A Review On Gas Turbine Gas-Path Diagnostics: State-Of-The-Art Methods, Challenges and Opportunities

1 Review
2 A Review on Gas Turbine Gas-path Diagnostics:

3 State-of-the-art Methods, Challenges and
4 Opportunities
5 Amare D. Fentaye, Aklilu T. Baheta, Syed I. Gilani and Konstantinos G. Kyprianidis†,*
6 
Mechanical Engineering Department, Universiti Teknologi PETRONAS, 32610 Tronoh, Malaysia
7 ([email protected] (ADF); [email protected] (ATB); [email protected] (SIG))
8 † School of Business, Society and Engineering, Mälardalen University, 883, SE-72123 Västerås, Sweden
9 * Correspondence: [email protected]
10 Abstract: Gas-path diagnostics is an essential part of gas turbine (GT) condition based maintenance
11 (CBM). There exists an extensive literature on GT gas-path diagnostics and verity of methods have
12 been introduced. The fundamental limitations of the conventional methods such as the inability to
13 deal with the nonlinear engine behavior, measurement uncertainty, simultaneous faults, and
14 available limited number of sensors remain the driving force for exploring more advanced
15 techniques. This review aims to provide a critical survey of the existing literature produced in the
16 area over the past few decades. In the first section, the issue of GT degradation is addressed aiming
17 to identify the type of physical faults that degrade a gas turbine performance, which gas-path faults
18 more significantly contribute to the overall performance loss, and which specific components often
19 encounter those faults. A brief overview is then given about the inconsistencies in the literature on
20 gas-path diagnostics followed by a discussion of the various challenges against a successful gas-
21 path diagnostics and the major desirable characteristics that an advanced fault diagnostic technique
22 should ideally possess. At this point, the available fault diagnostic methods are thoroughly
23 reviewed and their strengths and weaknesses summarized. Artificial intelligence (AI) based and
24 hybrid diagnostic methods have received a great deal of attention due to their promising potentials
25 to address the above-mentioned limitations along with providing accurate diagnostic results.
26 Moreover, the available validation techniques that system developers used in the past to evaluate
27 the performance of their proposed diagnostic algorithms are discussed. Finally, concluding remarks
28 and recommendations for further investigations are provided.
29 Keywords: Gas turbine performance; gas-path diagnostics; condition-based maintenance; fault

30 diagnostic methods; diagnostic method validation
31
32 1. Introduction
33 In today’s competitive business world, one way to increase profitability of a machinery
34 equipment or a process plant is reducing its operational and maintenance expenses while increasing
35 productivity. GT is one of the most expensive devices in aircraft and industrial applications, where
36 reliability and availability are the two most desirable attributes. In the past several decades, trillions
37 of dollars invested globally in GTs operation and maintenance [1, 2]. However, due to their rising
38 roles in the fast-growing industry, the market trend is still expected to be continued into the
39 foreseeable future. According to the International Air Transport Association (IATA) report, in 2014,
40 the world fleet count was 24,597 aircrafts. In this fiscal year, globally, airlines spent $62.1 billion on
41 Maintenance, Repair, and Overhaul (MRO), of which about 40% was for engine maintenance. In 2024,
42 the engines MRO is expected to reach over $36 billion, with a 3.8% increasing rate per annum [3]. One
43 can see how large these expenses would be if they are extended to include all types of GT applications.
Aerospace 2018, 5, x; doi: FOR PEER REVIEW www.mdpi.com/journal/aerospace
Aerospace 2018, 5, x FOR PEER REVIEW 2 of 50
44 Studies on GT market indicated that the market for other engine groups is much bigger than the
45 aircraft engines due to the rapid industrialization across the globe and the rising demand for power
46 generation, mechanical drives and propulsion [2, 4-6].
47 The GT fuel consumption and the likely increase in fuel price is another critical issue. For
48 example, the US Department of Defense (DOD) alone consumes 4.6 billion gallons of fuel annually,
49 which is 93% of the US government fuel consumption and the 34th largest fuel consumption in the
50 world, of which about 85% is for Air Force and Navy uses [7, 8]. On the other hand, in combined
51 cycle power plants (CCPPs), the fuel cost covers 75% of the total life-cycle cost (LCC) [6]. Therefore,
52 operating the GT as close to its clean conditions as possible might have a significant contribution in
53 reducing the engine operating expenses. This can be achieved via an improved maintenance policy
54 assisted by more advanced engine health monitoring (EHM) systems [9].
55 The gas turbine maintenance and operation costs are highly influenced by the performance of
56 the engine. Engine overall performance relies on the performance of the gas-path components
57 (mainly the compressor(s) and turbine(s)) and these components are the major problem areas due to
58 their exposure to different internal and external degradation causes [10]. Some of the major and the
59 most likely existing problems are drop in compressor efficiency due to fouling or erosion or an object
60 damage, loss in turbine efficiency due to blade erosion and blade Crip with subsequent tip of probe
61 and shroud damage, decrease in air flow capacity due to fouling, and an increase in flow capacity
62 due to turbine erosion. However, these faults are not directly measurable. The gas-path diagnostic
63 technology thus analyses the engine performance and identifies potential faults and provide an early
64 warning before they developed into more complex problems. An effective and reliable gas-path
65 diagnostic tool that could detect, isolate, and assess potential problems, based on the measurement
66 deviations, and suggest solutions well before they developed into more complex problems is
67 therefore very essential. This plays a major role in the investment by ensuring high-levels of GT
68 reliability and availability along with its best operating performance. There have been variety of gas
69 path diagnostic methods introduced so far beginning with the traditional model-based (MB) methods
70 (such as Kalman Filter (KF) and Gas Path Analysis (GPA)) to the most advanced artificial intelligence
71 (AI) based ones (such as Artificial Neural Network (ANN), Fuzzy logic (FL), Bayesian belief network
72 (BBN) and Genetic Algorithm (GA)) [9, 11]. In recent years, attention has been paid to hybrid
73 methods [12].
74 This paper aims to discuss the main gas-path faults that influence the GT performance, the
75 available challenges on an effective fault diagnostic system development that researchers of this field
76 were trying to deal with so far and some of the most desirable attributes that an advanced system
77 should ideally possess. The available MB, AI based, and hybrid methods are thoroughly reviewed
78 and their advantages and disadvantages regarding how effectively perform the diagnostic tasks,
79 undertake the challenges, and fulfill the desirable attributes are highlighted. Finally, some of the most
80 commonly used diagnostic method validation approaches are discussed followed by conclusions and
81 future research directions.
82 2. Gas Turbine Performance Degradation

83 GT performance can be degraded temporarily or permanently. The former can be partially
84 recovered during operation and engine overhaul while the latter requires replacement [13]. Fouling,
85 erosion, corrosion, and blade tip clearance are among temporary degradation causes. Whereas, airfoil
86 distortion and untwist and platform distortions lead to permanent deterioration (meaning that
87 residual deterioration exists even after a major overhaul). Deterioration can also be categorized as
88 recoverable (with washing), non-recoverable (cannot be recovered by washing during operation but
89 recoverable during overhaul), and permanent (recoverable neither by washing nor during overhaul)
90 [14]. Related to the service period of the engine or the evolution time frame of the deterioration,
91 performance deterioration can also be classified into short-term/rapid and long-term/gradual
92 deterioration [15]. Short-term/rapid deterioration happens at the early age of the GT engine as it starts
93 its operation or may be the result of a single event like an object damage at any time during engine’s
94 operation. Whereas, long-term deterioration is formed more gradually due to the ingestion and
95 accumulation of different contaminants and/or high operating temperature.
96 As shown in Figure 1, these physical faults cause changes in one or more of the independent
97 parameters (performance parameters) which describe individual gas-path component’s performance.
98 The performance parameters generally include compressor flow capacity, compressor isentropic
99 efficiency, turbine flow capacity, and turbine isentropic efficiency. Changes in the performance
100 parameters cause consequent changes in the measurement parameters (temperature, pressure, shaft
101 speed, and fuel flow), which are the fault indicators or symptoms in engine health monitoring.
102
103 Figure 1. GT physical faults, components’ characteristics, and measurements (adapted from [16]).
104 2.1. Fouling

105 Fouling is the adherence of different contaminants (such as sand, dust, dirt, ash, oil droplets,
106 water mists, hydrocarbons and industrial chemicals) on the surface of gas-path components [17, 18].
107 It leads to an increase in surface roughness and a change in airfoil shapes [19]. The end result is
108 performance deterioration. Compressor fouling causes a decrease in flow capacity and isentropic
109 efficiency [20]. However, as shown in Table 1, there is no consensus on the magnitude of the
110 percentage deviation of those parameters. For instance, according to Saravanamuttoo and
111 Lakshminarasimha [21], compressor fouling may result in a 5% loss in flow capacity and a 2.5% loss
112 in isentropic efficiency. Based on site test data, Diakunchak [18] reported a compressor fouling with
113 5% flow capacity and 1.8% isentropic efficiency reduction. On another study, it has been reported
114 that the change in flow capacity due to compressor fouling is equal to 1.25 times the associated change
115 in efficiency [13]. On the other hand, model simulation results reported by Aretakis et al. [22] showed
116 that flow capacity deviation by 3.1% reduced the isentropic efficiency by 0.906%. However, all studies
117 agreed that fouling influences more the flow capacity than the efficiency.
118 Table 1. Compressor fouling and its consequences according to different studies.
Compressor fouling consequences Ref.

ΓC ↓ by 5%, ηC ↓ by 2.5%, and power output ↓ by 10% [21,23]
ΓC ↓ by 5%, ηC by 1.8 %, power output ↓ by 7%, and heat rate ↑ by 2.5% [18]
A 1% reduction in Γc resulted in a 0.8% ηc reduction [13]
ΓC ↓ by 3.1% and ηC ↓ by 0.906% [22]
Power output reduces between 2% (under favorable conditions) and 15 to 20% (under adverse [24]
conditions)
ΓC ↓ by 5%, fuel consumption ↑ by 2.5%, and power output ↓ by 8% [25]
119 Compressor fouling is responsible for 70 to 85% of the total performance loss of a GT [18].
120 According to Diakunchak [18], a 5% flow capacity and a 1.8% isentropic efficiency reduction due to
121 compressor fouling, could result a 7% loss in power output and a 2.5% increase in heat rate. Whereas,
122 Lakshminarasimha et al. [26] reported that a 10% reduction in power output could result in a 5%
123 mass flow rate and a 2.5% efficiency reductions due to compressor fouling. This result agreed with
124 the result in [21]. According to Meher-Homji and Bromley [24], compressor fouling could result loss
125 of power output as high as 20% under adverse conditions. These changes are immediately corrected
126 by increasing the fuel consumption through the automatic engine control system. A 2.5% increase in
127 fuel consumption due to a 5% flow capacity reduction was reported by Zwebek and Pilidis [25].
128 Compressor fouling could also decrease blade tip clearance [27] and surge margin [28] and increase
129 turbine entry temperature (TET) [29].
130 Different studies on multistage axial compressor fouling declared that only the first few stages
131 are subjected to fouling and level of fouling is not uniform at different stages [30, 31]. Experiment
132 based studies on a 16 -stage axial compressor [32] showed that the number of stages affected by the
133 fouling reaches 5 to 6 and the degree of fouling diminishes from the suction end to the delivery end.
134 A similar study by Aker and Saravanamuttoo [30] revealed that the first 40-50% of stages of a 16-
135 stage axial compressor is exposed to fouling. Although the first few stages of the axial compressor
136 are subjected to the highest amount of foulant, during compressor washing the deposit moves to the
137 rear end stages and accumulates, and thereby influences the power output [33]. The degree of
138 compressor fouling and the extent of its impact on engine component’s performance depends on
139 several factors including the number of stages, surface roughness, airfoil loading, and the
140 contaminant nature [34].
141 Fouling based performance deterioration can be reversed by compressor washing using water
142 and/or detergents [24]. There are two types of compressor washing, namely, online and offline [35].
143 The former is performed during operation, while the latter needs to shut down and cool the GT. These
144 washing regimes are discussed in detail in [36]. Although the initial stage of fouling deposit doesn’t
145 cause an immediate degradation, once it has been accumulated, the deposit removal task is time
146 taking and costly [37]. Online washing is important to minimize the foulant deposit and reduce the
147 frequency of offline washing. The online washing alone is not effective to completely remove fouling,
148 while the offline scheme is capable. The frequency of both online and offline washing and the
149 duration between them depends on the operating condition of the engine [38]. The washing process
150 should be assisted by an optimized schedule taking into account economic and safety issues [39]. This
151 is because a frequent washing increases downtime and maintenance cost and sometimes it may also
152 lead to a premature blade surface erosion. On the other hand, a long duration may cause an
153 incomplete performance recovery. Fouling based performance deterioration is mostly recoverable if
154 the offline washing is performed when the reduction in compressor flow capacity reaches about 2-
155 3% [40].
156 2.2. Erosion

