University of Central Florida

STARS

Electronic Theses and Dissertations, 2020-

2021

Anomaly Detection and Failure Prediction in Gas Turbines

Vipul Goyal
University of Central Florida

Part of the Mechanical Engineering Commons

Find similar works at: https://stars.library.ucf.edu/etd2020
University of Central Florida Libraries http://library.ucf.edu

This Doctoral Dissertation (Open Access) is brought to you for free and open access by STARS. It has been accepted
for inclusion in Electronic Theses and Dissertations, 2020- by an authorized administrator of STARS. For more
information, please contact [email protected].

STARS Citation
Goyal, Vipul, "Anomaly Detection and Failure Prediction in Gas Turbines" (2021). Electronic Theses and
Dissertations, 2020-. 872.
https://stars.library.ucf.edu/etd2020/872
ANOMALY DETECTION AND FAILURE PREDICTION IN GAS TURBINES

by

VIPUL GOYAL
B.Tech. Aerospace Engineering, Indian Institute of Technology Bombay, 2017
M.Tech. Aerospace Engineering, Indian Institute of Technology Bombay, 2017

A dissertation submitted in partial fulfilment of the requirements

for the degree of Doctor of Philosophy
in the Department of Mechanical and Aerospace Engineering
in the College of Engineering and Computer Science
at the University of Central Florida
Orlando, Florida

Fall Term
2021

Major Professor: Jayanta S. Kapat

© 2021 Vipul Goyal

ii
 ABSTRACT

This study is based on time-series data taken from the combined cycle heavy-duty utility
gas turbines. For analysis, first, a multi-stage vector autoregressive model is constructed for the
nominal operation of powerplant assuming sparsity in the association among variables, and this
model is used as a basis for anomaly detection and prediction. This prediction is compared with
the time-series data of the powerplant test data containing anomalies. Granger causality networks,
which are based on the associations between the time series streams, can be learned as an important
implication from the vector autoregressive modelling. This method suffers from the disadvantage
that some of the variables are not stationary even after segmenting the working mode based on the
RPM. To improve the efficacy of the algorithm, the observations are further clustered into differ-
ent working modes, because of the heterogeneous behavior of the gas turbine parameters under
various modes. Then predicting the operational parameters is considered under each mode respec-
tively, via algorithms including random forest, generalized additive model, and neural networks.
The comparative advantage based on prediction accuracy and applicability of the algorithms is
discussed for real-time use and post processing. The advantage of this segmentation method is
that it achieves high predictive power and provides insight into the behavior of specific gas turbine
variables. Next, the long-memory behavior of residuals is modeled, and heterogeneous variances
are observed from the residuals of the generalized additive model. Autoregressive Fractionally In-
tegrated Moving Average (ARFIMA) and Generalized Autoregressive Conditional Heteroskedas-
ticity (GARCH) models are employed to fit the residual process, which significantly improve the
prediction. Rolling one-step-ahead forecast is studied. Numerical experiments of abrupt changes
and trend in the blade-path temperature are performed to evaluate the specificity and sensitivity
of the prediction. The prediction is sensitive given reasonable signal-to-noise ratio and has lower
false-positive rate.

iii
ACKNOWLEDGMENTS

iv
 TABLE OF CONTENTS

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

NOMENCLATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

CHAPTER 1: INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Role of Gas Turbines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Gas Turbine Maintenance Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Anomaly Detection in Gas Turbines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

CHAPTER 2: LITERATURE REVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Gas Turbine Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Condition-based Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Data-driven Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

CHAPTER 3: OBJECTIVE AND NOVELTY . . . . . . . . . . . . . . . . . . . . . . . . 14

CHAPTER 4: METHODOLOGY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Exploratory Data Analysis (EDA) and Data Selection . . . . . . . . . . . . . . . . . . . 16
Time-series Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
System and Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Performance Evaluation Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Data-based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Vector Auto-regressive Method (VAR) . . . . . . . . . . . . . . . . . . . . . . . . 21
Hidden Markov Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

v
 Supervised Learning Algorithms after Segmentation . . . . . . . . . . . . 27
Additive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Random Forests . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Additive Model with ARFIMA-GARCH Residuals . . . . . . . . . . . . . . . . . 31

CHAPTER 5: RESULTS AND DISCUSSION . . . . . . . . . . . . . . . . . . . . . . . . 38

Outlier Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Vector Autoregressive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Anomaly Detection from the VAR Model . . . . . . . . . . . . . . . . . . . . . . 43
Generalized Additive Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Supervised Prediction using Hidden Markov Model . . . . . . . . . . . . . . . . . . . . 47
GAM-ARFIMA-GARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Performance on Real Gas Turbine Operating Data . . . . . . . . . . . . . . . . . . 56
Performance on Gas Turbine Data with Simulated Perturbation . . . . . . . . . . . 61
Control Charts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Limitations of Data Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

CHAPTER 6: CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Error Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

LIST OF REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

vi
 LIST OF FIGURES

Figure 1.1: World net electricity generation by source (trillion kilowatthours) [1] . . . 2
Figure 1.2: An example of unexpected change in cooling steam pressure. . . . . . . . . 5

Figure 4.1: Schematic of one powerplant unit . . . . . . . . . . . . . . . . . . . . . . 17

Figure 4.2: RPM data used for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 4.3: Sample input and output data for gas turbine . . . . . . . . . . . . . . . . 19
Figure 4.4: Anomaly Detection Flowchart . . . . . . . . . . . . . . . . . . . . . . . . 22
Figure 4.5: Time history of normalized power generation . . . . . . . . . . . . . . . . 23
Figure 4.6: Illustration of a neural network . . . . . . . . . . . . . . . . . . . . . . . . 30
Figure 4.7: Training data non-dimensional temperature vs fuel flow . . . . . . . . . . 33
Figure 4.8: Training data scatterplot between Xt and Yt . . . . . . . . . . . . . . . . . . 34
Figure 4.9: Time-series plot of the residuals ξt of the GAM model . . . . . . . . . . . 35
Figure 4.10: Autocorrelation function (ACF) plot of the residuals ξt of the GAM model 36

Figure 5.1: Outlier detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

Figure 5.2: Non-dimensional one-step-ahead prediction of GT power . . . . . . . . . . 39
Figure 5.3: Non-dimensional error for one-step-ahead prediction of GT power . . . . . 39
Figure 5.4: Non-dimensional one-step-ahead prediction of blade-path temperature . . . 40
Figure 5.5: Non-dimensional error for one-step-ahead prediction of blade-path tem-
perature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Figure 5.6: Non-dimensional one-step-ahead prediction of blade-path temperature spread 41
Figure 5.7: Non-dimensional error for one-step-ahead prediction of blade-path tem-
perature spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Figure 5.8: Predicted vs observed GT power . . . . . . . . . . . . . . . . . . . . . . . 42
Figure 5.9: Predicted vs observed blade-path temperature . . . . . . . . . . . . . . . . 42

vii
Figure 5.10: Predicted vs observed blade-path temperature spread . . . . . . . . . . . . 43
Figure 5.11: Plots of the relationship between inputs and the outputs for gas turbine
power. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Figure 5.12: Plots of the relationship between inputs and the outputs for turbine tem-
perature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Figure 5.13: Plots of the relationship between inputs and the outputs for temperature
spread. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Figure 5.14: Predictions of the test set for gas turbine power, turbine temperature and
temperature spread, respectively. . . . . . . . . . . . . . . . . . . . . . . 48
Figure 5.15: The absolute deviation of the predicted values from the observed values for
gas turbine power, turbine temperature and temperature spread, respectively. 49
Figure 5.16: States estimated from the hidden Markov model for power and blade path tem-
perature spread. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Figure 5.17: Prediction of GT power using GAM. . . . . . . . . . . . . . . . . . . . . 50
Figure 5.18: Prediction of GT power using random forest. . . . . . . . . . . . . . . . . 51
Figure 5.19: Prediction of blade path temperature spread using neural networks. . . . . 51
Figure 5.20: Prediction of blade path temperature spread in State 4 (the “on” state). . . 52
Figure 5.21: Combined Prediction of GT power and blade path temperature spread. . . . 53
Figure 5.22: The nonlinear function relationship between the fuel flow and blade-path
temperature for the testing set. . . . . . . . . . . . . . . . . . . . . . . . . 57
Figure 5.23: Prediction of blade-path temperature using GAM . . . . . . . . . . . . . . 58
Figure 5.24: Prediction of blade-path temperature using GAM-ARFIMA-GARCH . . . 59
Figure 5.25: Standardized prediction residuals from the GAM and GAM-ARFIMA-
GARCH models in the testing set. . . . . . . . . . . . . . . . . . . . . . . 60
Figure 5.26: Squared residuals and predication variances from GAM . . . . . . . . . . 61
Figure 5.27: Squared residuals and predication variances from GAM-ARFIMA-GARCH 62

viii
Figure 5.28: Simulated perturbation: step-jump in the temperature. . . . . . . . . . . . 63
Figure 5.29: Simulated perturbation: increasing linear trend in temperature. . . . . . . 64
Figure 5.30: Prediction comparison for 5% step-jump simulation . . . . . . . . . . . . . 65
Figure 5.31: Prediction comparison for 0.1% per time-step ramp simulation . . . . . . . 66
Figure 5.32: Enlarged prediction comparison for 5% step-jump simulation . . . . . . . 67
Figure 5.33: Enlarged prediction comparison for 0.1% per time-step ramp simulation . . 67
Figure 5.34: Prediction comparison for 0.02% per time-step ramp simulation . . . . . . 68
Figure 5.35: Prediction comparison for 0.2% per time-step ramp simulation . . . . . . . 68
Figure 5.36: Standardized temperature residuals: abrupt jump sizes of 1%, 5% and 10%. 69
Figure 5.37: Standardized temperature residuals: linear trend with rates of 0.02%, 0.1%
and 0.2% per time-step. . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Figure 5.38: Quantile-quantile plot for the residuals . . . . . . . . . . . . . . . . . . . . 71
Figure 5.39: EWMA control chart for 1% manually introduced step jump. . . . . . . . . 72
Figure 5.40: Gas turbine performance - exhaust gas temperature . . . . . . . . . . . . . 73
Figure 5.41: Gas turbine performance - combustion pressure . . . . . . . . . . . . . . . 74
Figure 5.42: Power generation failure . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Figure 6.1: Components of Digital Twin with System Genome . . . . . . . . . . . . . 78

ix
 LIST OF TABLES

Table 1.1: Average forced outage % in F-class combined cycle subsystems [2] . . . . 4
Table 1.2: Gas turbine power generation and maintenance cost . . . . . . . . . . . . . 4

Table 2.1: Summary of literature on gas turbine diagnostics . . . . . . . . . . . . . . 8

Table 2.2: Summary of GPA and CBM methods . . . . . . . . . . . . . . . . . . . . 11
Table 2.3: State-of-the-art PHM methods . . . . . . . . . . . . . . . . . . . . . . . . 12

Table 5.1: Scaled RMSE of GT power generation . . . . . . . . . . . . . . . . . . . 52

Table 5.2: Scaled RMSE of turbine temperature spread . . . . . . . . . . . . . . . . . 53
Table 5.3: Comparison of performance of power prediction of the four predictors. . . 54
Table 5.4: RMSPE from the GAM and GAM-ARFIMA-GARCH predictions respec-
tively for various days of the testing set. . . . . . . . . . . . . . . . . . . . 59

Table 6.1: RMSE comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

x
 NOMENCLATURE

Superscripts and Subscripts

ACF Autocorrelation Function

ARCH Autoregressive Conditional Heteroskedasticity

ARF IM A Autoregressive Fractionally Integrated Moving Average

GAM Generalized Additive Model

GARCH Generalized Autoregressive Conditional Heteroskedasticity

GCN Granger Causality Network

GT Gas Turbine

HM M Hidden Markov Model

HRSG Heat Recovery Steam Generator

LASSO Least Absolute Shrinkage and Selection Operator

M LP Multi-Layer Perceptron

RM SE Root Mean Square Error

RM SP E Root Mean Squared Percentage Error

RP M Rotations per Minute

ST Steam Turbine

V AR Vector Autoregression

xi
 CHAPTER 1: INTRODUCTION

Role of Gas Turbines

Gas turbines are used to generate power by converting thermal energy into mechanical en-
ergy. Their main application lies in the field of aviation and electricity generation. These engines
are used in aviation because of their higher power-to-weight ratio and higher mass flow rates com-
pared to other available alternatives. On the other hand, in electricity generation, these engines
are used because of faster ramp response and their suitability in the topping cycle in the combined
cycle power plants.

