test quantity T0
in the FDI-system together with a data set where no fault is injected.

9.5.4 Sensor Fault Modes: F p3 , Ft3 , and Ft7

To detect fault in sensors measuring: (i) pressure p 3 , (ii) temperature t 3 , and (iii) tem-
perature t 7 can be little tricky since the health parameters compensate, in some sense, for
the faulty sensor value. The investigated step fault can be seen in the residuals during a
short period of time, but when the fault has decline it is difficult to discriminate the faulty
sequence from the non-faulty measurements since noise is present. A slow varying fault
is more difficult to isolate since the fault behaves similar to component deteriorations
and therefore are the tests not triggered. To reduce the problem, an upper and lower
limit of the health parameters is introduced according to Tab. 9.2. The purpose with the
limitations is to prevent the health parameters to compensate for the fault and hence
have residuals that can be discriminated from the non-faulty case. In Fig. 9.6–9.7, a
step fault with an amplitude of 4, and 3 % is injected in the temperature sensor y 5
measuring t 3 and the results are shown for the test quantity T2 which decouple a fault
in the sensor measuring pressure p 3 . Since T2 estimates a fault in p 3 , this test quantity
should trigger an alarm. When the 3 % fault is injected in the sensor (blue line), the
∆ΓT1 parameter reacts drastically and after some time the residuals go back to zero as
shown in Fig. 9.7. For better isolation performance, an upper and lower limit of the
health parameters are introduced for the two other fault scenarios which gives a more
distinct fault signature (red and green line) in the residuals. The limits of the health
parameters does not affect the signal used for detection, since the correct hypothesis test
164 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

0 0
-2 2

-4 4

(a) T2 : ∆η C1 % (b) T2 : ∆ΓC1 %

5
0
0
-4
-8 -5
30 90 150 30 90 150
(c) T2 : ∆ΓT1 % (d) T5 : fˆ5 %

Figure 9.6: Health parameter estimates corresponding to three different fault cases where a step
fault is injected (at day 70) in the sensor y 5 which measure the compressor discharge tempera-
ture T3 . The estimations are performed using a fault with an amplitude of 3 % (red line), with
an amplitude of 4 % (green line), and with an amplitude of 3 % (blue line). For the two first
fault cases, a maximum and minimum limit of the deterioration level is introduced to maintain
higher fault isolability using the residuals. The limits are not affecting the fault detectability
since test T5 estimates the fault correctly (d).

T5 estimates the injected fault correctly as shown in in Fig. 9.6d. The goal with the health
parameters is to decouple component deteriorations, and when the health parameter
hits the limits this ability can disappear which Fig. 9.7b indicates. In the figure, it is clear
that the component deterioration affects the residuals when the health parameters had
hit theirs limits. But here, only one residual in the test needs to exceed its threshold to
trigger an alarm. Thus, it is sufficient that residual r 8 reacts clearly.

9.5.5 Actuator Fault Modes: Fm f and FPA

Two actuator faults (mass flow of fuel m f and generated power PA ) are considered in the
simulation study. The investigation shows that faults in the input signals are difficult to
isolate from each other, but it is easier to isolate these faults from faults in the output
signals. However, to isolate faults in the input signals from each other require larger fault
amplitudes, since tests T9 , and T10 are not triggered for respective fault. The residuals
which give best reactions, for each fault with a positive and negative amplitude, are shown
in Fig. 9.8.
To summarize the test reaction, all test quantities are viewed in Fig. 9.9 where a step
fault with an amplitude of +5 % in the input signal u 1 is diagnosed. All tests alarm just
after the fault is introduced, but after a short period of time the test T9 stops trigger,
which results in the only valid hypothesis H 9 . At day 150, the component deterioration
has compensated for the fault in test T10 and the test is not alarming anymore. Therefore,
the fault size is increased to 10 % to get distinct diagnoses for the whole sequence.
9.5. Simulation Results 165

0.5 20

0.0 0

-0.5 -20
(a) T2 : r 2 bar (b) T2 : r 5 K

15 50

0 0

-15 -50
30 90 150 30 90 150
(c) T2 : r 6 K (d) T2 : r 8 K

Figure 9.7: A subset of the residuals corresponding to the health parameter estimation in Fig. 9.6.
With the introduced limits of the health parameters, the separation from the fault free case is
more distinct.

0.7 150

0 0

-0.7 -150
30 90 150 30 90 150
(a) T10 : r 2 bar (b) T9 : r 8 rpm

Figure 9.8: Two residuals corresponding to a fault in u 1 (a), and in u 2 (b). The amplitudes of the
faults are 5 % (green line), and 10 % (red line). The residuals are performed using the test
quantity T10 in (a), and test quantity T9 in (b).

9.5.6 Sensor Fault Mode: FnC1

In the final test case, a fault in the sensor measuring speed n C1 of the gas generator
is investigated. This fault case is interesting since the test T7 does not compensate
for deterioration in mass flow of the compressor. When the health parameter ∆ΓC1 is
considered in the model, it is difficult to get state variable estimates which not diverge
and the parameter is therefore omitted in the model. It is no problem to detect and
isolate a fault with an amplitude 5 %, and all tests except T7 are triggered.
The estimated fault fˆ7 is used to compensate for the fault in the inputs to the T0
test. The result of the estimations is shown in Fig. 9.10, where all health parameters
except ∆η C1 follow, more or less, the injected degradation trajectory. As expected, the
estimation of ∆ΓC1 in Fig. 9.10b is zero since the test T7 does not estimate the degradation.
Instead, the deterioration is captured by the fault parameters which Fig. 9.10e–f indicates.
The impact of the degradation in the fault estimate is much smaller than the injected
fault which not triggers any alarms, i.e., false alarms.
166 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

(a) T1 (b) T2 (c) T3 (d) T4 (e) T5

30 90 150 30 90 150 30 90 150 30 90 150 30 90 150

(f) T6 (g) T7 (h) T8 (i) T9 (j) T10

Figure 9.9: Logical decision of the test quantities T1 , . . . , T10 when a fault with amplitude 5 %
in the input signal u 1 is diagnosed. The logical test T10 is not triggered after day 150 since the
residual goes towards zero similar to the results in Fig. 9.8a.

0
0
-2
-2
-4
(a) T0 : ∆η C1 % (b) T0 : ∆ΓC1 %

0 0

-1
-2
30 90 150 30 90 150
(c) T0 : ∆η T1 % (d) T0 : ∆ΓT1 %

5 2
0
0 -2

(e) T7 : fˆ7 % (f) T7 : f 7  fˆ7 %

Figure 9.10: Health estimation by test T0 (blue line), and injected performance deterioration (red
line) corresponding to a fault case with a fault in the sensor measuring the speed n C1 of the
gas generator. All health parameters except ∆ΓC1 follow the injected component deterioration.
In (e), the estimated (green) and injected (red) fault is shown. In (f), the difference between
the estimated and injected fault is shown and represents, in some sense, the injected mass
flow deterioration.
9.6. Experimental Case Studies 167

9.5.7 Summary
The simulation study shows that all (step and ramp) faults presented in Tab. 9.7 are de-
tectable with the developed FDI-system simultaneously as the component deteriorations
are supervised. All faults are isolable from each other except faults in: (i) the sensor mea-
suring n T0 , and (ii) the input signal PA . Thus, it is not possible to discriminate between
a fault in the speed of the power turbine and the demand power for the simulated data.
Faults in the sensors measuring: (i) discharge pressure p 3 , (ii) discharge temperature t 3 ,
and (iii) exhaust temperature t 7 are more difficult to isolate than the other faults since
the fault signature may disappear in the residuals. To reduce the problem, an upper and
a lower limit of the health parameters are introduced. The disadvantages with using an
upper and a lower limit of the health parameters are: (i) it can be difficult to find/use
suitable values of the limits, and (ii) when the health parameter has reach the limit the
component deterioration is no longer decoupled in the test. For the second disadvantage,
this can be a problem only for the diagnosis procedure since the correct fault hypothesis
estimates the injected fault correctly.