157 Erosion is the gradual loss of materials from the surface of gas-path components caused by the
158 ingestion of contaminants such as sand, dust, dirt, ash, carbon particles, and water droplets [41].
159 Among these causes, sand is the most common due to its occurrence at most of the GT application
160 areas. The particulates that are causing erosion are usually 20μm or more in diameter [18]. Erosion
161 can attack all the gas-path components although the degree of influence is higher for turbines than
162 compressors. It can result an overall performance loss of about 5% [42]. Like fouling, performance
163 deterioration subject to erosion can be represented by flow capacity and isentropic efficiency changes.
164 Efficiency decreases during both compressor and turbine erosions because of an increase in blade
165 surface roughness and tip clearance and changes in airfoil profile. Whereas, flow capacity decreases
166 upon compressor erosion and increases upon turbine erosion [43]. According to Refs. [44], the ratio
167 of change of flow capacity to efficiency is 2:1. The effect of erosion is less for industrial GTs than
168 aircraft engines due to the presence of a more effective air filtration system [45].
169 2.3. Corrosion

170 Corrosion is an irreversible deterioration of components as a result of oxidation reaction or
171 chemical interaction with inlet air contaminants (sodium and potassium salts, mineral acids and other
172 chemically reactive elements including sodium, potassium, lead, and vanadium) and combustion
173 gases (for instance sulfur oxides) [46, 47]. It can be classified as cold and hot corrosion [48]. The
174 corrosion due to airborne contaminants in combination with water is called cold or wet corrosion and
175 affects especially of the compressor airfoils [47]. The hot corrosion occurs due to combustion gases
176 containing certain contaminants and/or molten salts, which affects especially of the turbines [49].
177 Corrosion due to hot gas contaminants is more severe and highly influenced by the gas temperature
178 [46]. Salt is the main cause of corrosion in both compressor and turbine components [50]. It decreases
179 compressor flow capacity, compressor isentropic efficiency, and turbine isentropic efficiency and
180 increases turbine flow capacity [51]. Corrosion effects can be prevented by a proper coating [41].
181 2.4. Foreign Object Damage (FOD)/Domestic Object damage (DOD)

182 Gas-path components are subjected to damage due to the foreign objects being injected into the
183 engine (such as birds or any other wildlife, stones, frost, snow, ice, and runway gravel) or domestic
184 objects (broken out engine parts like blade sections or large carbon particles from the fuel nozzles).
185 FOD is one of the most common problems, usually in aircraft engines [13]. The damage from foreign
186 objects varies from a non-recoverable deterioration to a catastrophic failure, as in the case of blade off
187 or large object ingestion in the engine [18]. It shows a rapid shift in the gas-path measurements. In
188 addition, engine vibration may come from unbalanced material loss or aerodynamic excitation from
189 blade distortion due to FOD [13]. FOD highly influences the components isentropic efficiency than
190 flow capacity due to its impact on the blade surface roughness and distortion [51]. The magnitude of
191 the loss depends on the type and nature of the FOD/DOD. If the damage causes a material loss on the
192 blade surface, the flow capacity will increase, or if the foreign object blocking of the gas-path the
193 opposite will be experienced [52].
194 2.5. Increase in Blade tip clearance

195 It refers to an increase in the clearance between moving blades’ tips and the casing or stationary
196 blades’ tips and the rotating hub due to the removal of materials caused by particulate ingestion,
197 thermal and centrifugal expansion, and erosion [13, 14]. It can also be caused by rotor assembly
198 vibration due to over speed during the starting cycle [18] or the rubs between the stator assembly and
199 rotor assembly due to thermal and centrifugal expansions [53]. It causes a non-recoverable
200 performance deterioration. The increase in clearances will increase the leakage and thereby a
201 performance deterioration [54]. The performance deterioration due to this fault can be represented
202 by efficiency and flow capacity reductions [55]. For example, a 0.6% efficiency reduction is reported
203 in [56] due to a radial clearance increase of 0.3% and 0.4% from the tips of rotors and stators,
204 respectively. In another study, it has been reported that an increase in tip clearance by 0.8% could
205 result up to a 3% and 2% reduction in flow capacity and isentropic efficiency, respectively [57].
206 According to Diakunchak [40], a 1% increase in blade tip clearance would lead to over 1% loss in
207 power output and overall efficiency. A 1% to 3.5% increase in blade tip clearance would also cause
208 up to 15% drop in the stage pressure ration as reported by Kurz and Brun [46]. Table 2 summarizes
209 contaminant types and their effects on the physical and thermodynamic characteristics of the gas-
210 path components of GTs.
211
212 Table 2. Summary of GT degradation causes, effects, and component performance change indicators.
Physical Contaminant/Cause Exposed Effect Performance Results References

fault component(s) change indication
Fouling Dust, dirt, sand, rust, ash, Compressor & - Increase in surface roughness - ↓Γ - Loss of power [17,19,38,58]
carbon particles, oil, unburned turbine - Changes in airfoil shape - ↓PR output/trust
hydrocarbons, soot, chemicals, - Increase the airfoil angle of attack - ↓η - ↑heating rate and EGT
fertilizers, herbicides fuel, etc. - Disrupt rotating balance
- Obstruct and plug flow path
Erosion Dirt, sand, dust, ash, carbon Compressor & - Airfoil profile changes - ↓ Compressor Γ - Loss of power [18,19,59]
particles, etc. turbine - Blade tip and seal clearances - ↑ Turbine Γ output/trust
increase - ↓Compressor PR - ↑heating rate and EGT
- Surface roughness increases - ↓η C&T
- Reduce the compressor and
turbine cross-sectional areas
Corrosion Salts, acids, nitrates, sulfates, Compressor & - Increase in blade surface - ↓ Comp. Γ & η - Loss of power [18,46,51,59]
etc. Turbine roughness - ↑ Turb. Γ & ↓η output/trust
- Alters blade profile change - ↑heating rate and EGT
Blade tip Rubs between rotor and stator Compressor & - Increased leakage - ↓η and Γ - Loss of power [40,53]
clearance blades caused by thermal turbine - Vibration - ↓ surge margin output/trust
expansion, FOD and erosion - Chock at lower flow - ↑heating rate and EGT
FOD/DOD Hailstones, runway gravel or Compressor - Increase in blade surface - ↓ η (C+T) - Loss of power [58,59]
birds, large carbon particles and turbine roughness - ↑/↓ Γ output/trust
- Removal of parts from blade - ↓PR - ↑heating rate and EGT
surfaces
213
214
215 3. Fault Diagnostics

216 There is inconsistency in the literature on the terminology and definition of fault diagnostics.
217 Some of the commonly used terminologies are fault diagnostics [60, 61], fault detection and isolation
218 (FDI) [62, 63], fault detection and diagnostics (FDD) [64, 65], fault detection, isolation, and
219 identification (FDII) [66], fault detection, isolation and accommodation (FDIA) [67, 68], fault
220 detection, isolation and recovery (FDIR) [69] and identification and fault diagnostics [70]. This makes
221 it difficult to understand the goals of the contributions and to compare the different techniques. For
222 example, the definition of the term ”isolation” in FDI and FDII is different in some papers. In the
223 former case, it refers the process of determining the fault type and location followed by estimating its
224 level whereas in the latter case it doesn’t include the fault level estimation. However, the broader
225 research community including the military and other industry sectors defines fault diagnostics as the
226 procedure of detecting, isolating and identifying an impending or incipient failure condition, during
227 which the affected component is still operational even though at a degraded mode [71]. Each element
228 in the fault diagnostic process further defined as:
229  Fault detection: Detecting the presence of an abnormal behavior, which may gradually lead
230 to the failure of the system or part of it.
231  Fault isolation: Determining the type and location of the fault(s).
232  Fault identification: estimating the magnitude of the fault(s).
233 Figure 2 shows the general conceptual model of performance-analysis-based GT fault
234 diagnostics, adapted from [72]. Usually, a complete fault diagnostics requires three basic activities;
235 data acquisition, data processing, and diagnostics. Each of these phases are equally significant and
236 critical in the attempt to provide a reliable and practically useful decision support mechanism. Data
237 acquisition is the process of collecting and storing the necessary engine performance data for fault
238 diagnosis. The second step, the data processing task, involves two basic activities: data screening and
239 analysis. Data screening is a process of filtering outliers and reducing noises followed by validation,
240 through an appropriate screening technique. This helps to minimize the effect of measurement
241 uncertainties on the fault diagnostic result. Feature extraction starts from baseline establishment that
242 represents a clean condition operation. Since the measurement deviations could be due to load or
243 ambient condition changes, establishing the baseline requires correcting the measurements against
244 these variations so that the deviations due to the actual engine faults and sensor problems can be
245 determined. Regardless of the other effects, the measurement deviations due to performance
246 degradation provide relevant information about the nature of the fault signatures in engine gas-path
247 fault diagnostics. Fault diagnosis is the decision making step, in which algorithms are applied to
248 detect, isolate and identify various faults.
249
250 Figure 2. Conceptual model of a gas-path fault diagnostics (adapted from [72]).
251 Fault detection is the very important step in the process of fault diagnostics. A trend shift
252 detection and binary decision approaches are the two commonly applied techniques [73]. This task
253 is performed based on the difference between the predicted and observed measurements or residuals
254 (Figure 3). Ideally, the residuals should be very close to zero when the engine is clean and deviate
255 noticeably from zero when a fault occurs in the system. But, in reality, due to measurement non-
256 repeatability and model uncertainty, a suitable threshold should be selected, to avoid false alarms.
257 After having an appropriate threshold selection, when the engine is running in a clean condition, all
258 the measurement residuals are expected to lie below the threshold. Conversely, when any kind of
259 abnormal condition occurs, one or more measurement residuals will probably deviate from the
260 selected threshold(s). On the other hand, in the case of the binary decision, the residual is considered
261 as a signal which is zero when the system is functioning properly and different to zero when some
262 abnormal behavior is observed. After a successful fault detection process, the location of the fault and
263 its type should be determined. This process may include separating different sensor faults [74],
264 distinguishing sensor and actual component faults, and classifying different component faults [62].
265 Like the detection, measurement residuals can be used in the isolation process based on proper
266 threshold selection [75] or the fault isolation problem can be treated as a classification problem, as
267 reported in [61, 76, 77]. However, the fault detection and isolation activities does not provide a
268 quantitative information about the health status of the engine. Hence, maintenance decision requires
269 knowing the severity of the deterioration. Usually, component’s isentropic efficiency and flow
270 capacity deviations (health indices) are used to represent the health status of engine gas-path
271 components. Hence, the progressive deviations of these parameters can be estimated using the
272 measurement deviations. The review of the available literature methods will be presented in the
273 method review section.
274
275 Figure 3. A general structure of residual based fault diagnosis procedure.
276 3.1. Challenges of a Successful GT Fault Diagnostics

277 In performance analysis based engine gas-path diagnostics, there are different factors
278 influencing the attempt to obtain a sufficiently accurate and practically useful solutions. The most
279 significant challenges are summarized as follows.
280 1. Nonlinearity of the diagnostic problem. The relationship between dependent parameters
281 (measurements) and independent parameters (performance parameters) is highly non-linear.
282 The complexity of the nonlinearity of the diagnostics problem increases as two or more
283 components are affected simultaneously and/or sensor and component faults exist together.
284 The diagnostic system to be proposed should thus be capable of dealing with the non-linear
285 nature of the engine behavior.
286 2. Measurement uncertainty. In reality, the data obtained from real engine operation cannot be
287 error-free [78]. This error may come from the sensor itself (due to improper installation,
288 miscalibration or malfunctioning), the operating environment, or the operator itself.
289 Measurement uncertainties provide wrong information about the nature of the fault
290 signatures, thereby cause misinterpretation during engine health assessment. Noise and bias
291 are the two categories of measurement uncertainty [79]. Noise is a measurement’s non-
292 repeatability due to the engine harsh operating environments. Whereas bias refers to a sensor
293 fault which is the difference between the average measurement and the actual value defined
294 by the National Bureau of Standards (NBS) [78]. It is a fixed error (can be higher or lower
295 than the actual value) usually occurs as a result of a flaw in the sensor itself. Sometimes, the
296 values of these uncertainties may reach to a level often comparable to the actual measurement
297 deviations caused by component deterioration. If this effect is ignored during the diagnostic
298 method development, the solution will be unrealistic. Conversely, engine fault diagnosis
299 using uncertain measurements may give an erroneous result, particularly, in MB methods.
300 Therefore, either the sensor problem should be treated and corrected prior to the component
301 fault diagnosis or the component fault diagnostic technique should tolerate these effects.
302 3. Availability of limited sensors. GT engines are packed with different sensors for different
303 purposes such as process control, health monitoring, and diagnostics. Measurement
304 parameters which are essential for engine performance analysis are known as standard
305 measurements [80]. For instance, pressure, temperature, fuel flow rate, and spool speed. The
306 deviations of those measurements provide relevant information about the nature and
307 severity of components’ performance deterioration. A careful measurement selection is