Gas Turbine Maintenance Cost

Gas turbines are some of the most expensive and vital components of the aircrafts and elec-
tricity generation plants. Moreover, these machines work under extreme thermal and mechanical
loading conditions and thus these machines need regular maintenance to deliver optimum per-
formance. Any forced outage imposes a hefty cost on the operator due to stranded capital during
downtime and the additional maintenance cost itself. Zhao [3] has reported that from 2006 to 2012,
in Chinese aviation sector 83% of problems associated with flights were related to the engines and
IATA (International Air Transport Association) has reported that in 2017 aircraft maintenance cost
reached $ 76 Billions and 42% of the maintenance cost is spent on the engines [4]. These numbers
are from different sources, so no exact conclusion can be drawn but these numbers are compara-
ble to other sources [5] and these do highlight the magnitude of the challenge and the associated
opportunity for the cost savings.
In the electricity generation sector, US electricity generation was a little over 4000 TWh
(Terawatt Hour) in 2019 and it is expected to grow at a rate of a little over 1% in EIA (Energy
Information Administration) reference-case scenario [6]. Out of the 5400 TWh electricity expected

1
 Figure 1.1: World net electricity generation by source (trillion kilowatthours) [1]

to be produced in 2050 [1], approximately 2000 TWh will be produced using solar PV (photo-
voltaic) and wind energy. Due to inherent intermittency of solar-wind energy and due to lack
of cheap energy storage system, natural gas is expected to play a huge role as a peaking power
provider due to its fast ramp-up and ramp-down capability. According to EIA [1], natural gas will
provide approximately 36% of electricity even in 2050.
On the other hand, global electricity generation is expected to double by 2050 with the ref-
erence growth rate of approximately 2.1%. And as nations grow richer, the demand for electricity
will also be buoyed by the fact that it is a cleaner form of energy compared to other form of en-
ergy (i.e. less carbon reaching to the end user), as outlined by BP (British Petroleum plc) [7] that
only 10% of world population remains without electricity which means a drop of 33% in last five
years. According to EIA [1], as shown in Figure 1.1, natural gas currently has 25% market share
in electricity generation and it will still be 16% in 2050 when the overall electricity generation is
expected to rise at least by a factor of 2. So, lowering of maintenance cost will have a huge impact
for decades to come.

2
 It is estimated that for heavy duty gas turbine the maintenance is performed after 8000-
48000 operating hours [8]. This wide band of inspection interval is due to the change in main-
tenance requirement depending upon whether the plant is a cycling or a base-load power plant.
According to NREL (National Renewable Energy Laboratory) [9], capital & maintenance cost for
a combined cycle power plant is 30-60 $/MWh-capacity. For a typical combined cycle of the con-
figuration of 2x2x1 (2 gas turbines, 2 heat recovery steam generators feeding steam to 1 steam
turbine) the variable O&M (operation and maintenance) cost is estimated to be 1.87 $/MWh in
2019 dollars [10]. For F-class, 240 MW turbine, the consumables cost (one part of O&M cost)
according to this report is is 0.60 $/MWh and the major maintenance cost is 18500 $/Start. Ac-
cording to some other EIA reports [2] the variable O&M cost are nearly double this amount at 3.25
$/MWh, so estimated annual variable O&M cost for a 250 MW plant is of the order of millions of
dollars if it is assumed that maintenance is performed once in three years with 2400 total number
of GT Starts.
So, significant cost savings can be achieved if this maintenance cycle can be optimized and
thus the focus of industry has been on the condition-based maintenance rather than fixed mainte-
nance schedule for last few decades which are extremely conservative in nature. Gülen [2] has also
provided the forced outage rates for gas turbine, steam turbine and HRSG system over the span of
more than a decade which are given in the table below. It clearly shows that gas turbine causes
higher forced outages compared to HRSG (Heat Recovery Steam Generator) and steam turbine
but these forced outages due to gas turbines are steadily coming down. The numbers for the gas
turbine in the table may include fault originated in accessories.

Anomaly Detection in Gas Turbines

The idea of anomaly detection method can be broadly explained with the help of Figure
1.2. In this figure cumulative combined-cycle power generation is shown along with cooling steam

3
 Table 1.1: Average forced outage % in F-class combined cycle subsystems [2]

Forced outage % 1995-1999 2000-2004 2005-2009

Gas Turbine Subsystem 2.67 2.09 1.59
HRSG Subsystem 0.11 0.28 0.21
Steam Turbine Subsystem 0.36 0.39 0.81

Table 1.2: Gas turbine power generation and maintenance cost

Author/Organization Ref. Year Type Topic

NREL [9] 2012 Technical Report Power plant cycling cost
EIA [6] 2021 Technical Report US energy forecast
EIA [1] 2021 Technical Report International energy outlook
BP [7] 2021 Technical Report World energy review
IATA [4] 2021 Technical Report Aircraft maintenance cost
EIA [10] 2020 Technical Report Powerplant capital cost
Gülen [2] 2020 Book chapter Maintenance cost combined-cycle
Gülen [8] 2019 Book chapter Maintenance cost gas turbine

pressure in one particular gas turbine unit. After extensive data-analysis, if it is ascertained a
priori that cooling steam pressure and power generation are highly correlated statistically then the
unexpected increase in the cooling pressure at the time of powering down can be flagged as an
anomaly much before pressure reaches to extremely high levels as seen at top-right corner of the
graph. Thus corrective action can be initiated before this anomaly causes unit failure. Since this
anomaly is developing progressively rather than a sudden spike in pressure, this can potentially
give early warning and thus prevent any system failure.
Engine health monitoring has made vast improvements over the years and most of these
methods used in the health monitoring rely on the physics-based model of the machine. The
deviation from the predefined performance path is what signifies the onset of an anomaly/failure.
Though these methods have served well, there are additional gains to be achieved if additional
insight can be gained into the engine performance and fault initiation which may have less to do

4
 Figure 1.2: An example of unexpected change in cooling steam pressure.

with the physics but more with the run-time performance of the engine. For example, if there is
circumferential variation in the fuel nozzles’ shape in the combustion chamber then over time that
may cause circumferential flow distortion in the turbine downstream and this can have adverse
effect on the engine. Though this circumferential distortion may be caused by several factors but
if, with the help of data analysis, a relation can be established among thermodynamic variables and
the variables which signify circumferential distortion then that relation can be used to monitor the
run-time degradation in the gas turbine and any deviation from that path can be attributed to a new
onset of anomaly.

5
 CHAPTER 2: LITERATURE REVIEW

Gas Turbine Diagnostics

The gas turbine literature uses various different definitions for diagnosis and prognosis and
these are not uniformly defined. So, for clarity some of the meanings associated with the specific
terms are given below and these have been adopted from Bechini [11]. Here, I associate diagnosis
and prognosis with exact meanings of the words, i.e. diagnosis means that a fault has occurred and
it has been identified and prognosis means that a scenario has been established which can lead to
failure.

1. Fault detection: Confirmation of any deviation from normal performance of gas turbine

2. Fault isolation: Confirmation of the source of fault

3. Fault identification: Assessment of the damage done along with the magnitude of the prob-
lem the fault can lead to

Fault diagnosis/anomaly detection is merely a combination of both fault detection and fault
isolation here; though some researchers consider fault diagnosis to be an even larger category
encapsulating even more steps [3].
Within this entire field of diagnostics there are four major categories of maintenance schemes
which have been considered till now as described by Zhao [3] - Breakdown maintenance, planned
maintenance, condition-based maintenance (CBM), and prognostics and health management (PHM).
If we are following breakdown maintenance or planned maintenance for gas turbine that means we
are not concerned about diagnostics. We use fail-safe subsystems or devise a very conservative
maintenance schedule and thus incur a heavy cost because of under-utilized life left in the com-
ponents. These two maintenance schedules are no longer used in the industry. Condition-based
maintenance has been the mainstay of the industry in the form of gas-path analysis (explained later)

6
and its various modifications while the PHM is an ongoing research area where various methods
are still being tried and perfected. Researchers are posing faith in the PHM because of remarkable
success of data-based algorithm in various fields and due to the fact that condition-based mainte-
nance itself works accurately when the engine health parameters can be obtained from the available
sensor information reliably which is often not the case.
There is no exact dividing line between what is considered CBM and PHM but generally
PHM is considered to be a sophisticated form of CBM. In CBM, the algorithms inform about
the deviation of the engine performance from the expected performance and hence it can help
us in determining the maintenance schedule relatively flexibly. In CBM, we don’t detect any
anomaly directly; rather, the deviation itself is taken as proxy of the anomaly and the source of
anomaly is not the direct unknown to be investigated. On the other hand, PHM is concerned
more with predicting directly the remaining useful life (RUL) and hence predict anomaly before it
occurs. According to Lee [12], the suitability of the algorithm depends upon the complexity and
uncertainty involved in the system. One factor that is known with reasonable confidence is that
CBM is not good when there are a lot of uncertainties involved in the measurements itself or in the
system response. And it is also known that CBM alone hasn’t been accurate enough [3].
All the faults in gas turbines can be either attributed to hot gas path flow or to structural
components. The reason these two have been identified separately here is that, for the first type of
the faults, there are some measurements available and they also tend to be more gradual in nature
while for the second type, not a lot of measurements are available, so the failures are sudden from
the analysis standpoint. So, only hot gas path faults can be considered for analysis.
There are two important steps that can be taken when it comes to improving the perfor-
mance of gas turbine maintenance schedule. One is gas turbine diagnostics and the other is gas
turbine prognostics. The prognostics makes sense only when the failure is not sudden from the
viewpoint of thermal data also; for example, all the structural failures may not be predictable be-
cause the structural weakness may not impact the gas path data in a statistically significant way.

7
On the other hand gradual deterioration in the performance of the gas turbine might signal some
anomaly in the hot gas path of turbine and here we can look for algorithms which can diagnose
the problem. So, in this study also, the focus is on hot gas path components; though the data may
contain faults which were related to structural components.
We are also specifically interested in hot gas path anomaly detection because as reported
by Zhao [3], close to a staggering 90% failures in aviation are associated with hot gas path and
these failures take up close to 60% of total maintenance budget. So, if hot gas path anomalies can
be brought down then, again, substantial cost savings can be achieved.

Table 2.1: Summary of literature on gas turbine diagnostics

Author Ref. Year Type Topic

Li [13] 2002 Review paper Gas turbine anomaly detection progress
Marinai et. al. [14] 2004 Review paper Gas turbine anomaly detection progress
Loboda [15] 2010 Book chapter Gas turbine condition monitoring
Romessis et. al. [16] 2013 Review article Condition-based maintenance
Stamatis [17] 2013 Book chapter Engine condition monitoring
Volponi [18] 2014 Review paper Review of GPA-based fault detection
Zhao et. al. [3] 2016 Review paper Gas turbine anomaly detection progress
Tahan et. al. [19] 2017 Review paper Review of health monitoring
Fentaye et. al. [20] 2019 Review paper Review of gas-path diagnostics challenges

The gas path analysis has been the mainstay of the industry and the broad idea of this
method can be described as -

Y = f (x, u) (2.1)

where,
Y- measurements taken from the engine
x - parameters denoting the health of the engine
u - operating point of the engine

8
 In classical GPA where we assume that changes in engine measurements are small and thus
vary linearly with the health parameters, the function f is described as a matrix which relates health
parameters to the engine measurements. This linearity assumption breaks down during transient
operation and large deviations. Thus, for large deviations, linear formulation is not accurate and
the remedies have been proposed which incorporate the non-linearities. The review by Li [13]
showed that GPA method had been enhanced by using nonlinear gas turbine model and the ge-
netic algorithms have been used to reach to a globally optimized solution. This review showed
that neural networks had potential and were promising at the time but the best outcomes were
expected from fuzzy logic and ES-based (expert systems) methods which were being developed
at the time. Though, none of these algorithms had incorporated transients. The sensor noise was
being remedied using data averaging which is now being reduced using neural networks also [20].
The review by Marinai et. al. [5] was along the similar line which showed that GPA had been
improved in research using nonlinear gas turbine model and AI-based methods, specifically expert
systems were showing promising results. The neural network training requirements were reported
to be extensive but these methods did not suffer from ’smearing’ problem where a fault is traced to
multiple components at the same time.
The review by Loboda [15] summarizes the model-based method (GPA and dynamic GT
model) and illustrates how data-driven methods, specifically neural network, can be used to detect
anomalies in the gas turbine and account for the inaccuracies in the sensor measurements them-
selves. The review by Romesis [16] summarized the various CBM methods which have been tried
in the field and have shown promise. This review stresses the fact that though various schemes have
shown to be working for a limited range of data there is no specific diagnostic tool available in the
industry which can integrate the information obtained from various techniques. This fact was also
highlighted by Volponi [18] that there is no single method available in the industry which can take
transients into account and still be applicable system wide due to inherent difference between the
engines.

9
 Condition-based Monitoring

Under CBM and PHM, the fault diagnosis methods can be broadly divided into three cat-
egories - model-based, data-based, information or knowledge-based. The third type (for example
- Expert Systems) did generate a lot of interest in the early 2000s but these methods couldn’t
deal with the complexity of the gas turbine engines and hence fell out of favor. The model-based
methods are still being used in the industry, specifically GPA and some advanced form of GPA
with information fusion. On the other hand, data-based methods are currently being investigated
extensively.
Continuous monitoring in CBM and PHM means that sensor information becomes ex-
tremely important. Thus a lot of research has gone into identifying the optimum set of sensors
to characterize the performance of gas turbine. Methods have also been identified to reduced the
sensor bias and noise, to generate performance curves when machine is working at part load, work-
ing with degraded performance. The parameter selection is an important problem and in most of
the model-based methods it is better to bring down the number of variables in the model to re-
duced complexity and also it requires less numbers of observation to deduce the unknowns form
the model. On the other hand, in data-based methods the data size is not exactly related to the
efficacy of the algorithm but researchers have developed statistics-based methods to find the most
important variables. Some of these methods have been summarized by Stamatis [21].
In model-based methods, the first systematic attempt to correlate the performance mea-
surements of the gas turbine to the health of engine was made by Urban [22]. The GPA method
of detecting anomalies by first quantifying the sensor measurement to health parameters and then
looking for deviations in health parameters has been mainstay of industry for a long time but the
method itself works only when the engine performance is close to linear (i.e. limited to small devi-
ations and thus negligible transients) and it is assumed that the fault can be captured by the limited
and available sensors. To improve upon the performance of this methods, various improvements

10
have been proposed. But as Tahan [19] has suggested, these improvements in GPA are still not able
to deal with very high transients which are specifically the case with aviation engines [3]. And this
is exactly where data-driven methods can help. According to Tahan [19], data-driven methods lack
robustness, so finally it is the hybrid of both physics-based methods and data-based methods which
might show the highest accuracy. Though there are methods available to dynamically model the
gas turbine as its performance deteriorates over time, the model inaccuracy still limits the efficacy
of GPA.