9.6 Experimental Case Studies

In this section, the design of the constructed fault tolerant FDI-system is evaluated
using experimental data from the gas turbine mechanical drive site. The objectives
of the evaluation are to: (i) detect and isolate faults in sensor and actuator signals,
(ii) supervise the health state of the engine, and (iii) use the FDI-system to suggest when
it is a good idea to, e.g., wash the compressor due to fouling. The main focus with the
investigation of the FDI-system is to evaluate how well the fault tolerant supervision
system performs in the presence of a single sensor or actuator fault. This question is
important since it is difficult, for some specific faults, to discriminate between a fault
and degradation in performance especially when the fault propagation is slow. At the
gas turbine site, the ambient air consists of a high grade of air pollutions such as: sand,
salt, oil, and other contaminations which affect the engine’s performance. The high
grade of air pollutions results in compressor fouling quite frequently, thus the service
personnel needs to shut down the gas turbine for maintenance, at least, each quarter.
In Scott (1986), a compressor wash is suggested when the efficiency of the compressor
has deteriorated 3 %. At SIT, a common procedure is to wash the compressor before the
isentropic efficiency and the mass flow have deteriorated 3 % from the nominal base line.
The focus in the present paper is not to decide a precise limit when it is time to wash the
compressor. Instead, a wash condition interval is chosen based on the value given by
Scott (1986) and the experimental data given by SIT where the fouling increases with the
decreased performance. Thus, the wash condition interval is chosen to be 2.0  3.5 %
for the efficiency health parameter ∆ηC1 and 2.0  3.5 % for the flow health parameter
∆ΓC1 . The band gap is 1.5 percent points for the both cases. In the experimental data,
the compressor is washed at day: 15, 56, and 144 which the estimation of performance
deterioration shown in, e.g., Fig. 9.12 (blue line) indicates. Since sensor and actuator faults
may affect the performance parameters in the test quantity T0 , it is desirable if these faults
168 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

can be detected and isolated as soon as possible to avoid unnecessary compressor washes,
or a compressor which is not washed when it is significantly fouled. If the floor limit of
the wash condition interval is increased, the time to detect the fault is reduced, and vice
versa. Thus, it is a trade of between the ability to detect the component deterioration and
the fault detectability of the FDI-system.
The evaluation of the FDI-system is divided into three separate parts: (i) investi-
gate the fault detectability and isolability of the system, (ii) investigate the performance
estimation of deterioration in test quantity T0 when fault f i is decoupled, and (iii) in-
vestigate how well the test quantity T0 perform when a step fault is detected, isolated,
and compensated for in the signal (fault tolerant supervision). In the investigation, the
FDI-system is tuned so there are no false alarms generated using the experimental data
sequences.

9.6.1 Evaluation 1: Fault Detection and Isolation

The fault detectability and isolability of the FDI-system are investigated using two types
of fault signatures which propagate: (i) immediately (step fault), and (ii) slowly (ramp
fault) to a specific fault amplitude. A fault that deviates much from the performance
deterioration is easier to isolate. Thus, a step fault is easier than a ramp fault to isolate (for
a given fault amplitude) since it is assumed that the performance deterioration propagates
slowly, .i.e., rapid changes in the health parameters are not allowed. For example, it is
possible to have a fault which has a major impact on the compressor efficiency and
behaves similar as a deteriorated component. This results in a fault which makes it
difficult to discriminate between the fault and the fouling of the compressor. For the
fault detectability of the two extreme cases, the opposite is true. It is easier to detect a
ramp fault then a step fault since it takes some time for the estimated faults to reach the
thresholds which trigger the alarm. How fast the fault is detected depends on the actual
tuning of the CGEKFs.

Step Faults
The step fault detectability and isolability of FDI-system are evaluated using the fault
amplitudes given in Tab. 9.8. The fault is injected at day 60, which are some days after the
compressor is washed. The fault is injected directly after the compressor wash to have
enough time to detect and isolate the fault before the compressor should be washed due
to fouling. The sign of the injected fault also affect the FDI performance as shown in the
simulation study in Sec. 9.5.
For each fault mode, the detection and isolation times are shown in the Tab. 9.8. For
the isolation time, the diagnosis chosen by the FDI-system is also the correct diagnosis.
Almost all faults are detected immediately where also all test quantities Ti are triggered
which means that the behavior can not be described by the isolation logic. After some
time, the amplitude of the estimated fault f i by the FDI-system is in the same range as
the injected fault which results in a fault which is decoupled in test quantity Ti and T0 .
This gives the time values in the right column in the table. All faults are uniquely isolable
except for the fault modes: Ft 7 and FPA , where it is, e.g., not possible to isolate a fault
9.6. Experimental Case Studies 169

Table 9.8: The detection and isolation time when a step fault is introduced in the signals. Fault
modes Ft 7 and FPA can not be uniquely isolated with the given fault size.

Step Fault Detection Isolation

Mode Size Time [days] Time [days]
FNF - - -
F p1 +4 % 0.0 1.8
F p3 +4 % 0.2 0.7
F p8 +4 % 0.1 0.3
Ft 2 +3 % 0.2 0.2
Ft 3 +3 % 0.3 1.4
Ft 7 +3 % 0.1 2.8 (Ft 7 , Fm f )
Fn C1 +5 % 0.1 0.6
Fn T0 +5 % 1.4 1.4
Fm f +5 % 0.2 0.2
F PA +5 % 1.1 1.1 (Fn T0 , Fm f , FPA )

in the sensor measuring m f without increasing the fault amplitude. Similar arguments
where shown in the simulation study which was performed in the previous section.

Ramp Faults
To investigate how the FDI-system responds to a ramp fault, the fault amplitudes shown
in Tab. 9.9 are added to the input and output signals. The fault starts at day 60, and the
duration is 50 days until it reaches the maximum amplitude. If the fault is detected and
isolated fast, it is an indication that a smaller fault signature can be used in this specific
case, e.g., for faults in sensors measuring p 1 and n C1 .
For each fault mode, three time values are shown in Tab. 9.9. These values are: (i) the
time when the fault is detected, (ii) the time and fault modes when only three (or less)
diagnoses remain in the possible diagnosis set, and (iii) the time when the fault is uniquely
isolated (if possible). As shown in the table, faults in sensors measuring: p 1 , p 8 , t 2 , and
n C1 are easy to detect and isolate, while faults in sensors measuring: p 3 , t 3 , and t 7 are
more difficult detect and isolate which also the simulation study indicated.
For the injected slow varying faults, the primary objects are to detect and isolate
all faults before: (i) the compressor is washed at day 144, and (ii) the performance
has degraded to a point where a compressor wash alarm is triggered. If the fault is not
detected before day 144, it can be an indication that the fault has violated the compressor
wash condition through compensation in the performance deterioration. On the other
hand, if the compressor wash alarm is triggered before day 144, the fault has assisted the
deterioration rate which results in a too early wash. As shown in Tab. 9.9, all faults except
for the sensors measuring n T0 and PA are detected and isolated before the compressor is
washed or the wash condition alarm is triggered. The fault modes Fn T0 and FPA cannot
be isolated uniquely since it is difficult to discriminate between a fault in the speed n T0
and in the power PA generated by the power turbine. The fault cases Ft 3 and Ft 7 which
170 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

Table 9.9: Detection and isolation time when the faults start at day 60 and ends at day 110. The
amplitudes of the ramp faults are the positive values given in Tab. 9.7.