308 crucial for an effective fault diagnostics, especially in the case of MB methods. On the one
309 hand an accurate gas-path analysis requires a large number of measurements since the engine
310 model is developed based on several instrumentation suites. In order to satisfy the
311 requirement for a determinate equation, the number of measurements (the dependent
312 parameters) has to be at least equal to the number of performance parameters (the
313 independent parameters). On the contrary, in real engine service, the number of instruments
314 available are limited due to weight and bulk issues (particularly in aircraft and marine
315 applications), sensor noise and bias problems, the need of a reduced sensors’ installation and
316 maintenance cost, and the absence of the gas generator turbine inlet sensors (since they
317 cannot withstand the very high operating temperature) [81, 82]. It is also impractical to
318 measure the air flow rate due to the absence of the technology. Therefore, the diagnostic
319 system is accountable to give the required solution using the available limited information
320 obtained from the minimum sets of measurements.
321 4. Occurrence of multiple faults simultaneously: In harsh engine operating conditions, the
322 occurrence of multiple component/sensor faults is a likelihood. Hence, a single fault
323 assumption can result an untrustworthy fault diagnosis in the presence of multiple faults.
324 The probability of the number of possible fault combinations grows exponentially depending
325 on the available number of engine components/sensors and as a result the complex of the
326 diagnostic problem increases. The performance of a gas-path fault diagnostics scheme highly
327 influenced by the number of simultaneous faults [83]. This is because, when two or more
328 components/sensors are affected together, there is a chance to produce a similar or obscure
329 fault signatures, thereby they can mask or compensate each other’s effects. For example, in
330 the case of double component faults (DCFs), when one of the components is lightly affected,
331 the combined effect may result a confusing pattern with that of a single component fault
332 (SCF). Likewise, if both components are severely affected, they may produce similar patterns
333 with that of a triple component fault (TCF), and as a result, the DCFs may wrongly be
334 classified as TCF or vice versa [83]. In general, as a multiple fault scenario concurrent
335 component faults, concurrent sensor faults, or concurrent sensor and component faults
336 possibly exist during the engine lifetime.
337 5. Operating condition variations. Due to load and/or ambient condition variations, engine
338 operating point may not be fixed. Therefore, operating point changes should be taken into
339 account for practicability. A common way to avoid the influence of operating conditions
340 variations is to form a “baseline” model, compute measurement deviations, and use them as
341 network inputs instead of measurements themselves. Usually, it requires the model of the
342 normal state to figure out the “baseline” [74, 84]. Different GTs have different baseline based
343 on their configuration and application environment. Hence, for a reliable fault diagnosis, an
344 accurate baseline establishment is critical.
345 6. Lack of standards in defining and representing fault diagnostic problems [85]. In the
346 literature there is no consistency in defining and representing GT fault diagnostic problems.
347 Majority of the available methods in the open domain considered different platforms, with
348 different levels of complexity, and applied different performance evaluation metrics. This
349 inconsistency causes difficulties in exchanging diagnostic ideas, information fusion between
350 fault diagnostic results of different engine systems, and a one-to-one comparison of different
351 techniques.
352 7. Unavailability of data in the required type, quality and quantity. Fault diagnostic method
353 developers require relevant and reliable operational data, which can sufficiently represent
354 the healthy and unhealthy engine conditions, to demonstrate and verify new algorithms.
355 However, because of the very limited access to engine operational data (owing to proprietary
356 and liability issues) and lack of deteriorated engine data due to the frequent washing actions
357 it is difficult to obtain the required data [81]. Performance data can be generated by either
358 intentionally ingesting different physical fault causes/contaminants into the operating GT or
359 implanting artificial fault patterns to the engine performance model [86]. The former
360 alternative is not recommended since it is not technically and economically feasible. Whereas
361 the latter, which is the most widely used alternative in this field, requires an accurate model.
362 8. Absence of Diagnostic Methods Validation Techniques: GT users need a practical tool to
363 evaluate the performance and effectiveness of a newly proposed algorithm in order to
364 incorporate to their plant. Up to know, there are no standards to effectively evaluate the
365 technical and economic feasibility of new algorithms [81]. The general procedures used by
366 the research community so far will be presented later in this paper.
367 3.2. Desirable Attributes of a Fault Diagnostic System

368 According to the previous studies on machinery health monitoring and diagnostics including
369 GTs [87-89], an effective fault diagnostic system is ideally expected to fulfill the following
370 characteristics. These desirable attributes could also be used as a selection criteria/standards of
371 various diagnostic approaches.
372 i. Fault diagnostic accuracy: For a correct maintenance decision the fault diagnostics technique
373 should able to detect, isolate and identify gas-path faults successfully. A fault detection task
374 commits two types of errors: false alarms and missed detections. Both the detection errors are
375 equally harmful. A false detection leads to an increased maintenance cost, which is to the
376 opposite of the aim of the fault diagnostics. Conversely, a missed detection may cause a
377 significant performance loss or even system/component failure. Hence, in the detection step,
378 the so-called normal class has to be distinguished from the abnormal class with a reasonably
379 acceptable accuracy. This is very important to avoid unnecessary or unexpected downtimes
380 and enhance reliability. Next to the fault detection, the diagnostic system should successfully
381 determine the fault type and location. In particular, a GT fault isolation algorithm is
382 accountable to separate sensor faults from actual engine component faults followed by
383 classification of different component faults. All the possible single and multiple sensor and/or
384 component fault cases are desired to be isolated correctly using the minimum instrumentation
385 suite. For a final maintenance decision, an accurate fault-level estimation is highly desirable so
386 that the operator could make a strategic maintenance schedule of possible maintenance actions.
387 ii. Robustness: For a practical implementation, diagnostic systems are highly required to be
388 robust/tolerant against measurement uncertainties.
389 iii. Explanation facility: To support engine users in the maintenance decision process, the fault
390 diagnostic tool is required to be able to explain the nature of the faults (i.e., their root cause,
391 current situation, and propagation) and justification of the recommendations.
392 iv. Simplicity/user-friendliness: The method should be simple to use and easy to understand by
393 the operators so that an urgent decision could be made without the presence of any expert. It
394 should thus be capable of providing a user-friendly interface.
395 v. Adaptability: GT performance is sensitive to ambient condition changes or load variations.
396 Therefore, a performance-based GT fault diagnosis system should be able to adapt to those
397 variations so as to maintain its performance.
398 vi. Memory and computational requirements: The storage capacity and computational
399 requirements (computational speed, time, and complexity) are the two basic features of a GT
400 fault diagnosis algorithm, particularly for online applications.
401 vii. Reliability. Concerns about the practicability of the method for an engine with limited
402 numbers of sensors and measurement errors. It should also be simple and cost-effective with
403 minimum downtime for repair and maintenance.
404 viii. Comprehensiveness. This is the measure of the ability of the method to incorporate
405 improvements when it is necessary and to be interfaced with other engine health management
406 systems through data fusion in order to obtain a complete condition based maintenance
407 framework.
408 ix. Flexibility. It measures the degree of capable of the method optimizing its configuration and
409 adapt/extend the system to work on different engines or on the same engine running at
410 different operating conditions. A low set-up time is desirable to implement this feature.
411 4. State-of-the-art: GT Gas-path Diagnostic Methods

412 In the field of GT diagnostics several methods have been devised by engine manufacturers and
413 the research community over the years [90]. As shown in Table 3, different authors categorized those
414 methods into different groups. In the present review, based on the type of information used in
415 modeling, the available methods are categorized into three main groups; model-based (MB), AI-
416 based, and hybrid methods [91]. Accordingly, the state-of-the-art on gas-path diagnostic methods
417 under each group has been undertaken. Different issues related to their working principles,
418 applications for gas-path diagnostics, capability of undertaking the challenges (Section 1.3.1) and
419 fulfilling the desirable attributes (Sections 1.3.2), and their advantages and limitations are reviewed
420 and summarized.
421 Table 3. Categorization of fault diagnostic methods presented in literature.
Author Ref. Year Classification Categories

Dash et al. [88] 2000 MB and Data-driven (DD)
Li [92] 2001 MB, AI based, and Fuzzy logic
Venkatasubramanian et al. [89] 2003 Quantitative, Qualitative, and DB
Ogaji & Singh [5] 2003 Conventional and evolving
Jew [85] 2005 MB, DD, and Hybrid
Jardine et al. [93] 2006 Statistical, MB, and AI based
Stamatis [10] 2014 MB, AI based, and Hybrid
Kong [94] 2014 MB and Soft computing
Zhao et al. [95] 2016 MB, DD, and Knowledge based
Tahan et al. [11] 2017 MB, DD, and Hybrid
422 4.1. Model-based Diagnostic Methods

423 MB diagnostics methods are the first generation GT CBM methods and they rely on the
424 thermodynamic model of the engine. According to this approach, the relationship between the gas-
425 path measurements and the performance parameters is determined by explicit mathematical and
426 thermodynamic equations. GPA and KF are the two most intensively investigated MB methods [92].
427 Engine manufacturers and military sectors have been using these methods since the past four decades
428 [96].
429 4.1.1. Gas-path Analysis

430 A GPA is a mathematical procedure that used to diagnose gas-path components based on the
431 measurement deviations. In this strategy, the diagnostic problem requires the search for a best match
432 between measurement changes and the associated performance parameter changes that cause the
433 measurement changes. According to [97, 98], the thermodynamic relationship between gas-path
434 measurements and components performance parameters can be expressed as:
  
435 Z  h(X, w) (1)

436 where, Z  R M is the measurable parameter vector and M is the number of measurement parameters,

437 X  R N is component performance parameter vector and N is the number of performance

438 parameters, w is the ambient condition and power setting parameter vector (called input vector),
439 and h( ) is a vector valued function determining the relationship between the dependent and
440 independent parameters, usually non-linear.
 M 
441 If sensor noise (vR ) and bias ( b  R M ) are considered:
442
    
443 Z  h(X, w)  v  b (2)
444 4.1.1.1. Linear GPA (LGPA)

445 LGPA is first introduced by Urban [97] upon the assumption of a steady state process with no
446 ambient condition and load variations and negligible measurement uncertainty effects (Eq. 3). The
447 relationship between the dependent and independent parameter changes was assumed to be linear.
448 Mathematically it can be expressed as:
449
 
450 Z  ICM (X) (3)
451

452 where  Z is the vector of measurement deltas, ICM is the so-called influence coefficient matrix, and

453  X is the vector of performance parameter deltas.
454

455 The estimation of  X is a reverse process performed using the inverse of the linear ICM which
456 is referred to as Fault Coefficient Matrix (FCM), as given in Eq. (4).
457
  1

458  X  FCM   Z  H  Z (4)
459
460 The relationship between ICM and FCM in matrix form can be presented as:
461
462
ICM
463 Degradation ∆Z , ∆Z , … ∆Z , … ∆Z ,
∆X , ∆X , … ∆X , … ∆X , ∆Z , ∆Z , … ∆Z , … ∆Z ,
Result in
464 ∆X , ∆X , … ∆X , … ∆X ,
⋮ ⋮ ⋮ ⋮
⋮ ⋮ ⋮ ⋮
465 ∆X , ∆X , … ∆X , … ∆X , ∆Z , ∆Z , … ∆Z , … ∆Z ,
466 ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮
467
Allow
∆X , ∆X , … ∆X , … ∆X , ∆Z , ∆Z , … ∆Z , … ∆Z ,
468 ∆X , ∆X , … ∆X , … ∆X , ∆Z , ∆Z , … ∆Z , … ∆Z ,
469 FCM
470
471
472 Based on the number of dependent and independent parameters, the estimation of FCM will
473 have three different cases [81].
474  Case 1. (When M = N): When the number of measurements and performance parameters are
475 equal, the number of unknowns and equations will be equal, and thereby the problem will be
476 determinable. In this case, the ICM is a square matrix and invertible.
477  Case 2. (When M > N): When the number of measurements is greater than the number of
478 performance parameters to be estimated, the problem will be over-determined. In this case, the
479 solution can be found applying the least square estimation method by replacing H-1 with the so-
480 called pseudo-inverse.
481
 