Table 2.2: Summary of GPA and CBM methods

Author Ref. Year Type Topic

Urban [22] 1972 Journal article Introduction to GPA
Ogaji et. al. [23] 2002 Journal article Parameter selection for NLGPA
Ogaji [24] 2003 PhD thesis GPA fault analysis using data-based methods
Marinai [5] 2004 PhD thesis GPA time-series analysis
Mathioudakis et. al. [25] 2006 Journal article Comparison of GPA schemes
Bechini [11] 2007 PhD thesis GPA fault analysis using data-based methods
Li et. al. [26] 2009 Journal article Condition-based health monitoring
Dragomir et. al. [27] 2009 Proceedings Prognostics in Condition-based maintenance
article
Romessis et. al. [28] 2014 Proceedings Condition-based maintenance methods
article
Lee et. al. [12] 2014 Journal article CBM vs PHM methods
Aretakis et. al. [29] 2015 Journal article Turbofan engine health assessment
Verbist [30] 2017 PhD thesis GSP GPA analysis

Data-driven Methods

All the methods that do not rely explicitly on the physics based model of a gas turbine
to determine faults can be broadly classified as data-driven methods. These methods rely on the
specific patterns within data to determine whether some fault has developed in the engine. This is
also a drawback when it comes to gas turbine industry because there are not a lot of data available in

11
the open domain which correspond to identical faults and which can be used to parameterize these
algorithms but still these data-driven methods are helpful because these methods can diagnose the
problem without relying on a rigid model of the machine.
Some of most famous in the field are neural networks and these neural networks were
first applied to the problem of gas turbine diagnostics by Dietz et. al. [13]. The dynamics neural
networks have now been established and these have been able to provide error rate of less than 1%
and these dynamic neural networks can correctly identify fault data from the healthy data [31].
A review of important prognostic metrics and the prediction of remaining useful life of
components using data-driven approaches has been given by Dragomir [27]. Though artificial
neural networks constructed from operational data can detect anomalies in the data containing
faults [32], the algorithm is not robust enough and due to the black-box nature of the algorithm, it
is not very amenable to quantitative changes. A more detailed description of use of neural networks
in gas turbine diagnostics is given by Loboda [15], [33].

Table 2.3: State-of-the-art PHM methods

Author Ref. Year Type Topic
Cory et. al. [34] 2007 Proceedings Autoregression based anomaly detection
article
Kumar et. al. [35] 2015 Proceedings EGT prediction using VAR
article
Chen et. al. [36] 2015 Journal article Measurement selection for fault diagnosis
Kiakojoori et. al. [37] 2016 Journal article Dynamic Neural Networks
Loboda [33] 2016 Book section Neural networks for gas turbine diagnostics
Borguet et. al. [38] 2016 Journal article Regression based fleet monitoring
JinFu et. al. [39] 2017 Journal article Fault detection in hot components
JinFu et. al. [40] 2018 Journal article Fault detection frequent pattern extraction
Bai et. al. [31] 2020 Journal article Nonlinear AR neural network
Rahmoune et. al. [41] 2021 Journal article Dynamic NARX VAR

Review by Fentaye [20] also outlines the correlation between different types of failures in
the gas turbines and its probable signature in the system data. Data non-availability also remains

12
a challenge due to security concerns and sparse numbers of failures in the gas turbines [20]. Still,
there are discrete applications of algorithms on isolated simulated data. The field of aviation has
made better progress due to availability of uniform data-sets provided by NASA (National Aero-
nautics and Space Administration) as summarized by [42]. Despite this there is no single method
that has proven effective system-wide and hybrid methods that includes both physics-based meth-
ods and data-based methods remain promising [19].
Cory [34] first showed as a proof-of-concept that the vector autoregressive methods can be
used to detect anomalies. In this work, the parameters of interest were turbine exit temperature and
compressor pressure ratio. Kumar [35] used a similar autoregressive model to predict the exhaust
gas temperature of a gas turbine and it achieved error rate of less than 1% in prediction but the
data was not high-dimensional in nature and this is exactly what I have tried to improve upon in
this study. Bai [31] proposed a NARX (Nonlinear autoregressive neural network) framework in
which dynamic neural network takes the fuel flow, ambient temperature and pressure ratio of the
gas turbines as input and achieves less than .04% error rate. The authors show that the algorithm
can correctly identify the faulty data but this paper details only the detection of anomaly not the
diagnosis. In my work here, I show that non-linear modeling can even do the diagnosis.

13
 CHAPTER 3: OBJECTIVE AND NOVELTY

The current work demonstrates the following advantages.

1. First, I have used real-time historical data for analysis instead of simulated data. The gas-
path parameters carry precious information about the power plant operational condition.
However, due to the infrequent anomaly observation and extremely expensive physical fault
simulation, data simulated from gas turbine models are more commonly used in the litera-
ture. These simulated data inherently carry bias, so the results lack accuracy.

2. Second, I have used VAR model as a part of this study that assists in understanding the asso-
ciation relationship among variables. In particular, the Granger Causality Network provides
information about the forecasting ability of one parameter to another.

3. Third, I have achieved variable selection and it relieves the computational cost in the online
prediction. Vector autoregressive model is well developed and widely applied in analysing
time series data. However, modeling high-dimensional time series of 200 variables simulta-
neously is a challenge. Curse of dimensionality arises in estimating the model. In particular,
ordinary least square regression no longer exists due to heavy collinearity among the vari-
ables for high dimensional datasets. As a remedy, I relied on LASSO estimation to construct
VAR models which selects the most important contributors from the high-dimensional pre-
dictors, given that only a small number of them are significant. This helps to narrow down
the search and discovers some of the causality which is least expected. By selecting the most
important predictors, the model addresses a crucial issue of having the large sampling time
of the data (in this case 30 seconds). From 200 variables, I have recovered a GCN where
each system variable is dependent only on a group of selected few variables and this can
substantially narrow down the search for original cause of the anomaly.

14
4. In the second part of the study, I have used the hidden Markov model which exploits the
joint information carried by RPM and other responses. The prediction accuracy from GAM
and Random Forest method after segmentation is comparable to that obtained from neural
network. More importantly, as is opposed to modeling with neural networks, the result can
be used to explain the underlying relationships. The GAM model is built assuming additive
nonlinear relationships between predictors and the response, the random forest is built from
decision trees, and they can both be used to select important variables.

5. One specific source of gas-turbine failure is turbine temperature distortion, which can lead
to severe adverse thermal loading of turbine blades. This circumferential distortion is mea-
sured using blade-path temperature spread. In the third part of the study, my aim is to predict
the blade-path temperature measured at each sensor and hence the temperature spread while
keeping the complexity of the model low and for this I have used nonlinear statistical analy-
sis. I had found in the first part of the study that in the gas turbine modeling, the blade-path
temperature depends on a number of variables but the bulk of the dependence can be cap-
tured using only the fuel flow. Variables that are known to be directly associated with the
blade-path temperature, such as combustion temperature and turbine inlet temperature, are
not available in many scenarios. On the other hand, the analysis shows that the effect ex-
plained by variables other than the fuel flow has a strong dependence on its own history. As
a result, in this work I have predicted the blade-path temperature by combining the nonlinear
function of the fuel flow and the time-series analysis. Specifically, first I have captured the
effect that the fuel flow has on the blade-path temperature, and then I have filtered residual
sequence using ARFIMA-GARCH model.

15
 CHAPTER 4: METHODOLOGY

Exploratory Data Analysis (EDA) and Data Selection

In this study the focus is on three supervised learning methods to predict the hot gas-path
variables. Once the sufficient accuracy was achieved in prediction, I applied these methods to the
fault containing data and tried to identify faults. I didn’t use unsupervised learning methods here
because there is a considerable degree of physics based understanding of gas turbines in order to
know which variables should be considered inputs, which ones should be considered outputs and
whether some particular variable can affect other given variables. Most of the algorithms that I
have used in this study are of the regression type where a numerical value is being predicted and
classification type algorithms are limited to identifying the state of the machine.
Since a lot of the machine learning algorithms suffer from bias-variance trade-off, the stan-
dard method of validation set can be used to ensure that error from the training set is not artificially
low which can lead to very high out of sample error in test set. In validation set, we use one set
of training data itself during the algorithm building process to ensure that the model has good pre-
dictability. But, because algorithms designed using validation set can give highly variable results,
k-fold cross validation has been used. In k-fold cross validation we partition the data in k separate
sets and repeatedly do the model training on (k-1) randomly selected sets. The remaining set is
used to test the error. Wherever necessary, the non-dimensionalization of the data has been done
using the methodology given by Volponi [43].

Time-series Analysis

In this study, I have used time series analysis and machine learning techniques. At first,
I estimated the model during the operation of the engine and also based on the offline data and
then I predicted the anomalies for the remaining data stream with the model. In future during the

16
 Figure 4.1: Schematic of one powerplant unit

operation of engine, potential anomalies will be detected using this model, and these anomalies
will be reported with controlled false alarms. These false alarms can then be further treated by
physics-based methods to determine whether an action is necessary.

System and Data Description

In the current study, I have used data that was taken from operational power-plant con-
sisting of identical gas turbine units which feed to a common steam turbine through HRSG units
and this steam turbine in turn provides the steam for the gas turbine units. Figure 4.1 shows the
schematic of one powerplant unit which has three independent gas turbine units. As per suitability,
I have selected different number of variables for analysis for each algorithm and I have analyzed
thermodynamic variables of the hot gas section of the gas turbines. The time-series variables were
sampled at 30 seconds. I have used long time-history of the orders of months to construct the
nominal performance model of the gas-turbine operation. The model is constructed based upon

17
 Figure 4.2: RPM data used for analysis

performance of only one gas-turbine unit and the same model was used to predict anomalies in all
three gas turbine units because these are identical.
For analysis, 8 different data sets are available with each having more than 200 variables
and more than 200,000 consecutive time-series values. These data sets are from different but
identical gas turbines and these data were not acquired simultaneously. These data-sets are not
known to have identical failure among them which makes it a very important consideration for any
machine learning based application because a large number of machine learning algorithms in the
initial phase of learning require significant amount of similar cases to be able to learn effectively.
Moreover, the number of failures are known to be sparse, i.e. every data set has no more than a
few failures and there is no specified definition/criterion of failure/anomaly for analysis, i.e. the
definition of failure is broad in nature in the sense that the system deviated significantly from ideal
behaviour.
I considered modeling of three outputs individually in this study: the gas turbine power, the
average blade path temperature and the blade path temperature spread. For algorithms requiring
stationarity, I considered only those data-points where the RPM is around 3600. See Figure 4.2
for illustration. A sample of input data is also shown in Figure 4.3, where GT power and exhaust

18
Figure 4.3: Sample input and output data for gas turbine

19
temperature have been non-dimensionalized.

Performance Evaluation Formulas

I have used the following measurements to evaluate the prediction error. First, I have used
Root Mean Squared Error (RMSE), which is a widely adopted measurement of the difference
between the prediction and the ground observations. Since RMSE is not invariant to the change
of scales of the response variables, I have introduced the scaled RMSE and the Residual Mean
Squared Percentage Error (RMSPE) for comparison among different data streams and different
sources.

1. The RMSE is defined as -

v
u n
u1 X
RM SE = t (yi − ŷi )2 , (4.1)
n i=1

where yi as the true response value of the ith observation in the testing set, ŷi is its predicted
value and n is the size of the testing set.

2. The scaled RMSE is defined as -

RM SE
scaled RM SE = ,
max(y1 , . . . , yn ) − min(y1 , . . . , yn )

3. RMSPE reports the rate of error as a percentage of the observations. That is, for the predicted
n points in the testing set -

v
u n 
u 1 X (yi − ŷi ) 2

RM SP E = t .
n i=1 yi

20
 Data-based Algorithms

I have first considered a vector autoregressive (VAR) model, which formulates 115 gas
turbine variables as linear functions of their history and another 23 exogenous variables. To use
the VAR model for anomaly prediction, I applied the model to the data set to achieve one-step-
ahead prediction. I then analyzed the prediction errors and the behavior was deemed as anomalous
if the prediction error had a magnitude that exceeded 3 standard deviations.
VAR has been specifically used to model the blade-path temperature, which is an important
parameter to monitor the gas turbine thermal operation. This model requires the assumption of
stationarity to be satisfied. However, from an examination of the blade-path temperature and the
major predictor fuel flow, i.e., Figure 4.8(b), it will be seen that linear assumption is not valid, even
when the gas turbine is operating in full or almost full load. As a remedy, in the third part of the
study, I have built a nonlinear model to capture the association, followed by time series modeling
of the residuals, which remarkably improves the prediction performance.