Ramp Fault Detection Multiple Diagnoses Isolation

Mode Size Time [days] Time [days] Mode Time [days]
FNF - - - All -
F p1 +4 % 2.9 2.9 F p 1 , Fn T0 2.9
F p3 +5 % 11.5 26.9 F p 3 , F t 7 , Fm f 44.6
F p8 +4 % 7.7 8.0 F p8 8.0
Ft 2 +3 % 12.9 13.0 Ft 2 13.0
Ft 3 +3 % 24.3 41.3 F t 2 , F t 3 , F PA 43.2
Ft 7 +5 % 11.3 29.5 F t 7 , F m f , F PA 53.5
Fn C1 +5 % 7.5 7.5 Fn C1 , Fm f 7.5
Fn T0 +5 % 29.3 40.3 Fn T0 , Fm f , FPA -
Fm f +10 % 10.5 14.8 F p 3 , F t 7 , Fm f 17.7
F PA +10 % 14.1 19.2 Fn T0 , FPA -

are most difficult to diagnose in the case study are presented in the remainder of this
section.
Fault Mode Ft7 : A fault with a positive amplitude of 5 % in the sensor measuring the
exhaust temperature t 7 of the power-turbine gives a reduction in the estimated value of
the mass flow throughout the compressor. As Fig. 9.11b indicates, this is compensated
through a reduction in the health parameter ∆ΓC1 . The reduction in ∆ΓC1 reduces the
ability to detect the fault in the residuals, thus the fault estimate in Fig. 9.11c is instead
used for fault detection.
In the figure, the health parameters ∆η C1 and ∆ΓC1 are shown for the faulty (red
line) and non-fault (blue line) case for the test quantity T0 (which has no extra fault
parameters). In Fig. 9.11c, the introduced fault is shown together with the estimated fault
by test quantity T6 (which decouples fault f 6 ). The estimate seems to match the injected
fault pretty well.
The fault is detected before the compressor wash alarm is triggered by the FDI-system,
and the fault is isolated about 40 days later. Without the fault detection alarm, the FDI-
system should not satisfy the condition to find the fault before the compressor wash alarm
is triggered. Since a fault alarm is generated, the service personnel have the choice to:
(i) continue the operation until a unique diagnosis is stated, or (ii) see which diagnoses
that are consistent with the measurements and analyze if it is possible exclude some of
the remaining diagnoses. An analysis can be to, e.g., exclude fault modes that represent
fault estimates which have not exceeded its threshold. Another technique is to see which
diagnoses that are left, e.g., about 20 days after the fault is detected, only the diagnoses
of: Ft 7 , Fm f , and FPA are left in the possible set of diagnoses. It can be useful to look at
the health parameters which are estimated by the tests: T6 , T9 , and T10 which have not
triggered, and see if some of them seem to be more likely than the others. Finally, the
fault is uniquely diagnosed after about 40 days. Thus, it is not possible to isolate the fault
9.6. Experimental Case Studies 171

0 Fault injected

-2
Wash condition
-4
(a) T0 : ∆η C1 %
0

-4 Wash condition
Fault isolated
Fault injected
Fault detected
-8
(b) T0 : ∆ΓC1 %

5 Fault detected

Fault isolated
0
30 90 150
(c) T6 : f 6 (blue line), and fˆ6 (red line) [%]

Figure 9.11: Estimation of performance deviation in the compressor for the cases of: no fault (blue
line), and sensor fault in y 6 (red line). The estimated (red line) and injected (blue line) fault
are shown in (c). The sensor fault affects the mass flow parameter at most.

uniquely before a wash condition is triggered. But the operator knows that a fault is
present before the wash condition alarm is triggered.

Fault Mode Ft3 : A fault with a positive amplitude of 3 % in the sensor measuring the
discharge temperature t 3 of the compressor gives a reduction in the estimated value of
the isentropic efficiency of the compressor. This is compensated through a reduction in
the health parameter ∆η C1 as Fig. 9.12a indicates.

The fault is detected in the lower limit of the wash condition interval where the ∆η C1
parameter suggests a compressor wash. The fault is isolated about 20 days after the fault is
detected. Since the fault is detected in the lower limit of the interval, the specific fault size
can be misinterpreted as a fouled compressor. Since the fault is not detected at this point,
it is not known if the ∆η C1 parameter represents the true deterioration or not which
is a disadvantage. If it is assumed that the both health parameters have to deteriorate
approximately in the same magnitude, a red flag could be raised when only one of the
two health parameters suggests a compressor wash. With this information, it is possible
to detect the fault earlier.
172 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

0
Wash condition
-4
Fault injected Fault detected
Fault isolated
-8
(a) T0 : ∆η C1 %

0 Fault injected

-2
Wash condition
-4
(b) T0 : ∆ΓC1 %

3 Fault detected

Fault isolated
0
30 90 150
(c) T5 : f 5 (blue line), and fˆ5 (red line) [%]

Figure 9.12: Estimation of performance deviation in the compressor for the cases of: no fault (blue
line), and sensor fault in y 5 (red line) when the test T0 is utilized. The estimated (red line) and
injected (blue line) fault are shown in (c). The sensor fault affects the efficiency parameter at
most.

9.6.2 Evaluation 2: Supervision Based on Fault Compensation in Test

Quantity T0
Before the results of the fault tolerant supervision technique are presented, it is interesting
to see how the test quantity T0 responds to fault free data where an estimation of fault f i
is used to compensate for a potential fault in the signals. The evaluation shows in some
sense how good the health estimation is when one sensor or actuator is removed, i.e.,
the fault is decoupled. Here, the fault free case is presented and the estimation of the
fault should ideally be zero. The evaluation is performed as follows: (i) assume that the
diagnosis based on fault mode F i is determined, (ii) use the estimated fault f i from the
test quantity Ti to compensate for the fault in the test quantity T0 , and (iii) compare the
health estimations with the non-faulty case. The key step here is to get a good estimation
of the fault, i.e., a small fault estimation error gives a small spread between the health
estimations for different fault cases. When the estimated fault is not zero, it will influence
the health parameters in T0 exactly in the same way as a fault.
The fault cases Ft 3 and Ft 7 have the largest deviation in health parameters and fault
estimations. These two fault cases are therefore shown in Fig. 9.13. With the chosen wash
9.6. Experimental Case Studies 173

-2
Wash condition
-4
(a) T0 : ∆η C1 %

-2
Wash condition
-4
(b) T0 : ∆ΓC1 %
2

-2
(c) T0 : ∆η T1 %

1
0
-1

(d) T0 : ∆ΓT1 %

1
0
-1
30 90 150
(e) T5 , T6 : fˆ %
Figure 9.13: Supervision of health for the fault free case (blue line) together with estimation of
faults based on: test quantity T5 (red line), and test quantity T6 (green line). For all three cases,
fault free data is used and the estimations are performed using test quantity T0 . In the last
figure, the estimated fault fˆi is viewed which should be zero in an ideal case.
174 Chapter 9. Diagnosis and Fault Tolerant Supervision of Industrial Gas Turbines

condition interval, the two health parameters for the compressor are inside the chosen
interval. When a fault in the sensor measuring the compressor discharge temperature t 3
is present, the estimation of efficiency degrades faster than for the non-faulty sequence.
Since the sensor which measuring the compressor discharge temperature is decoupled
and this sensor is important for the efficiency calculation, it is not surprising that the
estimation of efficiency differs from the fault free case. But when this sensor is decoupled,
it is possible to supervise the performance of the compressor using the fault tolerant
supervision system.