482 X  (HT  H)H 1HT Z (5)
483
484  Case 3. (When M<N): In the real situation of a GT operation neglecting the effect of sensor noise
485 and bias leads to un-realistic solution. Conversely, considering all these issues including model
486 uncertainty would result undetermined set of equations. The suitable solution for this problem
487 scenario is given by Volponi [81].
488
489 After Urban, LGPA has been studied by several researchers like those in [42, 99-101]. During the
490 early ages of gas-path diagnostics, it was used by engine manufacturers like Rolls-Royce [102]. It has
491 been shown that for deviation values higher than 1%, the LGPA provides an unreliable solution [103].
492 The reliability of this method highly influenced on the accuracy of the ICM, the level of noise and
493 bias, and the number of instrument suite considered [92].
494 4.1.1.2. Non-Linear GPA (NLGPA)

495 In a real GT engine health condition, the assumption of a linear relationship between
496 measurements and performance parameters becomes increasingly unrealistic, especially when the
497 component’s deterioration level exceeds the value assumed for LGPA and/or while the number of
498 gas-path faults increases [9]. The NLGPA scheme is capable of undertaking the nonlinearity of the
499 engine behavior. The thermodynamic relationship between the dependent and independent
500 parameters for a non-linear engine behavior is given as Eq. 6 [81].
501
 
502 Z  H  X (6)
503
504 where:

505   Z is vector of measurement delta and can be expressed as:
506
 Z1 
 Z 
 2 
    
 ( Z Measured  Z Baseline )  
507 Z    100   Z j 
Z Baseline
  
 
Z M 1 
 Z 
 M 
508

509   X is performance parameter delta vector and can be expressed as:
 Component1 
  
 X1   Component1 
 X   Component2 
 2   
     Component 2 
      
510 X   X k  , for example X   
 Compressor k 
  
    
Component k
X
 M 1   
 X    
 N    
 Component N 
 Component N 
511
512  H is the ICM, which determines the relationship between ∆Z⃑ and ∆X⃑. It is the percentage delta in
513 each measurement parameter for the corresponding percentage change in each performance
514 parameter. For an infinitesimal change in the independent parameters, the corresponding ICM is
515 the Jacobian.
516
 Z 1 Z 1 Z 1 Z 1 Z 1 
 X   
X 2 X k X N 1 X N
 1 
 Z Z 2 Z 2 Z 2 Z 2 
2
 
 X 1 X 2 X k X N 1 X N 
        
517  Z Z j Z j Z j Z j 
j
H    
 X 1 X 2 X k X N 1 X N 
        
 Z M 1 Z M 1 Z M 1 Z M 1 Z M 1 
   
 X 1 X 2 X k X N 1 X N 
 Z M Z M Z M Z M Z M 
 X  
 1 X 2 X M X N 1 X N 
518
519 Then, the corresponding performance change can be computed using the equation:
520
 
521  X  H 1   Z (7)
522
523 To consider the non-linear behavior of the engine, an iterative Newton–Raphson method could
524 be applied to the LGPA until the solution converges [101]. This is done by minimizing the error

525 objective function (Eq. 8), which is the difference between the predicted measurement vector ( Z )

526 and the actual measurement vector ( Z ). For the first iteration, a small delta on the component
527 performance is introduced and the corresponding ICM is generated. The FCM is then determined by
528 inverting the ICM. The performance parameter deviation vector is computed by multiplying the FCM
529 with the deteriorated engine measurements. From the calculated results, a new ICM and FCM are
530 generated and the procedure is repeated until the solution converges. The output of the first iteration
531 is the baseline for the second iteration, the output of the second iteration is the baseline for the third
532 iteration and so on, until the last iteration.
 
533 Objective function = (OF)   f  Z j  Z j  (8)
j  
534
535 The convergence of the solution can be evaluated using the error root mean square (RMS) value
536 as given in Eq. 9 [104]. When the RMS value reaches the target value, the iteration will be terminated.
537 The iterative procedure is illustrated in Figure 4.
538
2
M  Z j , predicted  Z j , actual 
  

j 1  Z j , actual 
539 RMS   (9)
M
540
Z  H  X
Y
 ( Z 2 )  ( Z 3 )
 ( Z 1 )
 ( X 1 )  ( X 2 )
X OLD X NEW X
541
542 Figure 4. Schematic illustration of Newton-Raphson based GPA methods (adapted from [105]).
543 The NLGPA approach is introduced by Escher [101]. Since then several diagnostic algorithms
544 with some improvements have been contributed by other authors [9]. Its effectiveness highly
545 influenced by the number and location of measurements on the gas-path. Ogaji et al. [86] used this
546 approach to investigate the effect of measurement selection on engine fault diagnostic accuracy and
547 suggest the best measurement sets corresponding to different fault scenarios. Recently, Li [106]
548 developed a novel GT performance and health status estimation method for a single-shaft aero
549 turbojet engine using adaptive GPA. He used nine gas-path measurements to assess five performance
550 parameters. The test results showed that the proposed method is capable of identifying gas-path
551 faults accurately even in the presence of measurement noise. The diagnostic effectiveness of three
552 different GPA methods have been investigated using different test fault cases for double shaft GT
553 engine by Stamatis [107]. Similarly, the fault diagnostics effectiveness of GPA and AI approaches
554 have been compared and their pros and cons identified based on case studies by Kong [94]. Emil
555 Larsson [108] developed a systematic design procedure to construct non-linear MB fault diagnosis
556 method for industrial GTs. On another study, Jasmani et al. [80], devised a new measurement
557 parameter selection scheme by combining analytical approach and measurement subset concept.
558 Likewise, Chen et al. [109] proposed an approach that can select the optimal number of engine
559 measurements for engine GPA purpose. However, GPA techniques can diagnose GT faults if and
560 only if noise and bias doesn’t exist [94].
561 4.1.2. The Kalman Filter

562 KF is a MB iterative algorithm that uses a set of equations and consecutive data inputs to estimate
563 the true value of the system parameter being measured when the measured values contain a certain
564 amount of uncertainty. It is initially developed by Rudolf Kalman [110], in 1960, and is basically a
565 predictor-corrector technique by which the state of a system is determined at time tk using only the
566 state at previous time step tk-1. The discrete time KF [110] and the continuous time KF [111] are the
567 two types of KF algorithms [112]. The complete KF procedure is composed of two phases; the
568 prediction phase and the correction or measurement update phase. In the prediction phase, the KF
569 produces estimates of the current state variables, along with their uncertainties. Once the outcome of
570 the next measurement is observed, in the correction phase, these estimates are updated using a
571 weighted average, with more weight being given to estimates with higher certainty. Figure 5
572 represents the block diagram of the discrete- time KF method.
573
574
575 Figure 5. Typical KF application block diagram (adapted from [113]).
576 The problem is defined mathematically as follows:

577
578 o System equation: Xk1  k1Xk  Gkuk  wk (10)
579 o Measurement equation: Zk  Hk Xk  vk (11)

580
581 where X ∈ RN is the system state vector, k is the time index, Φ ∈ RNxN is the transition
582 matrix/measurement matrix, u ∈ RM is the control vector, G is the input translation matrix, wk is the
583 system error matrix, Z ∈ RM is the measurement vector at time k, vk ∈ RM is the measurement error
584 (noise) matrix, and Hk ∈ RMxN is the model matrix.

585 The aim of the KF is to estimate the system Xk1 of Xk1 based on prior system knowledge and
586 the available noisy measurement, as a linear combination of all observations up to time k. The
587 following assumption should be satisfied:
588
589 o Initial condition

590 EX 0  X0 (12)
 
591  
E X0  X0  X0  X0   P
T
0 (13)
592 Ewk   0 (14)
593 Evk   0 (15)

594
595 where E   represents the expectation operator.
596 o The initial system state, system noise, and measurement noise are uncorrelated
597 o The system noise and measurement noise are white, independent, and Gaussian distributed
598 with known covariance matrices.
599
600 Though the predicted state is given by:
601
 
602 X k1/ k  Fk X k / k  Gk wk (16)
603 Pk1/ k  Fk Pk / k FkT  Gkuk (17)

604
605 According to [100, 114], a complete discrete KF scheme to solve this problem consists of the following
606 five equations:
607 1. State estimate extrapolation:

 
608 X k 1 k   k1 X k (18)
609 2. Covariance of the estimation error (State Covariance Extrapolation):
610 Pk 1 k   k 1 Pk Tk1  k (19)
611 3. Kalman Gain (KG) Computation:
612 
K  k 1   Pk 1 k  H kT1 H k 1 P k 1 k  H kT1  R k 1 
1
(20)
613 4. State Estimate Update
  
614 
X k 1  X k 1 k   K k 1 Z k 1  H k 1 X k 1 k  
1
(21)
615 5. Error Covariance Update
616 Pk1  Pk1 k   Kk1Hk1Pk1 k  (22)
617 where:
618 - X(k/k-1): An estimate of X at a time k based on data up to sample time k-1

619 - X k1 k : System state vector at time k+1 based on time k

620 - Xk : System state vector at time k
621 - | : Transition matrix at time k+1 based on time k
622 - | : System state vector at time k+1 based on data up to sample time k
623 - | : Kalman gain matrix at time k+1 based on time k
624 - | : Prediction covariance at time k+1
625 - | : System sate vector at time k+1 based on time k
626 - k : System error covariance at time k

627 - : Measurement noise matrix at time k+1
628 - : Estimation error at time k
629
630 Eqs. 18 & 19 represent the prediction part of the algorithm, while Eqs. 20-22 represent the
631 correction part of the algorithm. The prediction part simply consists of the dynamic model, which
632 predicts the next data of the system (at time k+1) based on the last data (at time k-1) or the current
633 data (at time k). The correction part takes the error between the current estimate and the predicted
634 output and uses it to correct the state estimates to obtain the best estimate of the system state Xk based
635 on an old observation data at the time k. The mixture of the prediction and correction is determined
636 by the Kalman gain. It is the ratio of the error in the estimate divided by the sum of the errors in the
637 estimate and in the measurement. This gain determines the extent to which the filter follows the
638 model or the measurement. The overall result is the best guess of the parameter to be determined,
639 which is obtained by combining these two different sources of information. The adjustment to the
640 previous estimate to come up with the new estimate depends upon the Gain. Based on the previous
641 estimate, the Gain will decide the relative weightage of the new measured value and the previous
642 estimate to update the new estimate. Once the current estimate is determined, the error in the estimate
643 should be determined so as to use in the next time round.
644
645 KF methods were introduced as a fault isolation and assessment technique in the late 1970s, and
646 the overall architecture is shown in Figure 6 [115]. They brought it into practice aiming to overcome
647 the two most GPA limitations: poor robustness against measurement uncertainties and the
648 underdetermined problem due to the presence of limited numbers of measurements. The success
649 attained in these early programs encouraged the use of these techniques in subsequent years [98, 100,
650 116]. The linear KF has reliability limitations on non-linear gas-path diagnostic problems. However,
651 the modified versions of this method, named extended KF (EKF) and Iterated EKF (IEKF) can solve
652 the problem by linearizing the current mean and covariance using Taylor series expansion [117, 118].
653 Further discussion on these versions is available in [151]. The well-known engine manufactures
654 (General Electric, Pratt & Whitney, and Rolls-Royce) have been utilized modified KF based fault
655 diagnostic methods since 1987 [119]. It is also integrated with the currently available GT gas-path
656 diagnostic tools such as Auto Analysis, MAPIII, TEAMIII, a self-tuning onboard real-time model
657 (STORM), a state variable engine model (SVM), GEM, COMPASS, an engine health management
658 (EHM) and ADEM [9, 112]. KF based fault diagnostic techniques are effective for engine problems
659 where performance influence coefficients are available as the model [112]. However, those methods
660 have reliability limitations. Most MB techniques which are relatively coping with measurement noise
661 and bias are developed utilizing this technique [120]. The potential of KF for a single gas-path
662 component fault isolation was evaluated by Volponi et al. [114]. Multiple KF models were used for
663 sensor and actuator fault detection and isolation purpose together with a component fault detection
664 in an aircraft engine by Takahisa et al. [121]. The effectiveness of KF on sensor selection for a reliable
665 engine performance diagnostics was also investigated by Simon and Rinehart [122] in comparison
666 with a maximum a posteriori (MAP). They considered a liner engine model affected with single
667 component faults and sensor biases. The fault detection and classification performance of the method
668 using 7, 8, and 9 sensors associated with 8 health parameters were tested. Borguet et al. [123]
669 attempted to dealt with one of the difficulties of MB methods, i.e., the existence of model biases, using
670 simulated transient data. A modular KF based single and double fault FDI algorithm was proposed
671 by Meskin et al. [124] for a jet engine application. Recently, the sensor FDII performance of multiple
672 hybrid KF based system was investigated by Pourbabaee et al. [66]. In this method, nonlinear
673 mathematical model of the system and multiple piecewise linear (PWL) models are combined to
674 accomplish the sensor FDI task followed by estimating the fault level using modified generalized
675 likelihood ratio (GLR) method. The capability of EKF to solve underdetermined engine diagnostic
676 problems was also evaluated by Lu et al. [125]. Besides, they compared the performance of three
677 different EKF estimators; basic EKF, underdetermined EKF, and resultant EKF. The test results
678 indicated that the method was able to solve the underdetermined problem with a promising accuracy
679 and robustness than the conventional linear KF scheme.
680
681 Figure 6. GT gas-path diagnostics using Kalman filter, adapted from [125].
682 4.2. Advantages and Limitations of MB Methods