Vector Auto-regressive Method (VAR)

Vector Autoregressive (VAR) model has been widely used in multivariate time series anal-
ysis in the areas including economics, finance, geometry, genetics, etc; see [44] for examples. The
idea of the VAR model is that the current value of a variable can be written as a linear combination
of previous values of itself and/or other variables, and a random error term that is used to represent
for all factors that can not be explained by the historical values.
To apply the model I have selected the time series such that the gas turbine is operating in
full or partial load, and I have assumed that the behavior of the data streams is stationary throughout
the selected time periods. In particular, I have built the model from the observations with RPM no
less than 3590 and power no less than 0.7 of the maximum load.
A total number of 115 gas turbine variables are modeled simultaneously, which includes

21
 Figure 4.4: Anomaly Detection Flowchart

power output, exhaust gas temperature, etc. The recent history of these 115 variables, as well as
another 23 gas turbine streams serve as the input, including the inlet air temperature, IGV angle as
a proxy for mass flow rate, fuel flow, etc.
I have used here the following VAR(p) models to model the multivariate data streams from
gas turbine.

I p
X X
Yt = α + β i xt,i + Φj Yt−j + ϵt . (4.2)
i=1 j=1

where

• Yt ∈ RK×1 is the vector of endogenous variables which has linear relationship with the
other system variables.

• α ∈ RK×1 is the intercept vector.

22
 Figure 4.5: Time history of normalized power generation

• xt,i , i = 1, . . . , I are the exogenous variables (variables whose values are independent from
other system variables).

• β i is the coefficient vector of the corresponding exogenous variables.

• Φj ∈ RK×K are auto-coefficient matrices.

• ϵt is the noise vector, with mean vector 0 and covariance matrix Σϵ .

• Zt is set to be the RPM of the gas turbine unit.

The assumptions for the model include -

1. The power output and other variables are stationary

2. The uncertainty (error) are white noise vectors

The dependence among variables cannot be assumed to be stationary, as it depends on the

operational stage of the gas turbine. For example, the turbine inlet temperature stays relatively

23
constant during steady-state operation but decays slowly when machine is shut down. Also, the
transition frequencies are non-constant and they could not be treated as a seasonal trend, as can be
seen in Figure 4.5. The operational stage, on the other hand, can be roughly indexed by the RPM
and the power. Therefore, in order to model the overall behavior of the gas turbine, I have split the
dataset and only modeled the stage where the machine is on, and operating in partial or full load.
That is, the model is valid only when the machine is running at full RPM and the power generation
stays above a threshold without excessive transients.

VAR Implementation

In the VAR model, I have considered K = 115 gas turbine parameters, with maximum lag
p = 4, and 23 exogenous variables. As expected, curse of dimensionality arises in estimating of the
model. In particular, ordinary least square regression no longer exists due to the high dimension
and heavy collinearity among the variables in the dataset. As a remedy, I have relied on LASSO
estimation to construct VAR models, which prevents the model from over-fitting and selects the
most important contributors from the high-dimensional predictors, given that only a small number
of them are significant.
Specifically, this is done by incorporating a penalty of the l1 norm of the regression coef-
ficients on random fluctuations. Assuming that the parameter space is sparse, i.e., many values of
βj and Φl (for j = 1, 2, . . . , J and l = 1, 2 . . . p) are zero, we estimate the model by solving the
following parallel K regularized least square optimizations. That is, solving (4.2) is the same as
solving the following individual problems:

I p K
X X X
yt,k = αk + βi,k xt,i + ϕj,k,l yt−j,l + ϵt,k . (4.3)
i=1 j=1 l=1

24
 For each k = 1, 2, . . . , K minimize

T I p
X X X (r)
Φj,k· Yt−j ]2 + λk pk ,

[yt,k − αk + βi,k xt,i + (4.4)
t=1 i=1 j=1

where Φj,k,· is the kth line of Φj , and the first p observations at the connecting point of the thresh-
olding do not participate in the summation. I have adopt the following LASSO penalty

J
X P
X
pk = |βk,j | + j 1/2 ∥Φj,·,k ∥1 ,
j=0 j=1

where ∥ · ∥1 is the l1 norm of a vector. Here I have applied a factor j 1/2 to the jth au-
toregression coefficient vector to encourage a short order of autoregressive model, i.e., a small p.
The tuning parameters λk are allowed to be different to allow for different sparsity level in vari-
ous measurement types. The tuning parameters are selected individually for all k, by block cross
validation.

Hidden Markov Model

In the previous part of the study, I applied regression based statistical algorithm and in
that analysis, in order to satisfy the condition of statistical stationarity, I took only that portion of
the data which correspond to machine performing close to steady state with low variation in ther-
modynamic quantities. The drawback of such model development is that the model showed poor
performance as soon as high transients were encountered which affected its predictive capability.
So, in this part of the study the goal is to apply a data-driven method that can take care of state
change in the data so that functionally dis-similar datasets are not combined for the analysis. This
step increases the predictive capacity of data-based algorithms.
It can be shown that data segmentation does improve the predictive capability of data driven
algorithm. Often, this data segmentation is taken care of using a dummy variable in statistical

25
algorithms but this works well only when the difference between multiple data streams is precisely
clear. The data from a gas turbine doesn’t provide any information about how the data should
be segmented which can be attributed to multiple states of the performance. To overcome this, I
have fitted hidden markov models to divide the data into three broad categories - GT not running,
GT running with steady state, GT running with transients. The data can be segmented in more
categories but that increases the complexity of the statistical analysis and increases the variance
among results of different data-sets, so I have restricted it to only 3 segments.
Hidden Markov Model (HMM) is a basic state-space model [45] that is used to segment
the data in different states of the operation. In hidden Markov models, suppose there are T ob-
servations Ot , t = 1, 2, . . . , T , and there is a categorical process St , t = 1, 2, . . . , T that indi-
cates the state of the system which is not observed. The observations Ot , t = 1, 2, . . . , T , are
temporally independent given St . For the state St , it is assumed that it is a stationary Markov
chain that takes value in {1, 2, . . . , m}. That is, St , the state of machine at time t depends on
Sk , k < t only through St−1 , and the marginal probability distribution of St is stable over time.
The probability of switching from status i to j in one step is denoted as a(i, j), and the matrix
A = (ai,j )1≤i,j≤m is called the transition matrix. For the observation process, it is modeled as
Ot := (powert , RPMt , temperature spreadt ) = (Ot,1 , Ot,2 , Ot,3 )⊤ , t = 1, 2, . . . T are identical
and independent multi-variate normal random vectors given the state St :

(Yt |St = i) ∼ N (µi , Σi ), t = 1, 2, . . . , T, i = 1, 2, . . . , m. (4.5)

Moreover, it is assumed that Σi is diagonal. In other words, given St , power, RPM and temperature-
spread are independent at each time t.
The HMM is fitted by the Baum-Welch algorithm, a.k.a. the Expectation–Maximization
(EM) algorithm. From the model, the posterior probability of St , t = 1, 2, . . . , T can be estimated
given the observations. In addition, the HMM enables us to predict the future status process given

26
the current and historical observations. For example, one can compute the one-step-ahead posterior
distribution P (ST +1 = i|o1 , . . . , oT ). I have skipped the details and interested readers are referred
to time series books such as [46].

Supervised Learning Algorithms after Segmentation

After segmenting the time series with the hidden Markov model, the observation process
is treated as identical and independent with each given state. This allows us to collect the observa-
tions in each state and recover the patterns in the relationship between the responses and predictors.
The following two responses are of interest: The GT power and the blade path temperature spread
with four predictors from the operational data. In this part of the study I have considered three su-
pervised learning algorithms: Generalized Additive Models, Random Forest and Neural Networks.
I have first partitioned the whole data set into a training set and a testing set. I have seg-
mented each of the training set and testing set according to their states, and I have built supervised
learning models in each of the training sets. I have then applied the algorithms to the testing
set for prediction and I have evaluated their performance by comparing their prediction with the
corresponding true values in the testing set.

Additive Model

To detect the anomalies in this part of the study, I constructed an additive model of the
output (e.g. temperature, power, etc.) based on the inputs, which are under control (e.g., fuel flow,
IGV angle, etc). More specifically, in this model, we can write the conditional expectation of the
output as a function of the inputs.

E(Yt |Xt,1 , Xt,2 , . . . , Xp ) = f (Xt,1 , Xt,2 , . . . , Xt,p ), (4.6)

27
where Yt are the output at time point t, a.k.a response or predicted, and Xt,j , 1 ≤ j ≤ p are the p
inputs at time t, a.k.a covariates or predictors.
With given f , one can calculate the expected value of outputs under normal operation from
the inputs, i.e., Yt∗ := f (Xt,1 , Xt,2 , . . . , Xt,p ). If at time t, the absolute value of εt := Yt − Yt∗
exceeds a certain threshold, an alarm should be raised for anomaly with confidence.
However, since f is unknown, a major challenge is to construct an estimation fˆ of f .
Anomalies are detected when the predicted et = Ŷt − Yt is sufficiently large in magnitude, where
Ŷt = fˆ(Xt,1 , Xt,2 , . . . , Xt,p ).
In this study, I have investigated such a prediction problem. Due to the model complexity,
I have further assumed that the relationship (4.6) can be further reasonably simplified into the
following form

Yt = β0 + f1 (Xt,1 ) + f2 (Xt,2 ) + · · · + fp (Xt,p ) + ϵt , (4.7)

where β0 is the intercept term. Comparing to (4.6), it is apparent that in (4.7), interactions between
the inputs are removed. The model (4.7) is known as a generalized additive model (GAM).
The idea behind estimating (4.7) is to obtain a compromise between the goodness-of-
fit and smoothness of fj ’s. In other words, from the training data (Xt,1 , Xt,2 , . . . , Xt,p , Yt ), t =
1, 2, . . . , N , the following quantity is minimized

N p p Z
X X X
(Yt − β0 − 2
fj (Xt,j )) + λj fj′′ (xj )2 dxj ,
t=1 j=1 j=1

where f ′′ is the second order derivative and λj ≥ 0 are the tuning parameters. Smoothness of each
function is controlled by including the second term in the above formula. The reason to control
the smoothness is to avoid over-fitting as in many statistical problems (which means to improve
the out-of-sample performance). The minimizer can be found by a linear combination of basis

28
functions (the natural cubic splines with knots at the observed values). The readers are referred
to [47] for details.
Although both VAR and GAM use regularization to avoid overfitting, the GAM controls
the model smoothness with the spline function which is essentially a ridge penalty; while in VAR,
LASSO is utilized to perform variable selection. I have done this because in VAR modeling I an-
alyzed more than 115 endogenous variables and 22 exogenous variables, while in GAM modeling
I employed only 4 predictors.

Random Forests

Random forest [48] is a well-known supervised learning method. It is developed on the

basis of bootstrapping decision trees. In each of the b = 1, 2, . . . , B steps, a bootstrap sample
is obtained from the training set and a tree, denoted as Tb , is grown on the bootstrapped sample.
In building Tb , a random subset Is ⊂ {1, 2, . . . , p} of predictors are used in splitting each of
the nodes, where the optimal predictor j ⋆ ∈ Is and the optimal splitting value is selected by
minimizing the residual sum of squares in the current region. For a new observation with predictor
x = (x1 , x2 , . . . , xp )⊤ , the the response ŷrf
B
is predicted as an average from all the bootstrapped
trees
B
X
B −1
ŷrf (x) =B Tb (x),
b=1

where Tb (x) is the predicted value from the bth tree.

An advantage of the random forest is that it can perform model selection from the obser-
vations that are not selected to train the tree in the bootstrap step, known as out-of-bag samples. I
have omitted the details here and the reader is referred to machine learning literature (e.g. [47, 49])
for details in the decision trees and random forests.

29
 Figure 4.6: Illustration of a neural network

Neural Networks

Neural Networks are widely used data-driven method for prediction. The underlying struc-
ture behind a neural network is a directed graph, which is also known as multi-layer perceptrons
(MLP), consisted of neurons (vertices) and synapses (edges) that connect between layers. A neural
network has at least three layers: an input layer, hidden layers and an output layer. See Figure 4.6
for illustration that has three input variables; one fully connected hidden layer with five neurons;
one dropout layer; one fully connected layer and the output neuron. This simple neural network is
cited only for illustration purpose. To describe the functional relationship between the response Y
and the predictor X = (X1 , X2 , . . . , Xp )⊤ , the MLP models Y as

Y = g(1, Z),

where Z ∈ RN is the neuron of the size-N hidden layer.

Zk = σ gk (1, X) , k = 1, 2, . . . , N.

30
Each of the hidden neurons Zk , k = 1, 2, . . . , N is the output of neurons or inputs from the pre-
vious layer, processed by an integration function gk : Rp+1 → R and then an activation function
σ : R → R. The integration function gk (1, X) = α⊤
k X computes a linear combination of the

inputs, while the activation function transforms the linear function to nonlinear and determines if
the corresponding neuron is “activated”. In this study the Rectified Linear Units (ReLU) activation
σ(u) = max(0, u) are used. The readers are referred to [47, 50] for details of neural networks.