9.6.3 Evaluation 3: Fault Tolerant Supervision of Performance

When a fault is diagnosed, the idea is to compensate for the fault in the measurements.
The compensation is done using the estimation of fault f i which is calculated by the test
quantity Ti . Thus, when the correct diagnosis is stated, the fault is decoupled in test
quantity T0 .
The fault tolerant FDI-system is evaluated using a step fault with an amplitude of 3 %
in the sensor measuring temperature t 3 . The fault is injected at day 60, and is immediately
detected by the FDI-system at the same time as all the test quantities alarm. After about
two days the correct diagnosis is stated and the fault parameter in T5 has estimated the
injected fault. The ∆η C1 parameter is shown in Fig. 9.14 for the three cases where the
parameter is: (i) compensated for the fault when it is isolated (blue line), (ii) based on

0 Fault isolated
Fault injected

-2

-4
(a) T0 : ∆η C1 %

0
30 90 150
(b) T5 : f 5 (blue line), fˆin (cyan line) %

Figure 9.14: Fault tolerant supervision of the ∆η C1 parameter (blue line), when fault f 5 is injected.
The estimation is compared to the fault free estimation (green line), and the estimation of T0
based on fault estimation of T5 (red line) in the fault free case. The measurement signal y 5 is
compensated with the signal fˆi to get the fault tolerant supervision of the ∆η C1 parameter.

fault free data (green line), and (iii) based on the fault free case together the fault which
is estimated by the test T5 (red line) are all shown. In the second figure, the injected and
9.7. Conclusion 175

estimated fault is shown. For this test case, it is easy to detect the fault and also isolate
the fault before the wash condition alarm is triggered.

9.6.4 Summary
The experimental case studies show that all considered fault modes can be detected
where also the most of them are possible to isolate from each other. The faults which are
most difficult to isolate are faults in the sensors measuring discharge pressure p 3 and
temperature t 3 after the compressor, and exhaust gas temperature t 7 . Faults in the two
actuator signals are also difficult to isolate uniquely from each other, but can be isolated
from the measurement sensors. Faults which behave like a step is easier to isolate than a
slow drifting bias fault. But the detection performance is the same for the two fault types
given the same amplitude. To increase the isolation performance, an upper and lower
limit of the health parameters is considered. If these limits are not considered, a slow
drifting bias fault can behaves similar to a fouled component.

9.7 Conclusion
In this work, an FDI-system for an industrial gas turbine is designed using a model
based approach. The FDI-system consists of: (i) a bank of constant gain extended
Kalman filters, (ii) a fault detection component, (iii) an isolation component, (iv) and a
compressor wash detection component. For each filter, a fault is decoupled. The filter
where the fault is decoupled can be used to supervise the performance in case of a fault,
and the other filters should have large estimation errors and triggered. The FDI-system
is evaluated using: (i) data generated from a simulation platform, and (ii) experimental
data collected from a gas turbine site. To receive better isolation performance, a lower
and an upper limit of the health parameters are introduced. These limits have to be
adjusted to the actual performance condition of the engine. A tighter band limit can
result in better isolation performance since the fault signature is more visible in the
residuals. When the band is too tight, the natural performance deterioration cannot be
captured. Thus, it is a trade of between fault isolation and the estimation of health. The
fault detection performance is not dependent of the chosen interval limits since all faults
are observable and thus estimated correctly by the right test quantity.
After the investigation, the conclusions are: (i) all considered faults are detectable,
(ii) faults in all eight output signals can be isolated correctly from each other when the
fault amplitude is 5 %, (iii) component deteriorations are compensated using health
parameters, and (iv) a fouled compressor can be detected even when a single fault is
present.
 Chapter 10

Conclusion

When a diagnosis and supervision system is designed using model based diagnosis tech-
niques, an important factor which affects the fault detection and false alarm probability
is the accuracy of the embedded models. To have good diagnosis performance, it is
valuable to have diagnosis test quantities based on models which have high accuracy
since this is also reflected in the diagnosis decision. Thus, when designing a diagnosis
and supervision system of an industrial gas turbine it is crucial to consider: (i) physical
relationships such as mass and energy balances, (ii) thermodynamic gas properties such
as enthalpy, heat capacity, and entropy, and (iii) performance characteristics of the gas
turbine components. To simplify the construction of the gas turbine model, it is suitable
to collect these aspects in a component library, here called GTLib, using an equation
based modeling language. The gas turbine model is then built using the predefined
components in GTLib. In GTLib, the thermodynamic gas properties rely on the the
NASA Glenn Coefficients which are encapsulated by the gas model. The performance
characteristics rely on look up tables, which are based on measurements performed by
the manufacturer. The benefit with using GTLib is the reduction in model equations
and state variables in the gas turbine model compared to the original gas turbine model
developed by SIT. The reduction depends largely on the usage of the air/fuel ratio concept
instead of the mass fraction of substances in the gas mixture as proposed in the original
model. The constructed gas turbine model can be used for performance analysis and
as a base (diagnosis model) when diagnosis tests are extracted. In the diagnosis model,
a number of health parameters are introduced to capture performance deterioration
in the compressor, and compressor-turbine. In the same way, faults in components
can be introduced in the diagnosis model. The component faults can represent, e.g.,
leakages or increased friction in bearings. Since an equation based modeling framework
is considered, the diagnosis model is in an index-2 DAE form and cannot be treated as
an ordinary state space model without modifications. It is also shown that unobservable
state variables are present, but these modes disappear when the over-determined part
of the diagnosis model is considered. The over-determined part is determined using

177
178 Chapter 10. Conclusion

structural methods. The equations which represent the over-determined part are imple-
mented as CGEKF based test quantities for a chosen operational point and for a suitable
choice of the noise matrices Q and R.
In the second part of the thesis, the methodology is evaluated using case studies. In
the first part, the CGEKFs are evaluated with focus on: (i) simulated data generated
from the reference simulation platform when the ambient conditions and demanded
power are varied, (ii) the flame temperature T f which is an important parameter for
the controller, and (iii) the performance deterioration due to compressor fouling. The
conclusion from the simulation study is that large deviations in the ambient conditions
affect the estimated health parameter in the same range as a fouled compressor, i.e.,
a deterioration of 1–2 %. To minimize the error depending on the change in ambient
conditions, the observers can compensate for the deviation in ambient conditions. In
the last two case studies, experimental data from a Middle East site are investigated. The
environment conditions at the Middle East site are tough due to contaminants in the
air, which results in frequently compressor washes. Using the estimated deterioration,
a decision of a fouled compressor should be taken. The investigation indicates that
it is possible to use the test in a FDI-system to detect a fouled compressor since the
deterioration of the compressor decreases about 2 % between the compressor washes.
Since the interval and offset of the deterioration are the same independently of the winter
or summer period, it is possible to have static thresholds for fouling detection in the
FDI-system. The flame temperature is estimated at the same time as the test quantity
compensates for the deteriorations in the compressor C1 and turbine T1.
Finally, to detect and isolate faults in actuators and sensors an FDI-system which
consists of a number of diagnosis test quantities is designed. The FDI-system can be used
to supervise the performance also when a fault has occurred. Slowly drifting bias faults
are more difficult than abrupt faults to diagnose, since the health parameters capture the
fault in some sense. Thus, it is difficult to discriminate the fault signature in the residuals
from the non-faulty case since measurement noise is present. With the constructed
FDI-system, the performance to detect a slowly drifting bias fault is the same (or little
better) as for an abrupt fault. To increase the fault isolation, a lower and an upper limit of
the health parameters are introduced. After the performed case studies, the conclusions
are: (i) all faults are detectable, (ii) faults in all eight output signals can be isolated
correctly from each other when the fault amplitude is 5 %, (iii) component deteriorations
are compensated using health parameters, and (iv) a fouled compressor can be detected
even when a single fault is present.
 References