683 Every method has its own advantages and limitations. Table 4 describes advantages and
684 limitations of GPA and KF methods related to GT fault diagnostics. In general, MB methods have
685 more advantages in terms of early fault detection and online fault diagnostics. They can also perform
686 both quantitative and qualitative fault diagnosis with an adequately good accuracy. Moreover, they
687 apply the real gas-path physics and have low model complexity and computational time.
688 Nevertheless, they suffer from model uncertainties, measurement noise, and sensor bias (even if this
689 problem partially addressed by KFs), and smearing effects which may lead to a misinterpretation and
690 false alarms. They require a large number of measurements on the gas-path to provide an accurate
691 diagnostic solutions. Installing additional sensors is almost impossible due to the reasons mentioned
692 in section 1.3.1. Besides, since a very limited information is available on the public domain due to
693 proprietary issues, an accurate GT performance model is very difficult to obtain.
694 Table 4. Advantages and limitations of GPA and KF related to GT diagnostics.
Method Advantages Limitations

GPA - Qualitative and quantitative assessment of the health status of - Reliability of GPA-based methods depends on the accuracy of the ICM. An accurate ICM
gas-path component(s) is possible using measurement requires knowledge about component characteristic changes due to different faults.
deviations [9, 106]. - The reliability of an LGPA highly influenced by the magnitude of the fault and there may
- It is capable of diagnosing multiple faults. also be solution convergence problem for large fault values [5].
- It is the most appropriate method for measurement parameter - A large number of sensors are desired [1]. This requires installation of additional sensors
selection [86]. which do not exist in practice, thereby increase sensor related problems [72].
- It helps to understand the physical, thermal, and aerodynamic - It is not coping with measurement uncertainty [92].
nature of the engine behavior. - Lack of access to modeling information due to proprietary issues.
- It is a cost-effective method with good early fault detection - It needs a detailed knowledge of the behavior of each component on the gas-path
capability [11]. - The diagnostic accuracy relies on the accuracy of the engine performance model [62]
- High computational speed [126]
KF - It can provide sufficiently accurate estimation results for linear - Even the EKF based methods can only handle problems with a limited amount of
problems. nonlinearity. The estimates for a nonlinear diagnostic problems are often biased and
- It has relatively low computational complexity suboptimal [127].
- Low storage and computing requirements - Prior knowledge and Tuning: The effectiveness of KF is affected by the unknown
- Since MB, the physical knowledge of the gas-path system can performance deterioration and measurement noise covariance matrices. Choosing the
be applied to solve the diagnostics problem. appropriate covariance matrix (called tuning) for an optimized KF performance, based
- Unlike GPA, the concern of measurement uncertainty is on prior knowledge, is a random and challenging task [128].
undertaken. The actual sensor noise can be represented by white - “Smearing” effect: although in practice, most of the time, only a limited number of gas-
Gaussian distribution, as it is desired by KF. They are good at path components and sensors are affected, the KF oppositely leans to spread of (“smear”)
outlier reduction and noise minimization [112]. the faults over multiple components’ performance parameters and measurement
- It is coping with sensor noise and bias [9] parameters. An attempt to estimate all component faults and sensor faults together using
- Unlike GPA, it has the potential to solve an underdetermined the available measurements results a highly nondeterministic problem [129].
diagnostic problems [125]. - Solution convergence problem due to model uncertainty and large sensor noise [130].
695
696
697
698
699 4.3. AI based Methods

700 The drawbacks of the MB methods as highlighted in Table 6 above forced the research
701 community to focus on AI methods. According to Konar [131], AI is defined as “the simulation of
702 human intelligence on a machine, in order to make the machine efficient to identify and use the right
703 piece of “Knowledge” at a given step of solving a problem”. There are many different AI methods
704 such as GA, ANN, BBN, FL, and ES. The most powerful and popular types of fault diagnostic
705 algorithms are from AI methods [12]. Demonstrating and validating AI based algorithms require
706 operational data with an appropriate quality, quantity, and type or model simulation data, in the
707 absence of operational data. Figure 7 illustrates the conceptual framework of AI based engine fault
708 diagnostics. It has two parts; developing the diagnostic mechanism and its implementation. The task
709 of developing the method includes acquisition of the required data and preprocessing it, training the
710 algorithm using the processed data, and evaluating its performance by applying the appropriate
711 evaluation approach, for instance, using a blind test case data as proposed by Simon [132]. The
712 potential of AI methods on GT FDII have been widely studied in the past several years. A
713 comprehensive survey of these methods including their strengths and weaknesses is presented
714 hereafter.
715
716 Figure 7. Schematics of the general AI based GT fault diagnostic procedures.
717 4.3.1. Artificial Neural Networks

718 An ANN is an artificial structure that processes information like a biological neuron does, except
719 this paradigm is mathematical instead. Its model contains sets of neurons connected each other in
720 layers. Knowledge is acquired from an input information (examples) through a learning process and
721 weights of connections between neurons are used to store the acquired knowledge. There are three
722 popular ANN learning paradigms, namely supervised, unsupervised and self-supervised [128]. If the
723 training is taken place using input and output examples it is called supervised, if it is performed
724 using the information derived from input data only it is called unsupervised, and if it is done utilizing
725 the same input and output information the learning is named self-supervised. There are various ANN
726 algorithms in literature such as MLP, AANN, RBFN, and PNN. These methods have been applied to
727 solve different engineering problems including prediction, pattern recognition/classification, and
728 clustering [133]. They are capable of providing efficient and reliable models if sufficient amount of
729 data is available [90]. ANN is a powerful tool in GT modeling for performance prediction and
730 diagnostics due to its capability to undertake the nonlinearity of the engine behavior [134]. This is
731 done without the need of the complex thermodynamic equations that relate the dependent and
732 independent parameters.
733 4.3.1.1. Multilayer Perceptron

734 A MLP network is a feed-forward neural network consisting of input and output layers with
735 one or more hidden layers in between [135]. It is called feed-forward because information from the
736 input neurons is passed to the next layer neurons and then they compute an output based on the
737 logistic equation and pass it to the next layer of neurons and so on until the end. The general
738 architecture of an MLP consists of an input layer, one or more hidden layers, and an output layer (see
739 Figure 8). The first layer is the input layer that the input data goes in while the output layer, located
740 at the end, computes the output value using the information coming from the hidden layers. The
741 optimal number of hidden layers and neurons is determined based on a convergence criterion and
742 the input-output mapping relationship characteristics.
743
744 Figure 8. General structure of MLPNN.
745 In the past decades, several studies have been done on GT diagnostics based on an MLP [114].
746 ANN-based user friendly GT fault identification system was provided by Kong et al. [136]. A multiple
747 fault detection system was developed by Matuck et al. [137] using this approach which is trained on
748 simulation data. They considered single, double, and triple component faults together with sensor
749 noise. However, this work was limited to fault detection only. Fast et al. [90], proposed a GT fault
750 diagnostic schemes using MLP in order to optimize the compressor washing schedule. They have
751 also indicated that ANN is a suitable approach to develop a performance prediction and fault
752 diagnostic techniques if an operational data is available in the required quality and quantity. To
753 answer the question why neural networks are more popular than the other AI methods, Patan et al.
754 [138] conducted a research work on two different feed-forward MLP algorithms taking into account
755 the nonlinear behavior of the GT together with modelling uncertainty. It was pointed out that those
756 fault diagnostic algorithms were having better early detection ability with smaller false alarms, high
757 fault classification rate, and efficient fault identification than the other AI techniques. Recently,
758 Tayarani-Bathaie et al. [139], Mohammadi et al. [140], Kiakojoori and Khorasani [141], and Vanini et
759 al. [62] proposed a dynamic neural network (DNN) fault diagnostic techniques for aircraft engine
760 applications using. More recently, an ensemble GT fault diagnosis system is devised by Amozegar
761 and Khorasani [142] using different types of MLP networks. Nested MLP networks were also used
762 to a fault detection and isolation application by Tahan et al. [143]. However, these methods are limited
763 to single and double faults only. Moreover, they used efficiency and flow capacity deltas separately
764 as a single component fault or in a pair as a double component fault although different studies on GT
765 performance degradation like [19, 21], indicated that deterioration can be most significantly
766 represented by changes of these parameters together.
767 4.3.1.2. Autoassociative Neural Networks

768 AANNs (also known as auto-encoder, bottleneck, replicator network, or sand-glass type
769 network [144]) are an important family of ANNs with three hidden layers; a mapping layer, a
770 bottleneck layer, and a de-mapping layer (Figure 9). They can reduce high dimensional data to a
771 lower dimension with an insignificant information loss based on the concept of principal component
772 analysis (PCA). Usually, the input and output layers contain equal numbers of neurons. The
773 bottleneck layer is the middle layer with the smallest number of neurons, where the important
774 features of the input data captured. The number of hidden neurons associated with each hidden layer
775 depends on the problem type and system complexity. In this regard, a more detailed information is
776 available in [145]. Compression and decompression are the two sub-networks of the general AANN
777 structure, where the former used to compress a high dimensional input data to a low-dimensional
778 feature and the latter tries to reconstruct the original data from the compressed version with
779 minimum information loss.
780
781 Figure 9. General architecture of AANN.
782 AANNs are widely used and very suitable for sensor data validation applications. Kramer [146]
783 introduced an AANN based technique for sensor validation that is coping with nonlinearity of the
784 data. He used the network residuals to detect and estimate sensor faults. An AANN based sensor
785 validation technique for a turbofan engine was proposed by Guo et al. [147]. Lu et al. [148, 149],
786 evaluated the performance of AANNs for sensor noise reduction and bias detection and correction.
787 While training the noise filtering networks, they used noisy data as an input and noise-free data as
788 an output. The networks, therefore, tried to provide an output as close to the desired noise-free data
789 as possible. Besides, the effect of the number of measurements on the accuracy of the proposed
790 methods was tested using 4 and 9 parameters and achieved almost similar success rates. AANNs are
791 better at outlier removal and noise reduction than the conventional filtering techniques [112]. A
792 multiple sensor fault detection and isolation method using a bank of AANNs together with a MLP
793 based fault identification technique was developed by Zedda and Singh [120] for a low-bypass-ratio
794 turbofan engine purpose. It has been shown that AANNs are capable of successful sensor failure
795 diagnosis, even in the presence of component faults. The effectiveness of using multiple hierarchical
796 AANNs to diagnose single and double sensor faults in a 2-shaft industrial GT engine was analysed
797 by Ogaji et al. [74]. They used three networks: the first one is to separate faulty and fault-free
798 measurements, the second to differentiate sensor and component faults and the last to estimate and
799 accommodate the amount of sensor faults. Such kind of diagnostic task division is important to share
800 diagnostic tasks, which may improve the accuracy significantly. A combined discrete wavelet
801 transform and AANN based diagnostic system was also developed by Tamiru et al. [150] for oil
802 system, vibration system, control system diagnostics purpose. Recently, an AANN was used for
803 single and double sensor and component fault diagnosis by Vanini et al [75]. However, the sensor
804 validation performance of AANN based techniques is influenced by the amount of the noise-level
805 and the threshold characteristics. Minimizing the number of false alarms and missed detections is
806 equally important, but they have opposite correlation (i.e., decreasing the number of false alarms by
807 increasing the threshold level may oppositely increase the rate of missed detections) [112].
808 4.3.1.3. Probabilistic Neural Network