I have used different structures of the neural networks in predicting power and temperature
spread in various states. In predicting power in state 1, I have processed the input variables (inlet
air temperature, fuel flow, bypass valve position and IGV angle) by two fully connected layers,
with 100 neurons in each layer, followed by a dropout layer, and then two fully connected layers
with 20 neurons. The neural network for state 2 has two fully connected layers, with 200 neurons
in each layer, a dropout layer, and two connected layers with 100 and 50 neurons respectively,
and another dropout layer. The structure of the network for state 3 is two fully connected layers
with 200 and 100 neurons respectively, a dropout layer, a fully connected layer with 50 neurons, a
second dropout layer, and a fully connected layer of size 5. In predicting the spread in state 3, in
the model structure I have included two fully connected layers of size 200 and 100 respectively, a
dropout layer and a fully connected layer of size 10. I have evaluated the training process of the
neural networks by RMSE (c.f. 4.1), which I have used as the early stopping metric.

Additive Model with ARFIMA-GARCH Residuals

A few practical issues arise in VAR model that was studied in part one. First, the linear
assumption is not valid for some variables, especially for the blade-path temperature, which is a
pivotal stream in monitoring the thermal operation of the gas turbine (c.f. Figure 4.8). Second,
some variables that are known to be influencing temperature are unavailable. For example, turbine
inlet temperature is one of the best predictor of the efficiency of the system but since this data is

31
not available to us, we can not use it for prediction. Also, a subset of the operational variables have
missing values during a small time period in the data set, partially as a consequence of sensor error.
Third, the assumption of stationarity is not fully satisfied. Moreover, there is a non-synchronization
issue of some of data streams.
In order to analyze and predict the blade-path temperature, I have built a generalized ad-
ditive model (GAM) from the fuel flow, which is the only predictor utilized in GAM. Due to the
lack of predictors, which are either unknown or unavailable, a large proportion of volatility in the
blade-path temperature is not captured by the additive model, resulting in a large magnitude of
prediction error. It is easily seen that the error term is serially dependent, with a long memory and
non-constant variance. This informed me to apply the ARFIMA and GARCH models to model the
errors, which significantly improves the prediction of the blade-path temperature and the anomaly
detection from the latter.

The Additive Model

Now let Yt be the blade-path temperature obtained at an individual sensor at time t. To

model this, the only exogenous covariate that is included in the model is the gas turbine fuel flow,
which is denoted as Xt . After examining the lag relationship between the fuel flow and temperature
(c.f. Figure 4.7 and Figure 4.8), I chose to include Xt as a nonlinear external predictor in the mean
function of Yt because the correlation between these two variables is -0.89 which means 80 percent
((−0.89)2 ) of variation in the temperature can be captured by its linear relationship with fuel flow.
A nonlinear relationship is revealed between Xt and Yt , as is shown in Figure 4.8. Therefore, I
have constructed a generalized additive model (GAM) between Yt and Xt .

Yt = f (Xt ) + ξt , (4.8)

32
 Figure 4.7: Training data non-dimensional temperature vs fuel flow

where f is a smooth function of Xt , which measures the conditional value of Yt given Xt . The
function f (t) is assumed as a smoothing spline and estimated from regularized regression from
natural spline basis. The term ξt accounts for the effect that cannot be explained only by the fuel
flow, which is a result of other variables, such as the combustion temperature and their interaction
with the blade-path temperature, whose measurement is unavailable. However, I found the ξt
process to be temporally dependent; see Figure 4.9. In other words, the sequence (ξt )t∈Z is not
independently distributed and its history can be used to predict the future.

AFRIMA-GARCH Model for the Residuals

As shown in Figure 4.9, the variances of ξt is non-constant (heteroskedastic). Instead,

there are time windows with high volatility and others with low volatility. A Generalized Au-

33
 Figure 4.8: Training data scatterplot between Xt and Yt .

toregressive Conditional Heteroskedasticity (GARCH) model [51, 52] is generally employed to

capture this effect. Moreover, the ACF plot in Figure 4.10 exhibits slow-decaying autocorrelations
ρ(h) = corr(ξt , ξt−h ), h = 1, 2, . . . , 50, and thus indicates a long-range dependence effect. In
Figure 4.8, lag of 50 corresponds to the 25 minutes of operation time. Autocorrelation function
(ACF) provides a quantitative measure of time-dependence between a variable and its own time-
history which is given by the term ”lag” h. For a time-series with short memory, the ACF satisfies
that ∞
P
h=0 |ρ(h)| < ∞, otherwise the sequence is said to have long-memory, which is the case

here. Therefore I have used an ARFIMA model [53, 54], which is a commonly used time-series
parametric method to model the long-range dependence.

34
 Figure 4.9: Time-series plot of the residuals ξt of the GAM model

I have modeled the sequence (ξt )t∈Z with an ARFIMA (0,d,0)-GARCH(1,1) model, that is

(1 − B)d ξt = ϵt , (4.9)

and

ϵ2t = σt2 Zt2 , (4.10)

σt2 = α0 , +α1 ϵ2t−1 + β1 σt−1

2
(4.11)

where B is the backshift operator; Zt ’s are assumed to be identically and independently distributed
Gaussian random variables; d, α0 , α1 , β1 are parameters to be estimated and d satisfy that |d| < 0.5.
The model (4.9) to (4.11) depicts both the long-memory effect and the heteroskedasticity

35
 Figure 4.10: Autocorrelation function (ACF) plot of the residuals ξt of the GAM model

effect. I have implemented the GAM and ARFIMA-GARCH models with the R package mgcv
[55] and rugarch [56] respectively.
After training the model, the prediction of Yt is composed of two steps: rolling-window
forecasting of ξt from the ARFIMA-GARCH model, and implementing the forecasted residual
process into the GAM model. Specifically, most recent histories are used to predict each point
in the testing set Ym+1 , Ym+2 , . . . , YT . Here m is the size of the rolling window. That is, to pre-
dict the temperature at a testing time point t, I first computed the residuals in the initial window
ξt−m , ξt−m+1 , . . . , ξt−1 by plugging Yt−m , . . . , Yt−1 , Xt−m , . . . , Xt−1 into the GAM model fitted
from the training data.
The one-step-ahead forecast of ξt and σt are obtained from ξt−m , ξt−m+1 , . . . , ξt−1 , which
I have added on top of the estimation from the GAM prediction in turn. I have applied the above

36
procedure to the prediction of each point Ym+1 , Ym+2 , . . . , YT in the testing set. As a result, the
beginning m points are not predicted, unless these are adjacent to the training set.
The intellectual merit of the work is that by using this method we can get an improvement in
anomaly detection by incorporating nonparameteric additive model with time series models. And
then we can predict blade path temperature spread using thermodynamic variables. Moreover, the
run-time residuals obtained from GARCH model can be used during operation with the help of
control charts.

37
 CHAPTER 5: RESULTS AND DISCUSSION

Outlier Detection

Outlier detection is the first step in detecting anomalies and preparing the data, in which
data-points farthest from the mean are considered to be anomaly. This rote learning method works
best when the deviation is large enough compared to the standard deviation. This step helps us
in finding sensor failures as well. I have shown one such anomaly in 5.1 where the data of the
majority of days appears in a cluster but on one particular day the data is significantly different
and thus with this simple analysis it can be inferred that an anomaly may have occurred. Though
useful in its own way, this method itself is not helpful in finding subtle pattern changes which can
be attained with the rest of the analysis.

Figure 5.1: Outlier detection

38
 Figure 5.2: Non-dimensional one-step-ahead prediction of GT power

Figure 5.3: Non-dimensional error for one-step-ahead prediction of GT power

Vector Autoregressive Model

This section is devoted to present the one-step-ahead prediction and error analysis results
for the power and exhaust temperature from the VAR model. In this part of the study, I have
partitioned the dataset used for modeling into training and testing set. Because time-series model
is deployed, I have selected only a pilot window at the beginning to train the model, and I have used

39
 Figure 5.4: Non-dimensional one-step-ahead prediction of blade-path temperature

Figure 5.5: Non-dimensional error for one-step-ahead prediction of blade-path temperature

the rolling prediction for the rest of days. The pilot training window contains sufficient number of
observations without an anomaly in order to obtain a statistically accurate estimation of the model.
I have calculated the deviations between the observed and predicted values, accompanied with
the sample standard deviations of the prediction error for each stage. The results for this vector
autoregression model are shown in Figure 5.2 to Figure 5.5.
The one-step-ahead prediction for power and the exhaust temperature are presented in Fig-

40
 Figure 5.6: Non-dimensional one-step-ahead prediction of blade-path temperature spread

Figure 5.7: Non-dimensional error for one-step-ahead prediction of blade-path temperature spread

ure 5.2 and 5.4 respectively. I have rescaled the data in the figures and I have taken J = 23
exogenous variables. I have selected the tuning parameters individually for k = 1, . . . , 115, by
minimizing the mean square error from the 10-fold block cross validation. The spiky nature of the
graph in Figure 5.2 and Figure 5.4 is due to discrete segments joined together by straight line.
In Figure 5.3 and Figure 5.5, I have presented the prediction errors, along with the interval
with length of 3 standard deviations of the corresponding stages. The errors exceed the 3 stan-

41
 Figure 5.8: Predicted vs observed GT power

Figure 5.9: Predicted vs observed blade-path temperature

42
 Figure 5.10: Predicted vs observed blade-path temperature spread

dard deviation threshold often at the transition. Since it is known that there are no failures at the
corresponding time instants, these alarms are false positives.

Anomaly Detection from the VAR Model

For the implementation in the modeling, I have taken J = 23 numerical exogenous vari-
ables. Also, I have taken a relatively large lag order p = 4 and in each regression, I have employed
LASSO to select from all the regression covariates. The order p = 4 seems to be sufficient as
very few covariates are selected with a more than one lag largely due to the fact that weights in the
algorithms are made to decay exponentially.
Compared to monitoring prediction error for a individual variable, simultaneously moni-
toring the prediction errors of the variables over a time period requires a wider margin of error
to control the false alarm rate. This is the reason that I have taken 3 standard deviations instead
of 2, the commonly used in applications. The grey lines in figures do not show the simultaneous

43
 Figure 5.11: Plots of the relationship between inputs and the outputs for gas turbine power.

confidence band, which involves distribution theory of the LASSO prediction of high-dimensional
time series, and I have left it as a future problem.
As can be seen in figures 5.8, 5.9 and 5.10, the algorithm predicts power and temperature
very accurately, the blade path temperature spread is not accurate given its exceedingly high false
positive rate. It can be seen that during transients the error is higher. This behavior is expected due
to non-stationarity which these transients introduce in the data.

44
 Figure 5.12: Plots of the relationship between inputs and the outputs for turbine temperature.

Generalized Additive Models

In this section, I have constructed generalized additive models (GAMs) which lend insight
to understanding of the non-linear relationship between the system inputs (e.g., fuel flow, IGV
angle, etc.) and outputs (e.g., temperature, power, etc.) of the system. Moreover, by applying the
GAM, the expected value of the outputs under normal behavior are estimated, and thus provide
a basis for anomaly detection. Due to the nature of the method, this analysis method will have
similar or better performance when it is used on similar data-set that is sampled at the interval of
one second or less.
From the initial dataset, I have used two thirds of the observations as the training set to

45
 Figure 5.13: Plots of the relationship between inputs and the outputs for temperature spread.

build the GAMs and the other one third as the testing set. For each output, I have constructed
the GAM with smoothing splines of each input with four degrees of freedom, with the R package
gam [57]. In Figure 5.11 to 5.13, I have plotted the relationship between the outputs and each
input, holding the other inputs fixed.
In Figure 5.11, fi , i = 1, 2, 3, 4 are the contribution to turbine power from fuel flow, IGV
angle, bypass valve position and combustion chamber pressure, respectively (c.f. (4.8)). The solid
lines show the fitted function and the dashed lines show fitted function plus/minus two standard
errors. Similarly, in Figure 5.12, fi , i = 1, 2, 3, 4 are the contribution to turbine temperature from
fuel flow, IGV angle, bypass valve position and combustion chamber pressure, respectively (c.f.
(4.8)). The same applies to Figure 5.13 where I have shown the contribution to temperature spread.

46
 The top left graph in Figure 5.11 indicates that fixing IGV angle, bypass valve position and
combustion chamber pressure, the GT power increases with the fuel flow almost linearly, until it
reaches a certain value (which is not revealed in the non-dimensional figure), where the growth
rate is slower.
It may be argued that the relationships in Figure 5.11 and Figure 5.12 are close to linear
and thus a linear model may perform reasonably well. Comparatively, the temperature spread is
more difficult to model. GAM is more advantageous in this case. Indeed, the highly non-linear
function (such as the sinusoidal relationship between IGV valve position and temperature spread)
makes it erroneous to predict with a linear model.
After training, I have applied the models to the testing set and I have exhibited the compar-
ison between observation and prediction in Figure 5.14. The black line displays the true values and
the blue line shows the predicted values. In Figure 5.15, I have plotted the absolute deviation of the
predicted values from the observed values for each output. Large deviations are signs of anomalies.
In fact, I confirmed that the identified ending period of the testing set is indeed a distinct period
that should be differentiated from the normal operation.