G. F. Aker and H. I. H. Saravanamuttoo. Predicting gas turbine performance degradation

due to compressor fouling using computer simulation techniques. Journal of Engineering
for Gas Turbines and Power, 111(2):343–350, 1989. doi:10.1115/1.3240259. URL http:
//link.aip.org/link/?GTP/111/343/1.

Uri M. Ascher and Linda R. Petzold. Computer Methods for Ordinary Differential
Equations and Differential-Algebraic Equations. Society for Industrial and Applied
Mathematics, Philadelphia, PA, USA, 1st edition, 1998. ISBN 0898714125.

J. Bird and W. Grabe. Humidity effects on gas turbine performance. In International

Gas Turbine and Aeroengine Congress and Exposition, Orlando, FL, 1991.

M. Blanke, M. Kinnaert, J. Lunze, and M. Staroswiecki. Diagnosis and Fault-Tolerant

Control. Springer, 2003.

S. Borguet and O. Léonard. A sensor-fault-tolerant diagnosis tool based on a quadratic

programming approach. Journal of Engineering for Gas Turbines and Power, 130(2):
021605, 2008. doi:10.1115/1.2772637. URL http://link.aip.org/link/?GTP/130/
021605/1.

Gary L. Borman and Kenneth W. Ragland. Combustion Engineering. McGraw-Hill,

1998. ISBN 0-07-006567-5.

Olaf Brekke, Lars E. Bakken, and Elisabet Syverud. Compressor fouling in gas turbines
offshore: Composition and sources from site data. ASME Conference Proceedings, 2009
(48869):381–391, 2009. doi:10.1115/GT2009-59203. URL http://link.aip.org/
link/abstract/ASMECP/v2009/i48869/p381/s1.

Arden L. Buck. New equations for computing vapor pressure and enhancement factor.
Journal of Applied Meterology, 20:1529, 1981.

E. Buckingham. On physically similar systems; illustrations of the use of dimensional

equations. Phys. Rev., 4(4):345–376, Oct 1914. doi:10.1103/PhysRev.4.345.

179
180 Chapter 10. Conclusion

Francesco Casella, Martin Otter, Katrin Proelss, Christoph Richter, and Hubertus
Tummescheit. The modelica fluid and media library for modeling of incompressible and
compressible thermo-fluid pipe networks. The 5th International Modelica Conference
2006, 2006.
Malcolm W. Chase, Jr, editor. NIST-JANAF Thermochemical Tables, volume Part 1
and Part 2 of Journal of Physical and Chemical Reference Data. American Institute of
Physics, 500 Sunnyside Boulevard, Woodbury, New York, fourth edition, 1998. ISBN
1-56396-831-2.
J. Chen and R.J Patton. Robust Model-Based Fault Diagnosis for Dynamic Systems.
Kluwer Academic Publishers, 1999. ISBN 0-7923-8411-3.
P. Dewallef, C. Romessis, O. Léonard, and K. Mathioudakis. Combining classification
techniques with kalman filters for aircraft engine diagnostics. Journal of Engineering
for Gas Turbines and Power, 128(2):281–287, 2006. doi:10.1115/1.2056507. URL http:
//link.aip.org/link/?GTP/128/281/1.
I. S. Diakunchak. Performance deterioration in industrial gas turbines. Journal of
Engineering for Gas Turbines and Power, 114(2):161–168, 1992. doi:10.1115/1.2906565.
URL http://link.aip.org/link/?GTP/114/161/1.
S.L. Dixon and Cesare Hall. Fluid Mechanics and Thermodynamics of Turbomachinery.
Elsevier, 6th edition edition, 2010. ISBN 978-1-85617-793-1.
D. L. Doel. Temper—a gas-path analysis tool for commercial jet engines. Journal of
Engineering for Gas Turbines and Power, 116(1):82–89, 1994. doi:10.1115/1.2906813. URL
http://link.aip.org/link/?GTP/116/82/1.
D. L. Doel. Interpretation of weighted-least-squares gas path analysis results. Journal
of Engineering for Gas Turbines and Power, 125(3):624–633, 2003. doi:10.1115/1.1582492.
URL http://link.aip.org/link/?GTP/125/624/1.
A. L. Dulmage and N. S. Mendelsohn. Coverings of bipartite graphs. Canadian Journal
of Mathematics, 10:517–534, 1958.
Thomas D. Eastop and Allan McConkey. Applied Thermodynamics for Engineering
Technologists. Prentice Hall, 5th edition, 1993. ISBN 0-582-09193-4.
Lars Eriksson. CHEPP – A chemical equilibrium program package for matlab. In
Modeling of Spark Ignition Engines, number 2004-01-1460 in SAE Technical paper series
SP-1830, 2004. doi:10.4271/2004-01-1460.
Erik Frisk, Mattias Krysander, and Jan Åslund. Sensor placement for fault iso-
lation in linear differential-algebraic systems. Automatica, 45(2):364–371, 2009.
doi:10.1016/j.automatica.2008.08.013.
Peter Fritzson. Principles of Object-Oriented Modeling and Simulation with Modelica
2.1. Wiley, 2004.
References 181