809 PNN is a statistical pattern classification multilayer feed-forward network based on the Bayes
810 pattern classification strategy (based on RBFs) [151]. The architecture of a typical PNN, as shown in
811 Figure 10, consists of three layers; an input layer, a pattern layer and an output/summation layer. The
812 input layer passes the input patterns to be classified to each of the nodes in the pattern layer. This
813 layer contains individual neurons corresponding to each pattern/example in the training data set. The
814 neurons in the pattern layer compute their responses based on Eq. 23 and feed into the output layer
815 neurons (Eq. 24). The output neurons stands for the desired output groups that the network is
816 expected to classify the input patterns into. Thus, all the connection weights between output layer
817 and pattern layer have a value of 1. PNN can be characterized as: they are simple in design, have a
818 similar training nature with backpropagation algorithm, have quite huge pattern layer, and low
819 computational speed [127].
820
821 Figure 10. PNN general architecture.
822 If Xij ϵ RMN (i = 1,2,…M and j = 1,2,…N) is the input vector with N number of input parameters
823 and M cases and K is the number of target classes, for an input pattern x and assuming Gaussian
824 probability density function (PDF), the pattern layer and output layer neurons’ outputs can be
825 computed, respectively, as [152]:
1  x  xij T x  xij 
826 ij  x   exp   (23)
2 M 2  M  2 2 
1 1 N  
 x  xij T x  xij  
827 f i x    exp   (24)
M
2   i
2 M N
j 1
 2  2 
 
828
829 where β is the smoothing parameter.
830
831 The capability of PNN for GT fault diagnostics was checked for the first time by Eustace and
832 Merrigton [153] by implementing it to a GE low-bypass F404 military engine. Romessis and
833 Mathioudakis [154] also used this method for sensor fault diagnosis in a deteriorated engine
834 condition. In another study, PNNs were used for sensor fault diagnostics in a jet engine for on-board
835 application [155]. Like the other ANNs, PNN applies the concept of pattern recognition technique for
836 fault isolation and identification tasks [156]. It uses a probabilistic measure to decide the type and
837 location of the fault and assess its magnitude. Nested PNNs have been used by Ogaji et al. [83] for
838 senor and component fault diagnostics in a 2-shaft aircraft engine. For this purpose, five PNN nets
839 are used: the first is to separate the fault and no-fault patterns; the second is dedicated to sensor and
840 component fault classification. The remaining three nets are used for component fault classification.
841 Then sets of radial basis networks are integrated to quantify the magnitude of the faults. The results
842 revealed that the proposed scheme was capable to diagnose all the considered fault scenarios with a
843 sufficiently high accuracy. Recently, the fault classification performance of PNN was compared with
844 MLP and RBF by Loboda and Robles [157], and obtained similar accuracies. In general, as per this
845 review, most of the previous PNN based GT diagnostic techniques were utilized for fault
846 classification tasks.
847 4.3.1.4. Radial Basis Function Networks

848 RBF also called a kernel function is a multivariate approximation function whose value depends
849 only on the distance from the origin or center c, i.e., the Euclidean distance [156]. It performs a
850 nonlinear transformation over the input vector before it is feed for classification. By using such
851 nonlinear function, it is possible to convert a linearly non-separable problem to a linearly separable.
852 RBF increases the dimensionality of the feature vector. The most commonly used types of RBFs
853 include Gaussian, multiquadric, and inverse multiquadric [158].
854 As shown in Figure 11, the general structure of an RBFN composed of three layers: an input
855 layer, a hidden layer, and an output layer [135]. Like other ANNs, the number of nodes of the input
856 layer is equal to the dimension of the input vector. The task of the hidden layer neurons is to project
857 the input vector into a higher dimensional vector. Hence, the number of neurons in the hidden layer
858 must be greater than the input layer. This is because if the feature vectors are linearly non-separable
859 in the input dimensional space, then it is more likely that those feature vectors will be linearly
860 separable when we cast them into a higher dimensional space. Once the feature vectors are linearly
861 separable in the M-dimensional space of the hidden layer, the linear combination of the outputs of
862 the hidden layer is likely to give the class it belongs to. The class combination is decided by the
863 connection weights from the hidden layer nodes to the output layer nodes. In general, in this process,
864 there are two important tasks: determining the receptor and the distribution of the function and the
865 connection weights. Every neuron in the hidden layer represents an RBF. Due to its useful analytic
866 properties, in addition to localization, the Gaussian function is more commonly used [158]. The
867 number of nodes at the output layer is the same as the number of desired classes. Like the other MLP
868 nets, RBF output units use linear summation functions.
869
870 Figure 11. General structure of RBF.
871 Although RBF networks can perform different tasks like function approximation, pattern
872 classification, and dynamic system modelling, they have been widely used in function approximation
873 [128]. It is capable of approximating any arbitrary function of the network layers utilizing the training
874 input dataset. The use of RBFs experiences many advantages over backpropagation based
875 feedforward neural networks, for example, rapid training, very low computational expense, and very
876 good at interpolation, generality and simplicity [128]. They are highly localized and need a huge
877 quantity of training data, thereby creates a large amount of nodes that allows a rapid training. In
878 addition, they work better than any other training technique and are able to approximate any
879 continuous function [159]. However, after training, their computational speed to perform
880 classification or approximation is low.
881 The Euclidean distance is computed from the point being evaluated to the center of each neuron,
882 and an RBF is applied to the distance to compute the weight for each neuron [156]. The further a
883 neuron is from the point being evaluating the less influence it has. Leonard et al. [160] suggested a K-
884 means clustering technique to determine cluster centers, a K-nearest heuristic technique to determine
885 the width of the RBF and multiple linear regressions to determine connection weights of the layers.
886 Assume a set of training feature vector and suppose C number of classes are required, for
887 Gaussian function the network output can be computed as [91]:
2
M  X  
j
888  
yk X   wkj exp 2
 (25)
j 0  2 j 
 
889
890 where yk is the kth output, wkj is the weight of the connection between the jth hidden unit and the kth
891 output unit, µ is the receptor/center of the function, σ (as shown in Eq. 2.26) is the standard deviation
892 of the function that decides what the spread of every RBF looks like, X is an M-dimensional input
893 vector, and X   j is the Euclidean norm.
1 p r 2
894  i    ni  nj   (26)
 p j 1 n1 
895
896 where p is the nearest number of RBF classes, r is the number of entries, µki and µkj represents the
897 receptors of the ith and jth hidden units of the nth entries.
898 Previously, a RBFN was often used for data cleaning prior to a fault diagnosis [91]. In this study,
899 it has also been shown that their proposed a RBFN based noise filtering technique reduced the
900 measurement noise by 75-81%. This performance is much better than the conventional linear filters.
901 The problem of measurement outliers and noise were undertaken by Roy et al. [161] using a RBFN.
902 The result showed that from 59-73% of the data outliers and noises were removed, which is still much
903 better than the traditional filtering methods. Ogaji et al. [5] used sets of RBFNs for gas-path
904 components fault approximation application. The proposed method comprises of three steps: fault
905 detection, fault isolation, and fault identification. The first two tasks were performed using sets of
906 PNNs. After the detection and isolation stages, RBFNN based techniques were applied to estimate
907 the level of component faults using their corresponding fault patterns coming from the associated
908 isolation networks. The applicability and performance of RBFN for a GT fault identification were
909 compared with a MLP by Loboda et al. [162]. They concluded that the RBF network resulted in a
910 little more accurate results than the conventional MLP network, however, the former requires much
911 more storage capacity and computational time. Recently, a similar work has been conducted using
912 RBFs [163]. As per [161], RBFNs work better than many other training techniques and require much
913 less training computational time and cost than backpropagation based algorithms.
914 4.3.2. Fuzzy Logic methods

915 FL is a nonlinear mapping of an input feature vector into a scalar output [164]. It is one of the
916 most widely used AI methods to approximate the relationship between dependent and independent
917 parameters based on a set of IF/THEN statements. The general FL approach consists of four basic
918 components: Fuzzy Rules (sets of IF---THEN statements), Fuzzifier (the mechanism which maps
919 numbers of input signals into the fuzzy set), Inference Engine (the technique used to determine the
920 ways in which the fuzzy sets are combined with each other) and Defuzzifier (the mechanism used to
921 calculate the output values) [94]. The schematic representation of a rule-based FL system is shown in
922 Figure 12. For a GAS TURBINE DIAGNOSTICS, sets of measurement parameter deltas are used as
923 an input to the FL system in order to compute the performance parameter deltas.
924
925 Figure 12. Framework of a rule based fuzzy system (adapted from [127]).
926 After the earliest use of FL by Fuster et al. [165] in 1997, several FL based GAS TURBINE
927 DIAGNOSTICS techniques have been devised by other researchers. Among these, Marinai [127, 166]
928 contributed a diagnostic model for a Rolls-Royce Trent 800 engine that can isolate both single and
929 multiple component faults in the presence of sensor noise and bias. Simulated data for clean and
930 faulty GT cases were used to test the fault detection performance of the model and results showed
931 that the detection based on filtered data was very accurate with negligible missed alarms and no false
932 alarms. However, the investigation of the method for multiple fault diagnosis was limited to dual
933 component faults only. Similarly, Ganguli [167], developed GT measurements’ trend shift detection
934 mechanism using median filters and FL. The test results revealed that the detection based on filtered
935 data was very accurate with negligible missed alarms and no false alarms. In order to undertake the
936 problem of availability of limited numbers of sensors on a real GT service, Ganguli [164] developed
937 a FL based single fault isolation system for a jet engine using only four commonly available sensors.
938 The proposed method can isolate 95% of the faults successfully. The accuracy increased with the
939 number of sensors and reached 100% for eight sensors. He has also been stated that FLs can work
940 with poor quality data. In another study, Ogaji et al. [168] proposed a diagnostic system for a modern
941 military turbofan engine that can identify single component faults with an accuracy of 92.5%.
942 Recently, Kyriazis et al. [169] developed a FL based GT compressor fault diagnostic system. Its
943 effectiveness was compared with pattern recognition and PNN methods. The results showed that the
944 FL method has as good generality and effectiveness in fault diagnostics as the other two methods.
945 4.3.3. Genetic Algorithm