Supervised Prediction using Hidden Markov Model

In this part of the study, I have modeled power and blade-path-temperature-spread by only
four predictors: inlet air temperature, fuel flow, bypass valve position and IGV angle. In modeling
state 1 and state 2, I have added a derived variable ’state-time’, which is the time length since the
machine switches to the current state. The data are representing 15 days operational time series.
Out of the 15 days data, I have selected 10 days as training data and the other 5 days as testing
data.
In this part of the study, I have performed hidden Markov modeling on power, RPM and
blade-path-temperature-spread with the R package depmixS4 [58]. I have used the Bayesian In-

47
Figure 5.14: Predictions of the test set for gas turbine power, turbine temperature and temperature
spread, respectively.

formation Criterion (BIC) as the model selection criterion. From the perspective of the BIC values
and the interpretation, I have selected the number of states to be three. We can interpret the states
as “transition state”, “off state” and “on state” respectively. Figure 5.16 shows the behavior of
power and blade-path-temperature-spread in each of the state.
I have segmented the training and testing data respectively according to their posterior
state probability from the hidden Markov model. I have then applied three supervised learning
algorithms - GAM, random forest, and neural networks to analyze the behavior of power and blade-
path-temperature-spread in each state. I have done the analysis using R packages mgcv [55, 59],
randomForest [60], and keras [61] respectively. I remark here that the one-layer MLP in Figure 4.6
is just for illustration. All the neural networks that I have used consist of multiple fully-connected
layers and drop-out layers, and the structure varies between states and responses.
I have shown predictions for each response, algorithm, and state respectively. Since the

48
Figure 5.15: The absolute deviation of the predicted values from the observed values for gas turbine
power, turbine temperature and temperature spread, respectively.

data is non-dimensional, I have summarized the scaled root mean square errors (RMSE) for each
of the models in Table 5.1 and Table 5.2.
In this section, I discuss the prediction performance. I have non-dimensionalized and re-
scaled both power and blade-path-temperature-spread to [0, 1]. In predicting the power, all three
methods work well in all the three states, and accordingly scaled RMSE is less than 1%, except for
neural networks in state 3 which is 1.77%. The neural network slightly outperforms its competitors

49
Figure 5.16: States estimated from the hidden Markov model for power and blade path tempera-
ture spread.

Figure 5.17: Prediction of GT power using GAM.

50
 Figure 5.18: Prediction of GT power using random forest.

Figure 5.19: Prediction of blade path temperature spread using neural networks.

51
 Figure 5.20: Prediction of blade path temperature spread in State 4 (the “on” state).

Table 5.1: Scaled RMSE of GT power generation

Method State-1 State-2 State-3

GAM 0.0046 1.4 × 10−5 0.0051
Random Forest 0.0063 1.9 × 10−5 0.0064
Neural Network 0.0034 2.5 × 10−5 0.0177

in state 1 while its performance is worse than that of GAM and random forest in State-3.
The prediction of temperature spread is more challenging, especially in the ”off state”. For
the temperature spread, only State 3, the “on state” is of interest. The prediction errors in this state
are around 2% for all three algorithms. Comparing between the performance of the algorithms,
I found that GAM gives slightly better performance than random forest and neural networks in
state-3.
From Figure 5.17 to Figure 5.20, I have shown contrast of predicted value and the observed

52
 Table 5.2: Scaled RMSE of turbine temperature spread

Method State-1 State-2 State-3

GAM 0.099 0.079 0.016
Random Forest 0.066 0.098 0.017
Neural Network 0.062 0.104 0.024

Figure 5.21: Combined Prediction of GT power and blade path temperature spread.

value in the testing set by each of the models. In Figure 5.17, in the top-left corner GAM prediction
for State 1 (the transition state) is shown, in top-right GAM prediction for State 2 (the off state) is
shown; In bottom-left corner GAM prediction for State 3 (the on state) is shown; and in bottom-
right corner the combination of GAM prediction from all three states is shown. The values of State
2 (off state) are so close to 0 that they are hidden behind points of other state in the combined plot

53
 Table 5.3: Comparison of performance of power prediction of the four predictors.

Predictor Combined GAM RF NN

RMSE 0.0043 0.0099 0.039 0.2

in the bottom-right panel. The same state-wise prediction applies to Figure 5.18 and Figure 5.19.
In Figure 5.20, on the left GAM prediction is shown; in the middle Random Forest prediction is
shown; and in the right panel Neural Networks prediction is shown. Since we are not interested in
predicting the performance in State 1 and 2 (transition and off state) the prediction is not presented
here.
I found that all three methods predict the power well for each state. It is because the values
of the power are so close to 0 in the “off state” that the predictions appear to deviate from the
observed power on the top right panels and they are not revealed in the bottom right panels in
Figures 5.17 to 5.19.
In Figure 5.21, I have presented the combined prediction for power and spread with the
three algorithms for a final testing set. That is, treating the aforementioned testing set as validation
set and RMSE as the criterion, according to Table 5.1 and Table 5.2, I have selected the optimal
algorithm in each of the states and I have predicted the power and spread for the final testing set
accordingly. In Figure 5.21, the dots represents the observed data and the green line represents
prediction. For State 1, power is predicted by neural networks; For State 2, power is predicted by
GAM; For State 3, power is predicted by GAM and spread is predicted by GAM. Spread in States
1 and 2 is not predicted because those are not of any interest.
From Table 5.3, I found that the combined prediction outperforms all the other predictors,
which are trained on the unsegmented training set. This is because after segmentation, the algo-
rithms in each state are adapted in parameter tuning, and the segmentation allows for the flexibility
to choose the best algorithm for each state.
In summary, in this part of the study I considered a two-step modeling of the power and

54
temperature-spread, where I first estimated the status of the machine from “on”, “off” and “tran-
sition”, according to the mean and volatility level of the RPM, power and temperature-spread,
and then I constructed supervised machine learning models to estimate the predictive relationship
between power and temperature-spread and their predictors.
There are some limitations of the method that I need to be point out. First, the motivation of
data segmentation is that it is believed that when the machine is under different modes, the response
variables (i.e., power and temperature-spread) will behave differently given their predictors. How-
ever, in (4.5), I assumed that the three observation variables RPM, power and temperature-spread
are synchronized and share the same phase-transition points. However, this is not true in practice.
There is a small delay. For example, RPM leads power for a few steps. Second, it is assumed
that the response variables are independent and identically distributed given the status. This is
a simplified assumption which does not account for the predictors. As a consequence, it can be
found from the magnified graph in Figure 5.16, that there are counter-intuitive estimation of status
near transition points. The following more sophisticated state-space model can be considered as a
remedy.

Yt |St = fSt (Xt,1 , Xt,2 , . . . , Xt,p ) + ϵt , (5.1)

where Yt is the response such as power and Xt,j , j = 1, 2, . . . , p are the predictors such as
fuel-flow. Their relationship f depends on the status of the machine. The error term ϵt is used to
model everything in the response that cannot be explained by the predictors and its distribution is
also assumed to be different for various status. I consider exploring the linear and nonlinear model
of (5.1) as future work.

55
 GAM-ARFIMA-GARCH

I have presented the prediction performance here for the GAM modeling and GAM-ARFIMA-
GARCH modeling respectively. I have consider both scenarios where the machine is operating
without an anomaly in its blade-path temperature, and where perturbation is introduced manually
in the numerical simulation. I have presented the results in the next two subsections respectively.
The performance of the time-series enhanced GAM model can be assessed by applying it
to the operational gas-turbine data in its original form. I have taken the dataset from a period when
it is known that no fault or anomaly occurred in the system. In the study, I have used observations
of two days to train the models, and the next 33 days as the testing set. The testing set is adjacent
to the training set and all the 35 days are consecutive with machine running at full RPM.

Performance on Real Gas Turbine Operating Data

In the first step, I have considered only the effect of fuel flow on the blade-path temperature
and I have used a generalized additive model to capture the effect of that dependence. It is known
that the fuel flow should be able to explain the majority of the variability in the blade-path tem-
perature. Indeed, the correlation between the blade-path temperature and the fuel flow is -0.89. In
other words, the linear relationship between the dependent and independent variables is able to ex-
plain approximately 80% of the variation in the blade-path temperature. As is seen from 4.8, their
relationship is not perfectly linear. In fact, The correlation between the blade-path temperature and
the fitted linear predictor from GAM, i.e., f (Xt )) in (4.8) is 0.924.
I have shown the fitted GAM model in Figure 5.22, with a confidence interval and a predic-
tion interval. The interval shown in green is a 95% Bayesian confidence interval of the function f
resulted from the estimated coefficients for the smoothing splines. It has been shown to match the
across-the-function coverage probability under technical conditions (c.f. [62, 63]). For this data,
the independent and identical assumption for ξt ’s is not satisfied and the radius is underestimated.

56
Figure 5.22: The nonlinear function relationship between the fuel flow and blade-path temperature
for the testing set.

This fact is not severe since there is a large training sample size and the green Bayesian confidence
interval is almost negligible. The interval shown in gray is a 95% prediction interval for new data
points with the corresponding fuel flow values. It accounts for uncertainty from both model es-
timation and the dispersion of ξt ’s, assuming independence and that ξt ’s are normal. Due to the
large sample size, the prediction interval almost solely resulted from ξt ’s, which is assumed with a
constant variance in the GAM model.
Due to the difficulty of measuring other combustion chamber variables in the gas tur-
bine, I have modeled the residual blade-path temperature variation from its own history using
an ARFIMA-GARCH model which is much more suitable to deal with the persistent and variable
variation in the residuals. In other words, after removing the effect of the fuel flow, I have modeled
the residual series using ARFIMA-GARCH model to capture the long-memory in mean and the

57
 Figure 5.23: Prediction of blade-path temperature using GAM

conditional variance on the history.

I have shown the GAM prediction of the blade-path temperature in Figure 5.23, while the
combined prediction is shown in Figure 5.24, along with a prediction band of 2 standard deviation.
Since the data is non-dimensionalized, I have used the root mean squared percentage error
(RMSPE) to evaluate the forecasting performance quantitatively. In Table 5.4, I have reported
the RMSPE for both models of prediction for the first, second and last day in the testing set, as
well as all the 33 days tested. The prediction range is only adjacent to the training set in the
first scenario. As is seen from Table 5.4, the time-series component significantly improves the
GAM prediction. Indeed, the GAM-ARFIMA-GARCH prediction achieves a low rate of 0.23%
in the scaled RMSPE throughout the 33 days. Moreover, the accuracy does not decrease as the
distance increases between the predicted time window and the training window in the rolling-

58
 Figure 5.24: Prediction of blade-path temperature using GAM-ARFIMA-GARCH

window prediction.

Table 5.4: RMSPE from the GAM and GAM-ARFIMA-GARCH predictions respectively for var-
ious days of the testing set.

RMSPE day 1 day 2 day 33 days 1-33

GAM 0.0038 0.0042 0.0041 0.0055
GAM-ARFIMA-GARCH 0.0022 0.0022 0.0020 0.0023

Apparently, the variance of ξt may not be constant. As can be seen from the plot, ARFIMA-
GARCH enhances the prediction in this case.
Moreover, the variance forecast adapts to the non-uniformity of the series and the shaded
region which mark the two standard deviation of the GAM-ARFIMA-GARCH forecast has a good
coverage of the true observations. The GAM prediction, on the other hand, apparently deviates

59
Figure 5.25: Standardized prediction residuals from the GAM and GAM-ARFIMA-GARCH mod-
els in the testing set.

from the true observation and it has a much larger prediction RMSPE. This is understandable,
because the rolling time-series prediction takes advantage of not only the fuel flow, but also the
recent history of the blade-path temperature before the time point being predicted.
I remark here that the 2 standard deviation bands do not serve as confidence bands in the
plot and thus do not serve directly for the purpose of anomaly monitoring. First the residuals are
not necessarily Gaussian. In fact, the QQ-norm plot (c.f. Figure 5.38) shows that they have heavier
tails than Gaussian. Second, the conditional variances are estimated individually.
In Figure 5.25, I have shown the standardized prediction residuals and in Figure 5.26 &
5.27, I have shown the squared residuals from the GAM and the GAM-ARFIMA-GARCH model
respectively. I found that the conditional variance forecast is in good consistency of the true resid-
ual squares. The conditional variances estimated from the GAM model, on the other hand, are

60
 Figure 5.26: Squared residuals and predication variances from GAM

estimated as a constant.