Ranjan Ganguli and Budhadipta Dan. Trend shift detection in jet engine gas path
measurements using cascaded recursive median filter with gradient and laplacian
edge detector. Journal of Engineering for Gas Turbines and Power, 126(1):55–61, 2004.
doi:10.1115/1.1635400. URL http://link.aip.org/link/?GTP/126/55/1.
Tony Giampaolo. Gas Turbine Handbook: Principles and Practice. The Fairmont Press,
4th edition edition, 2009. ISBN 0-88173-613-9.
Sanford Gordon and Bonnie J. McBride. Computer program for calculation of complex
chemical equilibrium compositions, rocket performance, incident and reflected shocks,
and chapman-jouguet detonations. interim revision, march 1976. Technical Report
SP-273, National Aeronautics and Space Administration (NASA), 1971. URL http:
//ntrs.nasa.gov/search.jsp?print=yes&R=19780009781.
Sanford Gordon and Bonnie J. McBride. Computer program for calculation of complex
chemical equilibrium compositions and applications i. analysis. Technical Report
RP-1311, National Aeronautics and Space Administration (NASA), 1994. URL http:
//www.grc.nasa.gov/WWW/CEAWeb/RP-1311.htm.
Fredrik Gustafsson. Adaptive filtering and change detection / Fredrik Gustafsson. Chich-
ester : Wiley, cop. 2000, 2000. ISBN 0471492876.
A. A. El Hadik. The impact of atmospheric conditions on gas turbine perfor-
mance. Journal of Engineering for Gas Turbines and Power, 112(4):590–596, 1990.
doi:10.1115/1.2906210. URL http://link.aip.org/link/?GTP/112/590/1.
E. Hairer, S. P. Norsett, and G. Wanner. Solving Ordinary Differential Equations II: Stiff
and Differential-Algebraic Problems. Springer, Berlin, 2nd revised edition edition, 1991.
M. Härefors. Application of hª robust control to the RM12 jet engine. Control
Engineering Practice, 5(9):1189 – 1201, 1997. ISSN 0967-0661. doi:10.1016/S0967-
0661(97)84358-4. URL http://www.sciencedirect.com/science/article/
pii/S0967066197843584.
R. Hermann and A. Krener. Nonlinear controllability and observability. Auto-
matic Control, IEEE Transactions on, 22(5):728 – 740, oct 1977. ISSN 0018-9286.
doi:10.1109/TAC.1977.1101601.
John B. Heywood. Internal Combustion Engine Fundamentals. McGraw-Hill series in
mechanical engineering. McGraw-Hill, 1988. ISBN 0-07-100499-8.
J. H. Horlock. Advanced Gas Turbine Cycles. Pergamon, revised edition edition, 2007.
A Idebrant and Lennart Näs. Gas Turbine Applications using Thermofluid. In Peter
Fritzson, editor, Proceedings of the 3rd International Modelica Conference, pages 359–
366. The Modelica Association, May 2003.
Thomas Kailath. Linear Systems. Prentice-Hall, Inc, Englewood Cliffs, New Jersey
07632, 1980.
182 Chapter 10. Conclusion

Thomas Kailath, Ali H. Sayed, and Babak Hassibi. Linear Estimation. Prentice-Hall,
Inc, Upper Saddle River, New Jersey 07458, 2 edition, 2000.

Takahisa Kobayashi and Donald L. Simon. Application of a Bank of Kalman Filters for
Aircraft Engine Fault Diagnostics. NASA/TM-2003-212526, 2003. URL http://ntrs.
nasa.gov/search.jsp?print=yes&R=20030067984.

Takahisa Kobayashi and Donald L. Simon. Evaluation of an enhanced bank of kalman

filters for in-flight aircraft engine sensor fault diagnostics. Journal of Engineering for
Gas Turbines and Power, 127(3):497–504, 2005. doi:10.1115/1.1850505. URL http://
link.aip.org/link/?GTP/127/497/1.

Takahisa Kobayashi, Donald L. Simon, and Jonathan S. Litt. Application of a Constant

Gain Extended Kalman Filter for In-Flight Estimation of Aircraft Engine Performance
Parameters. NASA/TM-2005-213865, 2005. URL http://ntrs.nasa.gov/search.
jsp?print=yes&R=20050216398.

Milton G. Kofskey and William J. Nusbaum. Performance evaluation of a two-stage

axial-flow turbine for two values of tip clearance. Technical Report TN D-4388, National
Aeronautics and Space Administration (NASA), 1968.

Mattias Krysander and Erik Frisk. Sensor placement for fault diagnosis. IEEE Transac-
tions on Systems, Man, and Cybernetics – Part A: Systems and Humans, 38(6):1398–1410,
2008.

Mattias Krysander, Jan Åslund, and Mattias Nyberg. An efficient algorithm for finding
minimal over-constrained sub-systems for model-based diagnosis. IEEE Transactions
on Systems, Man, and Cybernetics – Part A: Systems and Humans, 38(1), 2008.

R. Kurz and K. Brun. Degradation in gas turbine systems. Journal of Engineering

for Gas Turbines and Power, 123(1):70–77, 2001. doi:10.1115/1.1340629. URL http:
//link.aip.org/link/?GTP/123/70/1.

Rainer Kurz, Klaus Brun, and Meron Wollie. Degradation effects on industrial gas
turbines. Journal of Engineering for Gas Turbines and Power, 131(6):062401, 2009.
doi:10.1115/1.3097135. URL http://link.aip.org/link/?GTP/131/062401/1.

Emil Larsson. Diagnosis and Supervision of Industrial Gas Turbines. Technical report,
Department of Electrical Engineering, 2012. LiU-TEK-LIC-2012:13, Thesis No. 1528,
http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-75985.

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Fault isolation for an industrial
gas turbine with a model-based diagnosis approach. ASME Conference Proceedings,
2010(43987):89–98, 2010. doi:10.1115/GT2010-22511.

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Health Monitoring in an
Industrial Gas Turbine Application by Using Model Based Diagnosis Techniques.
ASME Conference Proceedings, 2011(54631):487–495, 2011. doi:10.1115/GT2011-46825.
References 183

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Fault Tolerant Supervi-
sion of an Industrial Gas Turbine. ASME Conference Proceedings, 2013(55188), 2013.
doi:10.1115/GT2013-95727.

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Gas turbine modeling for
diagnosis and control. Journal of Engineering for Gas Turbines and Power, 136(7):17
pages, July 2014a. doi:10.1115/1.4026598.

Emil Larsson, Jan Åslund, Erik Frisk, and Lars Eriksson. Diagnosis and fault tolerant
supervision of industrial gas turbines. Submitted for journal publication, 2014b.

Jonathan S. Litt. An optimal orthogonal decomposition method for kalman filter-based

turbofan engine thrust estimation. Journal of Engineering for Gas Turbines and Power,
130(1), 2007. doi:10.1115/1.2747254.

R H Luppold, J R Roman, G W Gallops, and L J Kerr. Estimating in-flight engine

performance variations using kalman filter concepts. In 25th Joint Propulsion Conference,
Monterey, USA, 1989. 2584-AIAA.

R. Martínez-Guerra, R. Garrido, and A. Osorio-Miron. The fault detection problem

in nonlinear systems using residual generators. IMA Journal of Mathematical Control
and Information, 22(2):119–136, 2005. doi:10.1093/imamci/dni010.

K. Mathioudakis and T. Tsalavoutas. Uncertainty reduction in gas turbine performance

diagnostics by accounting for humidity effects. Journal of Engineering for Gas Turbines
and Power, 124(4):801–808, 2002. doi:10.1115/1.1470485.

S. Mattsson and G. Söderlind. Index reduction in differential-algebraic equations

using dummy derivatives. SIAM Journal on Scientific Computing, 14(3):677–692, 1993.
doi:10.1137/0914043. URL http://epubs.siam.org/doi/abs/10.1137/0914043.

Bonnie J. McBride and Sanford Gordon. Computer program for calculating and
fitting thermodynamic functions. Technical Report RP-1271, National Aeronautics
and Space Administration (NASA), 1992. URL http://www.grc.nasa.gov/WWW/
CEAWeb/RP-1271.htm.

Bonnie J. McBride, Michael J. Zehe, and Sanford Gordon. Nasa glenn coefficients
for calculating thermodynamic properties of individual species. Technical Report
TP-2002-211556, National Aeronautics and Space Administration (NASA), 2002. URL
http://www.grc.nasa.gov/WWW/CEAWeb/TP-2002-21556.htm.

C.B. Meher-Homji. Compressor and hot section fouling in gas turbines – causes and
effects. In Industrial Energy Conference, Houstion, TX, 1987.