946 A GA is a method which mimics nature on the basis of Darwin’s evolutionary theory of ‘survival
947 of the fittest’ [164]. It begins with the population of randomly generated structures where each
948 structure encodes a solution to the task attempt and proceeds to evolve generations. In each
949 generation, the quality of every individual within the population is assessed by its fitness, multiple
950 individuals are selected from the current population (based on their fitness) and modified
951 (recombined and possibly randomly mutates). Since the fitness is a function of the objective function
952 (OF) to be optimized, the optimum solution is reached when the OF is approaching zero and the best
953 fitness is associated with the value of one. It is an iterative procedure each of which consists of three
954 operators; selection, crossover and mutation [170]. During each generation, the GA improves the
955 structures in its current population by performing selection followed by crossover followed by
956 mutation. It is looking for best solutions rated against fitness criteria. Thus, it avoids local optima and
957 searches for a global fitness. During selection, it duplicates higher fitness structures and deletes
958 structures with lower fitness. Crossover results in good components of good structures combining to
959 yield the better structures, and then recombines elements of good chromosomes from different
960 genomes. Mutation creates new structures that are similar to current structures with a small pre-
961 specified probability. Figure 13 shows the schematics of the cyclic process of the three GA operators.
962 GAs are different from the traditional optimization techniques through the following three
963 significant points [171]. First, since they search parallels from a population of points, they have the
964 ability to avoid being trapped in local optimal solution. Second, GAs work on the chromosome which
965 is an encoded version of potential solution parameters rather than optimizing the parameters
966 themselves. And third, GAs use fitness score which is obtained from OFs without an artificial over
967 engineer black-box mathematics. In the end, like the other methods, the user typically choses the best
968 structures of the last population as the final solution. A more detailed description of this method can
969 be found in [172].
970
971 Figure 13. Adaptive GA model organization (adapted from [170]).
972 GA is often applied as an effective optimization tool to obtain a set of component parameters
973 that produce a set of predicted dependent parameters, through a nonlinear GT model that leads to
974 predictions that best match the measurements [92]. The solution is obtained when the OF (which is
975 the measure of the difference between predicted and measured values) achieves the minimum value.
976 A simplified illustration of GA based GT fault diagnostics strategy is given in Figure 14.
977
978 Figure 14. Schematic diagram of GA based diagnostic strategy (adapted from [170]).
979 According to Zedda and Singh [120] and Singh [1], when measurement noise is assumed to be
980 Gaussian distribution, the suitable OF to be optimized is given as:
981
M
J  x  
z  h x, w
j j
2
982 (27)
j 1z w  
odj j
2
983
984 Or, if the absolute deviation is considered (this is suitable when measurement error distribution is
985 assumed to be other than Gaussian distribution and when modeling errors are inevitably present)
986 Eq. 2.28 is more suitable.
987
M z j  h j  x, w 
988 J x    (28)
j 1 z odj w   j
989
990 where J is the OF, M is the number of measurements, zj is the value of the jth measurement, h is a
991 vector valued function, w is power setting parameters, zodj is the value of the jth measurement in the
992 off-design clean condition, σj is the standard deviation of the jth measurement (noise value).
993 The application of GAs for a GT fault diagnosis was started in 1999 by Zedda [129], and have
994 been investigated by several scholars since then [173]. The problem of obtaining an accurate MB GT
995 fault diagnosis method in the presence of a limited number of measurements, far less than the number
996 of performance parameters, is undertaken by Zedda and Singh [120] using GA. The performance of
997 this method was evaluated by applying on Rolls-Royce (RR) RB-199, RR RB-211 and ICR-WR21
998 engine types. For this purpose, the performance models for each case engines were developed
999 utilizing the well-known GT engine performance code TURBOMATCH. Sampath et al. [170],
1000 developed a GA-based sensor and component fault diagnostic scheme that can deal with the
1001 nonlinearity nature of the diagnostic problem, for a double shaft GT engine application. In this work,
1002 the effects of sensor noise and bias and the number of sensor and component faults on the accuracy
1003 of component fault diagnosis were analysed using simulation data from a GT performance modelling
1004 tool, called Rolls-Royce Aerothermal Performance (RRAP). Six gas-path components (inner and outer
1005 Fans, HPC, HPT, LPT, and Nozzle), when affected individually or in pair (as a single component fault
1006 (SCF) and double component fault (DCF) cases), were analysed using 16 gas-path measurements in
1007 the presence of two and four concurrent sensor faults. For this purpose, 6 SCF and 15 DCF classes
1008 were considered. A GA based GT fault diagnosis method was used to find an optimal combination
1009 of a set of performance parameters and the corresponding set of best match measurement parameters,
1010 through a non-linear performance simulation model [109]. A generalized GA based GT fault
1011 classification method was proposed by Loboda et al. [174] using a thermodynamic model that is
1012 applicable to multiple operating point conditions for both steady state and transient cases. Li and
1013 Pilidis [104] and Li et al. [175] applied GA for a GT performance adaptation in order to assess the
1014 engine’s health status. For this purpose, the information from the measurements was used to estimate
1015 the component faults at a specified design point and off-design operating condition.
1016 Recently, the fault diagnosis effectiveness of an NLGPA and a GA based method was compared
1017 by Kong et al. [176], applying on a 2-spool turbofan engine. They showed that the diagnostic MB on
1018 GA is better than the NLGPA, particularly when sensor noise and bias are considered. In a similar
1019 manner, Kong [94] investigated the diagnostic effectiveness of GA in comparison with NLGPA and
1020 fuzzy-neuro techniques taking in to account measurement uncertainty and sensor fault effects. The
1021 test results indicated that the GA based method showed a reliable accuracy than the NLGPA,
1022 especially in the presence of sensor noise and bias.
1023 4.3.4. Bayesian Belief Network

1024 A Bayesian Belief Network is a graphical representation of a probability distribution AI based
1025 method that represents the cause and effect relationships among predisposing factors, faults and
1026 symptoms [128]. The graph consists of nodes which represent a set of random variables and directed
1027 edges indicating their dependencies. The degree of relationship between the variables is expressed
1028 in terms of conditional probability. In figure 15, an example of a BBN structure referred to a gas
1029 turbine fault diagnostics is presented. In this structure, the parent nodes are dedicated to the engine
1030 performance parameters and the child nodes to the measurement parameters. The given performance
1031 parameters and the measurement parameters of the case engine are related through sets of directed
1032 connections along with their associated probability values. According to the BBN approach, the
1033 engine gas path diagnostic problem can be expressed mathematically as:
P z / x Px 
1034 Px / z  
P z 
1035
1036 where P(x/z) is the probability of x given z, P(z/x) is the probability of z given x, x is the independent
1037 parameter (performance parameter), z is the dependent parameter (measurement parameter), P(x) is
1038 the probability of the independent parameter x, and P(z) is the probability of the dependent
1039 parameter z.
1040 The application of BBN for gas turbine diagnostics was started in the early 19’s by Breese et al.
1041 [177]. Their proposed technique relies on a model-based method that integrates a BBN with an expert
1042 system. It was implemented to assess failures on the engine oil cooling system, bearing, and bearing
1043 temperature sensors. A few years later, Palmer [178] developed a BBN based fault diagnostic system
1044 for the CF6 engine application, although model details were not provided. A more detailed BBN
1045 based aircraft engine gas path fault diagnostic procedure was provided later by Kadamb [179],
1046 Romessis et al. [180], Mathioudakis et al. [181], and Romessis and Mathioudakis [182] with the aid of
1047 an engine performance model. They also showed the capability of their proposed method dealing
1048 with engine diagnostic problems with measurements less than the performance parameters to be
1049 assessed. Lee et al. [183] suggested hierarchically arranged multiple BBN models based an offline
1050 fault diagnosis method for industrial gas turbine engines under steady-state operating conditions. It
1051 has been reported that the proposed method is capable of carrying out both qualitative and
1052 quantitative diagnostics under measurement’s uncertainty.
1053 In the presence of large data samples, BBN can be trained in a supervised or unsupervised
1054 learning manner, although majority of the past attempts focused on the supervised one [184].
1055 However, developing a BNN classifier based on expert knowledge is highly complex and time
1056 consuming, and it is subject to errors. Because it requires a classified training data sample based on
1057 the expert’s priory knowledge that can be used to model the BBN structure and generate its
1058 conditional probability table (CPT). In many gas turbine applications, it is difficult to obtain the
1059 required large data samples with the fault class information. For isolation of engine fault classes with
1060 a data that has no labels assigned to, the unsupervised learning based BBN might thus be preferable.
1061 Different algorithms exist for this purpose such as the Greedy Search-and-score, constraint based
1062 methods, and hybrid of these two [185].
COMPONENT 1  Measuremen t
Parameter 1
 COMPONENT 1  Measuremen t
Parameter 2
COMPONENT 2
 COMPONENT 2
 Measuremen t
Parameter m
COMPONENTn
 COMPONENTn
1063
1064 Figure 15. Typical BBN layout for engine fault identification, adapted from Kadamb (adapted from
1065 [179]).
1066 4.3.5. Expert Systems

1067 ESs are software programs which use to capture human expert knowledge in the form of facts
1068 and rules to solve problems or give advice as a human expert [92]. The architecture of ESs (Figure 16)
1069 comprises of four basic elements: user interface (used to acquire information and display results),
1070 inference engine (deals with all the reasoning operations of the system based on known facts and
1071 rules), knowledge base (contains facts and rules about the problem to offer the appropriate decision)
1072 and developer (stores information about current education) [94].
1073
1074
1075 Figure 16. Configuration of an expert system (adapted from [94]).
1076 The knowledge from the expert is first prepared in the form of knowledgebase by a knowledge
1077 engineer or programmer. The knowledge base contains data and facts in that specific area of
1078 application or knowledge domain. The information in the knowledgebase is intended to replace the
1079 human expert. The user interface presents questions to users, accepts information from them and
1080 then provides answers and sometimes the reasoning for those answers too. Interface engine has a job
1081 of matching the user’s input from the user interface with the data contained in the knowledgebase to
1082 find appropriate answers. This is done using interface rules, which describe how different items from
1083 data relate to each other and sometimes using probabilistic rules. ESs are programmed with a series
1084 of logical rules to find a solution. Very basic ESs use Boolean logic or decision trees. Boolean logic has
1085 two possible values, true and false, yes or no, etc. The problem is that Boolean logic has only two
1086 values making it difficult to represent real life problems. To avoid the problem of decision trees, ESs
1087 typically use inference rules and chaining to reach conclusion. Inference rules are written as IF …
1088 THEN statements which describe rules for a knowledge domain.
1089 Numerous researches have been done using ESs in the previous several years and many
1090 different ES based GT diagnostic techniques are available in the open literature [92, 186]. They can be
1091 broadly categorized into rule based, MB, and case based techniques [92]. The earlier forms of ESs
1092 based GT gas-path diagnostics applies pattern recognition/matching technique by comparing
1093 patterns of measurement deltas with performance parameter deltas/fault signatures obtained from
1094 OEM [9]. As listed out in [92], enormous ES based tools specific to different GT models and
1095 configurations were introduced by GT manufacturers and researchers such as ENGDOC, TEXMAS,
1096 HELIX, XMAN, TIGER, IFDIS, and SHERLOCK. However, knowledge from domain specific experts
1097 is usually inexact and reasoning on knowledge is often imprecise. An ES dealing with uncertainty
1098 and proved to be very efficient in fault diagnosis is Bayesian Belief Network (BBN). However, these
1099 systems require precise inputs and rely entirely on knowledge of experts and extensive database of
1100 rules.
1101 4.4. Advantages and Limitation of AI Methods

1102 Table 5 presents summary of the advantages and limitations of AI based methods reviewed
1103 above. This may help beginners of this field in selecting the appropriate diagnostic approach for
1104 variety of engine diagnostic problems. In general, unlike MB methods, AI based methods can handle
1105 the effect of sensor noise and bias, the possible existence of multiple faults simultaneously, the fault
1106 identification problem using a limited number of instrumentation suite, and the nonlinear
1107 relationship between the measurement parameters and the performance parameters. However, most
1108 of AI based techniques cannot give confidence limits on the output. In addition, they are not capable
1109 of diagnosing faults outside the domain of the data to which they have been exposed during training.
1110
1111 Table 5. Advantages and limitations of AI-based methods regarding GT diagnostics.

ANN - They can deal with the nonlinear nature of the engine performance. - Since they are unable to perform reliably outside the range of the data
- They are capable of learning with time. they are trained on, a huge amount of data that can sufficiently represent
- They are very good in data fusion [1]. the entire life of the engine condition is required.
- Training can be done by means of information extracted from performance data - They require high computational time (during training).
without knowing the details of the gas-path system. - If operational data is used to develop the methods, retraining will be
- They can handle multiple fault problems [83]. required during engine overhaul and/or its operating condition changes
- They can deal with measurement uncertainty (noise and bias) [9]. [5].
- They can be applied for onboard applications due to their high computational - Diagnostic error increases with an increase in the number of operating
speed [11]. points. But this problem can be solved by correcting the data towards
- Able to deal with large component faults in the presence of sensor faults [83]. ambient condition and load variations [5].
- They are more appropriate to handle diagnostics problems with limited numbers - The process in the hidden layers is not visible (they are black-box
of measurements [1]. models).
- They are relatively easy to implement for both engine performance modeling and
diagnostics [187].
FL - Like other AI methods, they are model-free, can handle diagnostic problems - Fuzzy rules depend on the knowledge of subject expert and diagnosis
which are difficult to be described and solved mathematically. accuracy depends on the available rules.
- Fuzzy rules are derived from the available GT data for training. This makes them - Large quantity of rules and training data sets are required [2].
particularly suited for finding solutions to problems for which there are no exact - They are not as effective in quantitative fault diagnostics as ANN [5]
solution, but only with a large number of examples [127]. - Like other AI systems, they are poor in recognizing new datasets that
- They are capable of generalizing from examples [127]. the model doesn’t see during training [127]. Hence, for a better
- They can handle multiple fault problems partially [9]. accuracy, a massive amount of data is necessary.
- They are able to deal with the non-linear nature of gas-path problems [11]. - Does not admit model robustness and
- They are coping with measurement noise and sensor bias [126]. - Difficult to define exact queries that identify specific faults
- Like ANNs, FLs could perform data-fusion [2].
- Like ANN, FL based systems could be utilized for online diagnosis application
due to the fast computational capability in inference mode [11].
GA - It is able to deal with measurement noise and bias. - It requires long computational time than traditional optimization
- It provides good results when integrated with other MB as well as DD systems techniques especially as population and generation numbers grow [106].
[1]. - As the number of simultaneous faults increases, the convergence time
- Like ANN, it can deal with engine diagnostic problems with limited increases[128].
instrumentation suites available. - In order to check the consistence in the GA optimization results,
- It can optimizing the engine performance functions without the need of solving multiple runs are often required [5].
complex equations like the other mathematical optimization approaches [175].
- It can perform both qualitative and quantitative diagnostics [5].
- It can perform simultaneous fault analysis [188].
- It can deal with the non-linear nature of the engine behavior [175]
1112 Table 5. (Continued…).