Performance on Gas Turbine Data with Simulated Perturbation

In order to assess the effectiveness of the algorithm for anomaly detection, I manually
introduced perturbations to the real data. The criteria for the online monitoring is that the algorithm
can detect the change soon after it occurs, while not raising false alarms frequently before there is
a true anomaly. I remark here that this work focuses on the prediction, which is a necessary step to
generate the noise for health monitoring. I have presented the sequential analysis and testing based
on the obtained white noise process in the next section.
Various types of anomalies are found in practice [64] which are broadly segmented into
point, contextual and collective anomalies. For this study, I have simulated the perturbations with

61
 Figure 5.27: Squared residuals and predication variances from GAM-ARFIMA-GARCH

various signal-to-noise ratios and I have considered two schemes for analysis: abrupt jumps and
linear trends. Specifically, for the abrupt change case, I have manually added 1%, 5%, 10% step-
jumps to the original data at a change point in the testing data respectively in each simulation
scenario, as shown in Figure 5.28. Whereas for the slow-varying trend, I have simulated a linear
increase and I have added these to the data lasting for 100 observations, with a rate of 0.02%, 0.1%
and 0.2% per time-step respectively, as shown in Figure 5.29.
In Figure 5.24, I have plotted the nondimensional blade-path temperature against opera-
tional time with no perturbation and the GAM-ARFIMA-GARCH rolling predication performed
much better than the GAM prediction (Figure 5.23). In Figure 5.30, I have shown the prediction
when there is an artificial step-jump is added to the real data. I have shown the enlarged versions
of figures 5.30 & 5.31 in Figure 5.32 & 5.33. As expected, the GAM predication, which only con-

62
 Figure 5.28: Simulated perturbation: step-jump in the temperature.

siders fuel flow, detects this deviation. This can be inferred from the presence of large prediction
residuals.
The GAM-ARFIMA-GARCH, on the other hand, also captures the sudden jump in the
variance, although after some time it takes the increased history as inputs and thus no longer
considers this as anomaly. As can be seen in Figure 5.32, the ARFIMA prediction adapts to the
jump in approximately 10 points. This is due to the fact that the ARFIMA forecast of ξˆt+1 =
PT
j=0 ϕj ξt−j is a linear process of ξt , ξt−1 , . . . ξt−T , with a large integer T and decaying weights

ϕj ’s in magnitude. It adapts quickly to the recent changes. Thus, the artificial introduction of
anomaly is captured by both the methods here but ARFIMA-GARCH has the advantage because
in the absence of anomaly, the accuracy of GAM is low compared to GAM-ARFIMA-GARCH.
But the capacity of the algorithm to identify the deviation depends on the size of the jump against

63
 Figure 5.29: Simulated perturbation: increasing linear trend in temperature.

the noise level without the jump, i.e., the signal-to-noise ratio.
In Figure 5.31, I have shown the prediction when a linear trend response is added as an
anomaly to the real data. Again, GAM predication easily detects this deviation because inputs
to the GAM remain the same (fuel flow) but the output has changed. The ARFIMA-GARCH,
in this particular case, does exhibit deviations in the forecast and observation. Whether or not
GAM-ARFIMA-GARCH is able to detect this trend as an anomaly depends upon the steepness of
this deviation, i.e., the signal-to-noise ratio. In the graph, it is identified as an anomaly because it
lies outside the two standard deviation confidence interval to ARFIMA-GARCH prediction. The
higher the signal-to-noise ratio is, the easier for the signal to be detected. When the signal is so
weak that it is disguised by the noise (e.g., Figure 5.34) it is hard to be detected by any algorithm.
Here, by signal I mean the strength of the change, i.e., the size of the step and the slope of the

64
 Figure 5.30: Prediction comparison for 5% step-jump simulation

ramp. Note that the current procedure has a low power in detecting ramp. A preprocessing step of
differencing is expected to enhance the power. In other words, after the GAM modeling, one needs
to take first order difference on its residuals ∇ξt = ξt − ξt−1 and construct the ARFIMA-GARCH
model on ∇ξt ’s. I have not pursued it in the current work and I have left it as a future problem.
I have shown the standardized residuals of time-series enhanced prediction (GAM-ARFIMA-
GARCH) in Figure 5.36 & 5.37, i.e.,

zt = (Yt − Ŷt )/σ̂t

In Figure 5.36, I have shown the residuals for the case when the step-jump anomaly is added to the
data and in Figure 5.37, I have shown the residuals for the case when the slowly increasing anomaly
is added to the data. Under the assumption that there is no anomalies, zt in (4.11) are uncorrelated

65
 Figure 5.31: Prediction comparison for 0.1% per time-step ramp simulation

with a mean zero and a variance of 1. The residuals zt that we obtain after pre-whitening and
standardization will serve as important data for statistical online tests for anomaly detection and
health monitoring.
Although I have not conducted an online test on the residuals in this work, it can be seen
from Figure 5.36 & 5.37 that the residuals are high at the point of anomaly introduction, which
indicate an alarm to be raised that the operation deviates from that of the training set.

Control Charts

Using the standardized residuals, I utilize the control charts in order to monitor the onset of
anomaly in an objective manner. Specifically, standardized residuals zt from the GAM-ARFIMA-
GARCH are input of the exponentially moving weighted average (EWMA) control chart. The

66
 Figure 5.32: Enlarged prediction comparison for 5% step-jump simulation

Figure 5.33: Enlarged prediction comparison for 0.1% per time-step ramp simulation

67
Figure 5.34: Prediction comparison for 0.02% per time-step ramp simulation

Figure 5.35: Prediction comparison for 0.2% per time-step ramp simulation

68
 Figure 5.36: Standardized temperature residuals: abrupt jump sizes of 1%, 5% and 10%.

reason for averaging is to remove noisy observation from the systematic anomalies.

EW M At = λzt + (1 − λ)EW M At−1 , (5.2)

where, EW M A0 is mean of the past data; λ is the memory tuning parameter and it is kept
between 0 and 1; zt is standardized residual at time t.
I had set the in-control average run length (ARL0 ) as 12000 time steps. The manually
introduced jump was detected promptly, 1 step after the occurrence, with only one false alarm at
the beginning of the monitoring process.

69
Figure 5.37: Standardized temperature residuals: linear trend with rates of 0.02%, 0.1% and 0.2%
per time-step.

Limitations of Data Based Methods

It is now believed that no technique is expected to work in a standalone manner to give

good results but a combination of multiple techniques gives superior results. In case of Anomaly
failure, three important questions need to answered-

• What kind of failure/anomaly has occurred? In this case we try to look for some informa-
tion/identification in the time history of variables to find any signature which can help us to
pinpoint the type of failure itself.

• How significant is the problem? In this case we try to gauge the extent of the problem. This
can help us in the real time decision making and for similar future occurrences.

70
 Figure 5.38: Quantile-quantile plot for the residuals

• Why did the failure/anomaly occur? In this case we try to answer why the system devi-
ated from the ideal performance and assess the root cause of the failure in order to better
understand the nature of the problem itself and to prevent it from happening in the future.

The physics based methods can answer all three questions - ‘why’, ‘what’ and ‘how much’
- when it comes to diagnosis but they are relatively slower, whereas statistical methods can only
tell ‘what’ and ‘how much’ rather than answering ’why’, these methods can do so relatively faster.
Even with this advantage, one of the biggest drawback of these methods is that the past history of
the variable is the ’true engine’ for these methods. These methods can’t differentiate whether an
observed trend in the past itself is an anomaly if it has been consistent. For example, in Figure 5.40
& 5.41 below the exhaust temperature at the higher power is lower compared to low power. Using
physics based knowledge it can be explained that since at high power, pressure ratio is higher and
thus turbine works more efficiently, taking more power out and leaving less temperature in the
exhaust. But with data based approach one can only discover this pattern exists, but can never

71
 Figure 5.39: EWMA control chart for 1% manually introduced step jump.

know whether or not it should exist. It is the combination of both physics-based and statistics-
based approach that uncovers the underlying behavior of the system and prevents accidents from
happening.
Moreover, the failure shown in Figure 5.42 can easily be detected (once it has occurred)
by lag one algorithm. These drastic failures don’t even need a statistical method for confirmation.
They can be known even by a simple comparison of thresholds of output values. These kinds
of failures may may not be detected beforehand from the variable itself but on the other hand
progressive failure as given in Figure 1.2 can be predicted with a good statistical model.
The vector autoregrssive method may be able to give crucial insight when it is applied on
the simulated data where sensor error can be excluded, and a lot of characteristics of the data can be
controlled. But since we are dealing with real data, there are additional problems. These methods
automatically behave better when the information about the system and the data itself increases but

72
 Figure 5.40: Gas turbine performance - exhaust gas temperature

unfortunately that is not the case with practical systems. For example, the turbine inlet temperature
is one of the best predictor of the efficiency of the system but since this data is not available, it may
not be predictable.

73
Figure 5.41: Gas turbine performance - combustion pressure

Figure 5.42: Power generation failure

74
 CHAPTER 6: CONCLUSION

In this study, I first constructed a threshold vector autoregressive model on historical data
from a power plant. I then analyzed prediction errors to detect anomalous behavior. In the next
step, I also constructed generalized additive models to predict GT power, GT temperature and the
temperature spread. I constructed GAMs by smoothing splines that allows for non-linearity. With
the model, I successfully identified a distinct working mode from the testing data. The high false
positive rate of the model is due to non-stationarity of the data.
To alleviate this, in the next part of the study, I conducted a predictive analysis of power
and blade-path temperature spread. Assuming that the behavior of the responses and the RPM
are synchronized, periodically stationary and independent given the states, I segmented the status
of the machine into “on state”, “off state” and “transition state” via a hidden Markov model by
exploiting the joint information carried by the three variables. After segmentation, I applied statis-
tical learning algorithms for prediction and pattern recognition: GAM and random forest, for each
state and each response power and blade path temperature spread. I then compared the prediction
performance of the models to the widely accepted machine learning algorithm neural networks.
I showed that the method works with a satisfactory performance. In predicting the power,
all three methods worked well in all three states, where the error rate was less than 1% when I
assembled the algorithms. The prediction of temperature spread was more challenging. I found
that the random forest has a slightly better performance than its competitors in the transition state,
while in the on state, neural networks outperforms in terms of the RMSE. The predictions in “off
state” is the worst among all states, but it is less interesting. Neural network is popular in prediction
for gas turbine data but it is well known to be difficult for interpretation. In this study, I found that
the GAM and random forest, which are strong in interpretation, have comparable performance as
neural networks after the HMM segmentation.
The VAR models were able to predict power but their accuracy was little less in the case of

75
predicting the exhaust gas temperature. This was largely due to the fact that linear assumptions of
VAR were not satisfied. Thus, a nonlinear additive model was required to predict the blade-path
temperature. Therefore, in the third part of this study, I employed time-series modeling to enhance
the machine learning prediction of the blade-path temperature. The only exogenous variable that I
used in this model was the fuel flow, as other input factors such as combustion chamber temperature
and turbine inlet temperature were unavailable and it also reduced the complexity of the model. I
captured the long memory and the conditional heteroskedasticity effects in the ARFIMA-GARCH
model. In the numerical study, the inclusion of the ARFIMA-GARCH component significantly
improved the forecast from GAM as the most up-to-date information was included in the rolling
window forecast.
In terms of anomaly detection, I showed in the simulation study that the proposed model has
the ability to discern changes in the blade-path temperature as a consequence of issues/operations
other than the fuel flow. The sensitivity though depends on the signal-to-noise ratio.

Error Quantification

For comparison, I list here in Table 6.1 the predicted RMSE from data-based methods,
reported in the recent literature. I also estimated the range of the predicted variables and the
approximated scaled RMSE, and in Table 6.1, I am comparing them with this study.

Using GAM-ARFIMA-GARCH, I got RMSPE error of 0.0020 for temperature. As can be

seen from the table that the RMSPE achieved using GARCH method is competitive to the state-of-
the-art methods. I remark here that I only used one predictor in this analysis, that is the fuel flow.
Moreover, I obtained the pre-whitened residuals zt ’s, which are important inputs for the EWMA
control chart to perform anomaly detection and to monitor the operation of the machine.

76
 Table 6.1: RMSE comparison

Author Year Variable Reported Approximate Reported Method Used

RMSE scale- Response
RMSE Range
Bai et. 2020 Temperature 1.1 0.0012 780-850 K Neural Network
al. [31]
Liu et. 2018 Temperature 1.329 0.0016 750-920 K Support Vector
al. [65] Machine
Kiakojoori 2015 Temperature 2.358 0.0025 1400-1425 K Neural Network
et al. [37]
Olsson [66] 2021 Power 50.32 0.0333 2-3.5 kW Regression
Tufekci 2014 Power 3.787 0.0095 440-480 MW Bagging Tree
[67]
Saeed [66] 2020 Efficiency 0.126 0.0013 0-100 Neural Network
Bettocchi 2007 Efficiency .007 0.0070 0-1 Neural Network
[32]
This study 2021 Temperature - 0.0020 - GAM-ARFIMA-
GARCH

Future Work

In this work, I predicted the temperatures individually. An important future work is to con-
sider the multi-variate time-series modeling that captures the spatio-temporal dependence among
the measurements of temperature from local sensors. Also, in this study, I included four impor-
tant input variables which were selected based on physics-based knowledge. Since the model was
additive, I did not consider their interactions. With another group LASSO regularization [68, 69],
interactions may be included as well as other inputs in the generalized linear model, and the im-
portant variables may be selected by data. This is also left for future study.
The dataset available in this study was limited in scope but I expect that if the application of
the algorithm is tried using combustion chamber parameters, and all the circumferential blade path
temperature measurements are considered to predict spread rather than relying on maximum and
average value of temperature, good accuracy in prediction can be achieved. These desired variables

77
 Figure 6.1: Components of Digital Twin with System Genome

were not available in the present dataset but a lot more appropriate labeled data is expected from
industries, which is more appropriate for solving the problem at hand. It will be good if the distri-
bution of blade path temperature can be correlated with some circumferential varying combustion
chamber parameters but on the basis of the current study itself I can not say this conclusively. By
looking at the enhanced data, we can attain better understanding about whether additional sensors
in the gas turbine can help in future machine learning algorithms.
As I presented one approach of data analytics in this study and discussed the potential
limitations, I believe that a multi-faceted approach will be necessary to have a robust “digital twin”
where failures can be anticipated in advance so that corrective actions can be taken. In other words,
pure data-driven approach, referred to as “AD” in Figure 6.1, may not be enough. A low-order,
physics based model, referred to as “DSM” in Figure 6.1, must be operated in tandem based on
the latest system parameters, in order to enhance and interpret the findings from the data-driven
process.
As this continuous process of comparisons, validations and verifications continues, unex-

78
plainable discrepancy between DSM and ADL, will require in-depth localized simulation (referred
to as “LDS” in Figure 6.1). Such simulations can only be run off-line but these can help in iden-
tifying the root causes of the discrepancies between ADL and DSM. The results of LDS may lead
to addition of critical sensors at strategic locations to the sensor network, referred to as “DSN”.
As the power plant continues to operate, there will be degradations in system parameters
used in ADL, and hence the data set that defines the system parameters and referred to as System
Genome in Figure 6.1, must be continuously updated, either through measurement or more often
through parameter estimation applied to data from sensors.
All five components, DSM, DSN, ADL, LDS together with System Genome, must be in-
tegrated together for a robust “Digital Twin”, that cannot be just based on ADL or data driven
processes.