Walter C. Merrill, John C. Delaat, and William M. Bruton. Advanced detection, isolation,
and accommodation of sensor failures - real-time evaluation. Journal of Guidance,
Control, and Dynamics, 11(6):517–526, 1988. doi:10.2514/3.20348.
184 Chapter 10. Conclusion

Modelica Association. The Modelica Language Specification version 3.0, 2007. URL
http://www.modelica.org/.

E. Naderi, N. Meskin, and K. Khorasani. Nonlinear fault diagnosis of jet engines by

using a multiple model-based approach. Journal of Engineering for Gas Turbines and
Power, 134(1), 2011. doi:10.1115/1.4004152.

H. Nijmeijer and T.I. Fossen. New directions in nonlinear observer design. Lecture notes
in control and information sciences. Springer, 1999. ISBN 9781852331344.

Ramine Nikoukhah. Innovations generation in the presence of unknown inputs: Appli-

cation to robust failure detection. Automatica, 30(12):1851 – 1867, 1994. ISSN 0005-1098.
doi:10.1016/0005-1098(94)90047-7.

Per Öberg. A DAE Formulation for Multi-Zone Thermodynamic Models and its Ap-
plication to CVCP Engines. PhD thesis, Linköpings universitet, 2009. URL http:
//www.fs.isy.liu.se/Publications/PhD/09_PhD_1257_PO.pdf.

E.S. Page. Continuous inspection schemes. Biometrika, 41:100–115, 1954.

Constantinos C. Pantelides. The consistent initialization of differential-algebraic systems.

SIAM J. Sci. Stat. Comput., 9(2):213–231, 1988.

R.J. Patton and M. Hou. Design of fault detection and isolation observers: A matrix pen-
cil approach. Automatica, 34(9):1135 – 1140, 1998. ISSN 0005-1098. doi:10.1016/S0005-
1098(98)00043-0.

Randal T. Rausch, Kai F. Goebel, Neil H. Eklund, and Brent J. Brunell. Integrated in-
flight fault detection and accommodation: A model-based study. Journal of Engineering
for Gas Turbines and Power, 129(4):962–969, 2007. doi:10.1115/1.2720517. URL http:
//link.aip.org/link/?GTP/129/962/1.

M. Safonov and M. Athans. Robustness and computational aspects of nonlinear stochas-

tic estimators and regulators. Automatic Control, IEEE Transactions on, 23(4):717 – 725,
aug 1978. ISSN 0018-9286. doi:10.1109/TAC.1978.1101825.

H.I.H. Saravanamuttoo, G.F.C. Rogers, and H. Cohen. Gas Turbine Theory. Longman,
5th edition edition, 2001. ISBN 0-13-015847-X.

J.N. Scott. Reducing turbo machinery operation costs with regular performance testing.
In International Gas Turbine Conference and Exhibit, Dusseldorf, Germany, 1986.

R. Shields and J. Pearson. Structural controllability of multiinput linear systems. Au-

tomatic Control, IEEE Transactions on, 21(2):203 – 212, apr 1976. ISSN 0018-9286.
doi:10.1109/TAC.1976.1101198.

S. Simani and Ron J. Patton. Fault diagnosis of an industrial gas turbine prototype using
a system identification approach. Control Engineering Practice, 16(7):769 – 786, 2008.
ISSN 0967-0661. doi:http://dx.doi.org/10.1016/j.conengprac.2007.08.009.
References 185

Dan Simon and Donald L. Simon. Aircraft turbofan engine health estimation using
constrained kalman filtering. Journal of Engineering for Gas Turbines and Power, 127
(2):323–328, 2005. doi:10.1115/1.1789153.

Donald L. Simon and Jeffrey B. Armstrong. An integrated approach for aircraft engine
performance estimation and fault diagnostics. Journal of Engineering for Gas Turbines
and Power, 135(7), 2013. doi:10.1115/1.4023902.

Donald L. Simon, Jeff Bird, Craig Davison, Al Volponi, and R. Eugene Iverson.
Benchmarking gas path diagnostic methods: A public approach. ASME Confer-
ence Proceedings, 2008(43123):325–336, 2008. doi:10.1115/GT2008-51360. URL http:
//link.aip.org/link/abstract/ASMECP/v2008/i43123/p325/s1.

Gregory N. Stephanopoulos, Aristos A. Aristidou, and Jens Nielsen. Metabolic Engineer-

ing: Principles and Methodologies. Academic Press, 1 edition, 1998. ISBN 0126662606.
URL http://www.worldcat.org/isbn/0126662606.

N. Sugiyama. System identification of jet engines. Journal of Engineering for Gas

Turbines and Power, 122(1):19–26, 2000. doi:10.1115/1.483172.

Carl Svärd and Mattias Nyberg. Observer-based residual generation for linear
differential-algebraic equation systems. In IFAC World Congress, Seoul, Korea, 2008.

Michael Tiller. Introduction to Physical Modeling with Modelica. Kluwer Academic

Publishers, Norwell, MA, USA, 2001. ISBN 0792373677.

Stephen R. Turns. Thermodynamics : concepts and applications. Cambridge University

Press, Cambridge, 2006. ISBN 978-052-185-242-1.

L. A. Urban. Gas Turbine Engine Parameter Interrelationships. Windsor Locks, 1969.

L. A. Urban. Gas path analysis applied to turbine engine condition monitoring. Journal
of Aircraft, 10(7):400–406, 1972.

L. A. Urban and A. J. Volponi. Mathematical methods of relative engine performance

diagnostics. Journal of Aerospace, 1992. doi:10.4271/922048.

A. J. Volponi. Sensor error compensation in engine performance diagnostics. ASME

Conference Proceedings, 1994(58), 1994.

A. J. Volponi. Gas turbine parameter corrections. Journal of Engineering for Gas

Turbines and Power, 121(4):613–621, 1999. doi:10.1115/1.2818516.

A. J. Volponi. Foundation of Gas Path Analysis (Part I and II). von Karman Institute
Lecture Series: Gas Turbine Condition Monitoring and Fault Diagnosis, January 2003.

A. J. Volponi. Gas turbine engine health management: Past, present, and future trends.
Journal of Engineering for Gas Turbines and Power, 136(5), 2014. doi:10.1115/1.4026126.
186 Chapter 10. Conclusion

A. J. Volponi, H. DePold, R. Ganguli, and C. Daguang. The use of kalman filter and
neural network methodologies in gas turbine performance diagnostics: A compara-
tive study. Journal of Engineering for Gas Turbines and Power, 125(4):917–924, 2003.
doi:10.1115/1.1419016.
J. W. Watts, T. E. Dwan, and C. G. Brockus. Optimal state-space control of a gas
turbine engine. Journal of Engineering for Gas Turbines and Power, 114(4):763–767, 1992.
doi:10.1115/1.2906654. URL http://link.aip.org/link/?GTP/114/763/1.
 Appendix A

Mole/Mass Conversions

A.1 Mole/Mass Fraction Calculation

Mole, and mass fraction conversion appears in different parts of the thesis. Here will
these calculations be summarized.

Mass and Mole Concentrations

The mole concentration x̃ i of species i is defined
ni
x̃ i (A.1)
n
where n i is the number of mole of species i and n is the total number of mole in the
mixture. The mass concentration x i of species i is defined
mi
xi (A.2)
m
where m i is the mass of species i and m is the total mass in the mixture.