79
 LIST OF REFERENCES

[1] EIA, 2021. International energy outlook 2021. Tech. rep., U.S. Energy Information Admin-
istration, Department of Energy.

[2] Gülen, S. C., 2020. Gas Turbine Combined Cycle Power Plants. CRC Press.

[3] Zhao, N., Wen, X., and Li, S., 2016. “A review on gas turbine anomaly detection for im-
plementing health management”. In ASME Turbo Expo 2016: Turbomachinery Technical
Conference and Exposition, American Society of Mechanical Engineers.

[4] IATA, 2018. Airline maintenance cost executive commentary. Tech. rep., MCTF, IATA.

[5] Marinai, L., 2004. “Gas-path diagnostics and prognostics for aero-engines using fuzzy logic
and time series analysis”. PhD thesis, Cranfield university, July, 2004.

[6] EIA, 2021. Annual energy outlook 2021. Tech. rep., U.S. Energy Information Administration,
Department of Energy.

[7] BP, 2021. Statistical review of world energy. Tech. rep., BP plc.

[8] Gülen, S. C., 2019. Gas Turbines for Electric Power Generation. Cambridge University
Press.

[9] Kumar, N., Besuner, P., Lefton, S., Agan, D., and Hilleman, D., 2012. Power plant cycling
costs. Tech. rep., Intertek APTECH.

[10] EIA, 2020. Capital cost and performance characteristic estimates for utility scale electric
power generating technologies. Tech. rep., U.S. Energy Information Administration, Depart-
ment of Energy.

80
[11] Bechini, G., 2007. “Performance diagnostics and measurement selection for on-line moni-
toring of gas turbine engines”. PhD thesis, Cranfield university, December, 2007.

[12] Lee, J., Wu, F., Zhao, W., Ghaffari, M., Liao, L., and Siegel, D., 2014. “Prognostics and health
management design for rotary machinery systems—reviews, methodology and applications”.
Mechanical Systems and Signal Processing, 42(1), pp. 314–334.

[13] Li, Y. G., 2002. “Performance-analysis-based gas turbine diagnostics: a review”. In Pro-
ceedings of the Institution of Mechanical Engineers, Part A: Journal of Power and Energy,
Volume 216, Issue 5, Institution of Mechanical Engineers.

[14] Marinai, L., Probert, D., and Singh, R., 2004. “Prospects for aero gas-turbine diagnostics: a
review”. Applied Energy, 79.

[15] Loboda, I., 2010. Gas Turbine Condition Monitoring and Diagnostics. IntechOpen.

[16] Romesis, C., 2013. “Condition-based maintenance for gas turbine plants”.

[17] Stamatis, A. G., 2013. Engine Condition Monitoring and Diagnostics. IntechOpen.

[18] Volponi, A. J., 2014. “Gas turbine engine health management: Past, present, and future
trends”. In Journal of Engineering for Gas Turbines and Power, Volume 136, Issue 5, Amer-
ican Society of Mechanical Engineers.

[19] Tahan, M., Tsoutsanis, E., Muhammad, M., and Karim, Z. A. A., 2017. “Performance-
based health monitoring, diagnostics and prognostics for condition-based maintenance of gas
turbines: A review”. Applied Energy, 198, pp. 122–144.

[20] Fentaye, A., Baheta, A., Gilani, S., and Kyprianidis, K., 2019. “A review on gas turbine
gas-path diagnostics: State-of-the-art methods, challenges and opportunities”. Aerospace,
6(7).

81
[21] Stamatis, A., Mathioudakis, K., and Papailiou, K., 1992. “Optimal measurement and health
index selection for gas turbine performance status and fault diagnosis”. Journal of Engineer-
ing for Gas Turbines and Power, 114.

[22] Urban, L. A., 1972. “Gas path analysis applied to turbine engine condition monitoring”.
Journal of Aircraft, 10.

[23] Ogaji, S., Sampath, S., Singh, R., and Probert, S., 2002. “Parameter selection for diagnosing
a gas-turbine’s performance-deterioration”. Applied Energy, 73.

[24] Ogaji, S. O. T., 2003. “Advanced gas-path fault diagnostics for stationary gas turbines”. PhD
thesis, Cranfield university.

[25] Mathioudakis, K., and Kamboukos, P., 2006. “Assessment of the effectiveness of gas path
diagnostic schemes”. Journal of Engineering for Gas Turbines and Power.

[26] Y.G. Li, P. N., 2009. “Gas turbine performance prognostic for condition-based maintenance”.
Applied Energy.

[27] Dragomir, O. E., Gouriveau, R., Dragomir, F., Minca, E., and Zerhouni, N., 2009. “Review
of prognostic problem in condition-based maintenance”. In European Control Conference
(ECC), IEEE.

[28] Romesis, C., Aretakis, N., Roumeliotis, I., Alexiou, A., Tsalavoutas, A., Stamatis, A., and
Mathioudakis, K., 2014. “Experiance with condition-based maintenance related methods
and tools for gas turbines”. In In Proceedings of the Future of Gas Turbine Technology 7th
International Gas Turbine Conference.

[29] Aretakis, N., Roumeliotis, I., Alexiou, A., Romesis, C., and Mathioudakis, K., 2015. “Tur-
bofan engine health assessment from flight data”. Journal of Engineering for Gas Turbines
and Power, 137.

82
[30] Verbist, M., 2017. “Gas path analysis for enhanced aero-engine condition monitoring and
maintenance”. PhD thesis, Delft University of Technology.

[31] Bai, M., Liu, J., Chai, J., Zhao, X., and Yu, D., 2020. “Anomaly detection of gas turbines
based on normal pattern extraction”. Applied Thermal Engineering, 166, p. 114664.

[32] Bettocchi, R., Pinelli, M., Spina, P. R., and Venturini, M., 2007. “Artificial intelligence for
the diagnostics of gas turbines— part i: Neural network approach”. Journal of Engineering
for Gas Turbines and Power, 129, pp. 711–719.

[33] Loboda, I., 2016. Neural Networks for Gas Turbine Diagnosis. IntechOpen.

[34] Joshua Cory, B. B., and Cross, C., 2007. “Autoregression-based turbine engine anomaly
detection”. In Joint Propulsion Conference & Exhibit, American Institute of Aeronautics and
Astronautics.

[35] Kumar, A., Srivastava, A., Goel, N., and McMaster, J., 2015. “Exhaust gas temperature data
prediction by autoregressive models”. In 2015 IEEE 28th Canadian Conference on Electrical
and Computer Engineering (CCECE), pp. 976–981.

[36] Min chen, Liang Quan Hu, H. T., 2015. “An approach for optimal measurements selection
on gas turbine engine fault diagnosis”. Journal of Engineering for Gas Turbines and Power.

[37] S. Kiakojoori, K. K., 2016. “Dynamic neural networks for gas turbine engine degradation
prediction, health monitoring and prognosis”. Neural Computing and Applications.

[38] S. Borguet, O. Leonard, P. D., 2016. “Regression-based modeling of a fleet of gas turbine
engines for performance trending”. Journal of Engineering for Gas Turbines and Power.

[39] Jinfu, L., Jiao, L., Jie, W., Zhongqi, W., and Daren, Y., 2017. “Early fault detection of hot
components in gas turbines”. Journal of Engineering for Gas Turbines and Power, 139.

83
[40] JinFu, L., LinHai, Z., YuJia, M., Jiao1, L., WeiXing, Z., and DaRen, Y., 2018. “Anomaly
detection of hot components in gas turbine based on frequent pattern extraction”. Sci China
Tech Sci.

[41] Rahmoune, M. B., Hafaifa, A., Kouzou, A., Chen, X., and Chaibet, A., 2021. “Gas turbine
monitoring using neural network dynamic nonlinear autoregressive with external exogenous
input modelling”. Mathematics and Computers in Simulation, 179, pp. 23–47.

[42] Basora, L., Olive, X., and Dubot, T., 2019. “Recent advances in anomaly detection methods
applied to aviation”. Aerospace, 6(11).

[43] Volponi, A. J., 1998. “Gas turbine parameter corrections”. In International Gas Turbine &
Aeroengine Congress & Exhibition, American Society of Mechanical Engineers.

[44] Lütkepohl, H., 2005. New introduction to multiple time series analysis. Springer Science &
Business Media.

[45] Kalman, R. E., 1960. “A new approach to linear filtering and prediction problems”. Journal
of basic Engineering, 82(1), pp. 35–45.

[46] Shumway, R. H., and Stoffer, D. S., 2017. Time series analysis and its applications: with R
examples. Springer.

[47] Hastie, T., Tibshirani, R., and Friedman, J., 2009. The Elements of Statistical Learning: Data
Mining, Inference, and Prediction, Second Edition. Springer Series in Statistics. Springer
New York.

[48] Breiman, L., 2001. “Random forests”. Machine learning, 45(1), pp. 5–32.

[49] Wood, S., 2017. Generalized Additive Models: An Introduction with R, 2 ed. Chapman and
Hall/CRC.

84
[50] Ripley, B. D., 2007. Pattern recognition and neural networks. Cambridge university press.

[51] Engle, R. F., 1982. “Autoregressive conditional heteroscedasticity with estimates of the
variance of united kingdom inflation”. Econometrica: Journal of the Econometric Society,
pp. 987–1007.

[52] Bollerslev, T., 1986. “Generalized autoregressive conditional heteroskedasticity”. Journal of

econometrics, 31(3), pp. 307–327.

[53] Granger, C. W., and Joyeux, R., 1980. “An introduction to long-memory time series models
and fractional differencing”. Journal of time series analysis, 1(1), pp. 15–29.

[54] Hosking, J. R. M., 1981. “Fractional differencing”. Biometrika, 68(1), 04, pp. 165–176.

[55] Wood, S. N., 2011. “Fast stable restricted maximum likelihood and marginal likelihood
estimation of semiparametric generalized linear models”. Journal of the Royal Statistical
Society: Series B (Statistical Methodology), 73(1), pp. 3–36.

[56] Ghalanos, A., 2020. rugarch: Univariate GARCH models. R package version 1.4-4.

[57] Hastie, T., 2015. gam: Generalized Additive Models. R package version 1.12.

[58] Visser, I., Speekenbrink, M., et al., 2010. “depmixs4: an r package for hidden markov mod-
els”. Journal of Statistical Software, 36(7), pp. 1–21.

[59] Wood, S. N., 2004. “Stable and efficient multiple smoothing parameter estimation for gener-
alized additive models”. Journal of the American Statistical Association, 99(467), pp. 673–
686.

[60] Liaw, A., and Wiener, M., 2002. “Classification and regression by randomforest”. R News,
2(3), pp. 18–22.

85
[61] Allaire, J., and Chollet, F., 2019. keras: R Interface to ’Keras’. R package version 2.2.5.0.

[62] Nychka, D., 1988. “Bayesian confidence intervals for smoothing splines”. Journal of the
American Statistical Association, 83(404), pp. 1134–1143.

[63] Marra, G., and Wood, S. N., 2012. “Coverage properties of confidence intervals for general-
ized additive model components”. Scandinavian Journal of Statistics, 39(1), pp. 53–74.

[64] Chandola, V., Banerjee, A., and Kumar, V., 2009. “Anomaly detection: A survey”. ACM
Computing Surveys (CSUR), 41.

[65] JinFu, L., LinHai, Z., YuJia, M., Jiao, L., WeiXing, Z., and DaRen, Y., 2018. “Anomaly
detection of hot components in gas turbine based on frequent pattern extraction”. Science
China Technological Sciences, 61, p. 567–586.

[66] Olsson, T., Ramentol, E., Rahman, M., Oostveen, M., and Kyprianidis, K., 2021. “A data-
driven approach for predicting long-term degradation of a fleet of micro gas turbines”. Energy
and AI, 4, p. 100064.

[67] Tüfekci, P., 2014. “Prediction of full load electrical power output of a base load operated
combined cycle power plant using machine learning methods”. International Journal of
Electrical Power & Energy Systems, 60, pp. 126–140.

[68] Ravikumar, P., Lafferty, J., Liu, H., and Wasserman, L., 2009. “Sparse additive models”.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 71(5), pp. 1009–
1030.

[69] Huang, J., Horowitz, J. L., and Wei, F., 2010. “Variable selection in nonparametric additive
models”. Annals of statistics, 38(4), p. 2282.

86