Mole Mass
The mole mass of a mixture with mole concentration x̃ > Rn , is defined as the sum of the
elements
M Q
M i x̃ i (A.3)
i
where M i is the mole mass for species i, and the sum over all species is Pi x̃ i 1. The
mole mass expressed in mass concentration x > Rn can be written

M Q M i P 1~xMj ~M
i
xi
Pi x i 1
(A.4)
i j j P x
j j ~ M j P j j ~M j
x
where (A.6) is used.

187
188 Appendix A. Mole/Mass Conversions

Mole fraction Mass fraction

The conversion from mole fraction x̃ i to mass fraction x i is

mi Mi ni Mi ni 1~n Mi Mi
xi x̃ i x̃ i (A.5)
m Pj Mjnj P j M j n j 1~n P j M j x̃ j M

where m i is the mass of species i, n i is the number of mole of species i, m is the total
mass, and n is the total number of mole. In the last step, (A.3) is used.

Mass fraction Mole fraction

The conversion from mass fraction x i to mole fraction x̃ i is

ni m i ~M i 1~m 1~ M i
x̃ i xi (A.6)
n P j m j ~ M j 1~m P j x j ~M j

where n i is the number of moles of species i. Using (A.4), the mole fraction can be
written
1/M i
x̃ i xi (A.7)
1/M

A.2 Stoichiometry Matrix Expressed in Mass

The stoichiometry matrix, expressed in mole, repeated from (4.32) is:

’0 0 0 0 0“
– 1 2 3 1 0—
– —
S̃ –2 3 4 0 0—
– —
–0 0 0 0 1—
”2 3.5 5 0 0•

where the rows represent the species: Ar, CO 2 , H 2 O, N 2 , and O 2 according to the air
vector x a in (4.30). The columns represent the species: CH 4 , C 2 H 6 , C 3 H 8 , CO 2 , and N 2
according to the fuel vector x f in (4.30). Expression (A.5) can now be utilized for each
matrix element to get:

’ 0 0 0 0 0“
MC O2 M M
–
– M C H4
2 M CC OH2 3 M CC OH2 1 0—
—
2 6 3 8
– MH O M M —
S –2 2
3 M CH 2HO 4 M CH 2HO 0 0— (A.8)
– M C H4 —
– 2 6 3 8 —
–
–
0 0 0 0 1—
—
MO M O2 M O2
”2 M 2 3.5
MC H
5
MC H
0 0•
C H4 2 6 3 8

where the stoichiometric matrix now is expressed in masses instead of moles.

A.3. Determination of Stoichiometric Air/Fuel Ratio 189

A.3 Determination of Stoichiometric Air/Fuel Ratio

Reaction formula (4.34) represents a combustion, and is expressed in masses:

ma xa  m f x f m a x a  m f Sx f (A.9)

where S is the stoichiometric matrix from (A.8). The stoichiometric air/fuel ratio appears
when it is just enough oxygen in the reaction, i.e., the last row in (A.9) that represents
oxygen is equal to zero and can be solved. The last line can be written:

MO2 MO2 MO2

m a x a,O 2 m f Š2 x f ,C H 4  3.5 x f ,C 2 H 6  5 x f ,C 3 H 8 
MC H4 MC2 H6 MC3 H8

which can be rewritten in form:

x x x
ma 2 Mf ,CC HH 4  3.5 Mf ,CC 2HH 6  5 Mf ,CC 3HH 8 ma
4 2 6
x a,O 2
3 8

Š  (A.10)
mf M O2
mf s

which is the definition of ˆA~F s .

 Appendix B

Measurement Plots

In this appendix, additional experimental data plots are viewed. These plots are the
ambient temperature T0 , the ambient pressure p 0 , the shaft speed of the gas generator
n GG and the generated power by the external application P.

191
192 Appendix B. Measurement Plots

B.1 Ambient Temperature T0

t
Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[K]

K
Nov
310

300

290

280

310

300

290

280

(a) T0 for the first six-month period. (b) T0 for the second six-month period.
Figure B.1: In the figure, the ambient temperature T0 is shown. The mean of the temperature
varies according to winter and summer half year. In the end of February, the temperature
rises for a couple of days, which affect the calculated measurement temperature deltas in, e.g.,
Figure 7.5.
B.2. Ambient pressure p 0 193

B.2 Ambient pressure p0

t
Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[kPa]

[kPa]
Nov
104

103

102

101

104

103

102

101

(a) p 0 for the first six-month period. (b) p 0 for the second six-month period.
Figure B.2: Ambient pressure p 0 for one year of experimental data.
194 Appendix B. Measurement Plots

B.3 Shaft Speed nC1 of the Gas Generator

t
Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[rpm]

[rpm]
Nov
9600

9400

9200

9000

9600

9400

9200

9000

(a) n C1 for the first six-month period. (b) n C1 for the second six-month period.

Figure B.3: Shaft speed nGG of the gas generator for one year of experimental data.
B.4. Generated Power PA by the Application 195

B.4 Generated Power PA by the Application

t
Oct
Apr

Sep
Mar

Aug
Feb

Jul
Jan

Jun
Dec

May
[MW]

[MW]
Nov
25

20

15

25

20

15

(a) PA for the first six-month period. (b) PA for the second six-month period.

Figure B.4: Generated power PA by the application for one year of experimental data.
196 Appendix B. Measurement Plots
Notes 197
198 Notes
 Linköping Studies in Science and Technology, Dissertations
Division of Vehicular Systems
Department of Electrical Engineering
Linköping University

No 1 Magnus Pettersson, Driveline Modeling and Control, 1997.

No 2 Lars Eriksson, Spark Advance Modeling and Control, 1999.

No 3 Mattias Nyberg, Model Based Fault Diagnosis: Methods, Theory, and Automotive
Engine Applications, 1999.

No 4 Erik Frisk, Residual Generation for Fault Diagnosis, 2001.

No 5 Per Andersson, Air Charge Estimation in Turbocharged Spark Ignition Engines,

2005.

No 6 Mattias Krysander, Design and Analysis of Diagnosis Systems Using Structural

Methods, 2006.

No 7 Jonas Biteus, Fault Isolation in Distributed Embedded Systems, 2007.

No 8 Ylva Nilsson, Modelling for Fuel Optimal Control of a Variable Compression En-
gine, 2007.

No 9 Markus Klein, Single-Zone Cylinder Pressure Modeling and Estimation for Heat
Release Analysis of SI Engines, 2007.

No 10 Anders Fröberg, Efficient Simulation and Optimal Control for Vehicle Propulsion,
2008.

No 11 Per Öberg, A DAE Formulation for Multi-Zone Thermodynamic Models and its
Application to CVCP Engines, 2009.

No 12 Johan Wahlström, Control of EGR and VGT for Emission Control and Pumping
Work Minimization in Diesel Engines, 2009.

No 13 Anna Pernestål, Probabilistic Fault Diagnosis with Automotive Applications,

2009.

No 14 Erik Hellström, Look-ahead Control of Heavy Vehicles, 2010.

No 15 Erik Höckerdal, Model Error Compensation in ODE and DAE Estimators with
Automotive Engine Applications, 2011.
No 16 Carl Svärd, Methods for Automated Design of Fault Detection and Isolation Sys-
tems with Automotive Applications, 2012.

No 17 Oskar Leufvén, Modeling for control of centrifugal compressors, 2013.

No 18 Christofer Sundström, Model Based Vehicle Level Diagnosis for Hybrid Electric
Vehicles, 2014.

No 19 Andreas Thomasson, Modeling and control of actuators and co-surge in tur-

bocharged engines, 2014.