BBN - It is graphical, easy-to-visualize the model variables - High computational requirement [179].
and its diagnostic results - Gathering the required information for setting a BBN model is tedious and time consuming [128]
- Capable of performing multiple simultaneous fault - It requires a well-trained developer to set it up [189].
analysis [189] - As the number of nodes and edges increase the model complexity and computational and storage
- Can perform data fusion [2] requirements increase [190].
- Can provide better, more flexible and robust diagnostic - Measurement uncertainties and operating condition variations affect its accuracy [191].
solutions [182]
- Unlike ANN, it requires a short retraining time because
gas turbine model hardware changes can be easily
entered [127]
- Unlike ANN, BBN can include generic faults that are
not considered during the training process of the
diagnostic system [189].
- After completing the BBN model utilizing a gas turbine
simulation model, it does not need to run the model
gaining in computational speed [127]
- It is more realistic to make diagnosis expressing the
probability of whether a fault occurred or not, than
expressing a deterministic answer. [10]
ES - Preferable when the diagnostic problem is well - The knowledge base may suffer from measurement uncertainties and incomplete or missing data.
understood, stable, and human experts are available to This problem could have a significant effect on reliability and accuracy of answers provided by the
develop the knowledgebase. system.
- Relatively, simple to develop and easy to understand. - Like any software program, the inference engine running missing rules or incorrect data process
- More suitable technique for stable and predictable GT influences its accuracy and reliability.
operating conditions, if potential faults can be defined - They are highly dependent on the knowledgebase, which covers only a small amount of knowledge,
easily. and are incapable of dealing with problems outside this domain. If experts are wrong, the output
- The knowledgebase does not represent only a single will also be wrong. Thus, it requires up-to-date knowledge of the human experts along with a
human expert knowledge, rather a group as a whole. significant set of rules.
This helps to eliminate biases from individuals [59]. - The knowledge acquisition task is time consuming.
- It is more important for applications in remote areas that - They require high establishment cost due to luck of human experts [59].
maintenance experts cannot be found in nearby distance
[59].
1113
1114 4.5. Hybrid Methods

1115 In spite of the fact that it is impossible to find single-technique which can undertake all the
1116 concerns of diagnostics, it would be fascinating to combine two or more methods in an attempt to
1117 offset the limitations of one with the advantages of another [1]. With regard to this, many combined
1118 AI techniques are proposed by different authors including Genetic-neural network, Genetic-fuzzy
1119 logic, fuzzy–neural networks (FNNs), and neural–fuzzy systems [192, 193]. Green and Allen [194]
1120 discussed the benefits of combining ANN with other AI methods for diagnostics and prognostics.
1121 GA is applied as an effective optimization tool to obtain a set of components parameters that produce
1122 a set of predicted dependent parameters, through a non-linear GT model that leads to predictions
1123 which best match the measurements [92]. In FL based techniques, optimal selection of fuzzy sets with
1124 the appropriate membership functions (MFs) is very essential. However, in the conventional FL
1125 applications, there is no defined function to determine the number of fuzzy sets and MFs [112]. It is
1126 indicated in here that GAs are capable of selecting the optimal number of fuzzy sets and MFs
1127 automatically, thereby enhancing the performance of the FL. Kobayashi and Simon [195] investigated
1128 the effectiveness of a MB hybrid neural network GA technique for a turbofan engine application. The
1129 neural network scheme was utilized to diagnose GT component faults while the GA scheme is used
1130 for sensor fault evaluation. The sensor and component fault diagnostic potential of a hybrid ANN
1131 and GA based model was also developed and implemented on advanced cycle Intercooled
1132 Recuperated (ICR) WR21 engine by Sampath and Singh [193]. In this model, the task of the ANN
1133 module was to pre-process and validate the GT data whereas the GA module was to isolate and
1134 identify faults. It has been shown that the accuracy, reliability, and consistency of the results obtained
1135 from the hybrid technique are better than ANN and GA based algorithms. A RBFN was applied for
1136 a GT fault identification purpose, integrated with a MB KF fault diagnosis scheme by Simani and
1137 Fantuzzi [159]. In this algorithm, the KF based scheme was accountable to detect and isolate the faults
1138 using measurement residual, while the RBFN quantified the magnitude of the faults based on pattern
1139 recognition approach.
1140 Table 6 presents the summary of these methods regarding their capability of undertaking the
1141 challenges discussed in section 1.3.1 and fulfilling the desirable attributes of an effective diagnostic
1142 system stated in section 1.3.2. It gives a clear view of the pros and cons of the aforementioned
1143 diagnostic methods, and it is important to choose an appropriate diagnostics system for a particular
1144 diagnostics problem.
1145
1146 Table 6. Comparative summary of the major diagnostic methods concerning their capability towards the desirable attributes.
Category Coping with Ability to deal No. of sensors Data fusion Computational SFI capability Ability to
noise & bias with problem required [9] ability [2, 9] speed [2, 126] [9] provide
Method
[9, 126] non-linearity [9, quantitative
126] solutions [2]
LGPA MB No Incapable M≥N L H Capable Capable
NLGPA MB No Capable M≥N L FH Capable Capable
LKF MB Partially Incapable N<M AA H Capable Capable
NLKF MB Partially Capable N<M AA H Capable Capable
ANN DD Yes Capable N<M FH H Capable Capable
GA MB/AI Yes Capable N<M FH L Capable Capable
FL DD Yes Capable N<M H H Capable Partial
BBN DD Yes Capable N<M H L Capable Capable
ES DD Yes Capable N<M Fairly high FH Capable Capable
Method System MFII Explanation Adaptability Memory Online/offline User friendly Flexibility [2]
complexity Capability [9, facility [88] [88] requirement [9, application [9] interface [11]
[126] 88] 88]
LGPA L Capable - NF L Offline F H
NLGPA L Capable - NF L Offline F H
LKF FL Capable NF NF L Both F AA
NLKF Mm Capable NF NF L Both F AA
ANN FH Partial NF F H Both F L
GA FH Partial - F H Offline NF AA
FL H Partial - F H Both F A
BBN H Partial NF F H Offline F F
ES H Partial NF F H Offline F A
1147
1148 L, low; Mm, medium; AA, above average; FL, fairly low; FH, fairly high; H, high; F, favorable; NF, not favorable; N, number of performance parameters; M, number
1149 of measurement parameters.
1150 5. GT Diagnostic Methods Validation Techniques

1151 In the field of GT diagnostics, method validation a critical issue. GT users need a practical tool
1152 to evaluate the performance and effectiveness of any algorithm before they decide to incorporate into
1153 their plants. However, there are no standardized commercial tool for this purpose. The evaluation
1154 approaches that the research community is using until now are:
1155 1. Performance metrics approach: Performance metrics can be used to measure the detection,
1156 isolation, and identification performance of a fault diagnostics algorithm [196]. The detection metric
1157 measures how accurately the detection algorithm detects abnormal operating conditions. The
1158 isolation metrics evaluates how successfully the isolation part of the diagnostic framework
1159 distinguishes the fault types and their locations. And, the identification metrics measures how
1160 accurately the diagnostics system estimates the magnitude of the faults. A more detailed description
1161 about this concern together with sample performance metrics is available in [197]. Majority of the
1162 fault diagnostic methods available in the open domain are evaluated based on this approach. A fault
1163 diagnostics performance metric associated with the fault detection and isolation reliability of four
1164 different methods for aircraft engines is presented by Simon et al. [198].
1165 2. Benchmark Fault Cases Approach: The generally accepted and implemented solution to obtain
1166 the required performance data for diagnostic method development and validation is implanting fault
1167 cases corresponding to the possibly existing faults into the GT performance model [114, 199].
1168 However, there exist some inconsistency concerning the range of the performance parameter losses
1169 that different gas-path faults are represented by [200]. The issue of using benchmark fault cases has
1170 also been thoroughly studied by the OBIDICOTE (On Board Identification, Diagnosis and Control of
1171 GT Engines) Project conducted by the European research community [201]. The project identified a
1172 set of benchmark fault cases, which have been used by several researchers so far to evaluate their
1173 engine fault diagnostic methods [182, 193, 202]. The effectiveness of using some sets of benchmark
1174 fault cases to evaluate the performance of a diagnostic system is further investigated by the engine
1175 health management industry review (EHMIR) established under The Technical Cooperation Program
1176 (TTCP) [202]. TTCP is a collaboration forum for defense science and technology (DST) between five
1177 nations, namely, UK, USA, Canada, Australia and New Zealand. The effort of this forum resulted in
1178 a reference engine problem together with a recommendation of an evaluation environment for
1179 different diagnostics algorithms. Based on the recommendation of the aircraft engine health
1180 monitoring community, recently, NASA’s research team developed a public benchmark gas-path
1181 fault diagnostics techniques’ performance evaluation software referred to as the Propulsion
1182 Diagnostic Method Evaluation Strategy (ProDiMES), applicable to aircraft engines [198]. In this
1183 software four different methods are included; Weighted Least Squares, PNN, Performance Analysis
1184 Tool, and Generalized Observer/Estimator. Moreover, the field GT condition monitoring and
1185 diagnostics has been studied for many years by the research team of the Laboratory of Thermal
1186 Turbomachines of the National Technical University of Athens (LTT/NTUA) [12] and Cranfield
1187 University (CU) [1] and proposed many different techniques. These groups showed the effectiveness
1188 of using benchmark fault cases to develop and evaluate the performance of diagnostic algorithms.
1189 3. Comparison of Methods Approach: As a third alternative, there are also some self-conducted
1190 comparative evaluation (a one-to-one comparison of methods) based research works to assess the
1191 diagnostic performance of different themes [114, 199]. In this category, previously published papers
1192 are used as a benchmark methods to compare the performance of the newly developed algorithm.
1193 However, this approach has limitations due to the reason that most of the available GT diagnostic
1194 methods targeted different engine problems and degree of complexity under variety of diagnostic
1195 conditions (i.e., operating modes, measurement system, deterioration profile etc.). [202].
1196 6. Conclusions and Future Research Directions

1197 Gas turbine (GT) gas-path fault diagnostics is a key element of an overall engine condition-based
1198 monitoring (CBM) system providing an enhanced safety, reliability, and availability along with
1199 optimal operation and maintenance costs. In consideration of this role, an effective and reliable gas-
1200 path diagnostic technique is critical. This paper was devoted to discussing various issues related to
1201 gas-path diagnostics including engine physical faults, challenges and desirable attributes of a gas-
1202 path diagnostics, state-of-the-art methods, and verification and validation approaches. The past
1203 efforts on gas-path diagnostics have focused on the aspects of data filtration, sensor validation, and
1204 component fault diagnostics. Variety of methods associated with these aspects, beginning with the
1205 conventional methods to the most sophisticated artificial intelligence (AI) based ones, were reviewed.
1206 Due to their remarkable capability of handling the available challenges and meet the majority of the
1207 desirable attributes, recent efforts paid more attention to AI methods. Particularly, artificial neural
1208 networks (ANNs) have been widely used for both qualitative and quantitative diagnostic
1209 applications, although the majority of the investigations were limited to single sensor fault and/or
1210 single component fault analysis. However, in order to avoid the barriers between system developers
1211 and engine users and get their interest to invest and incorporate gas-path diagnostic technologies
1212 into their plants, there are two main issues that the gas-path diagnostics research community should
1213 give attention: improving the effectiveness and reliability of the available fault diagnostic systems
1214 and developing practical tools to evaluate the effectiveness of the proposed techniques. With that in
1215 mind, the following further studies should be carried out.
1216  The need for a standardized gas-path diagnostic problem definition. According to this survey,
1217 there is no consensus between researchers in defining and representing gas-path diagnostic
1218 problems (terminologies, component fault representation, ranges of sensor/component faults,
1219 and the number and type of faults corresponding to different engine configurations that possibly
1220 exist in the engines life time). This inconsistency may confuse young researchers of the field,
1221 create barriers in exchanging gas-path diagnostic related ideas/solutions and performing a one-
1222 to-one comparison of different algorithms.
1223  The review on GT performance deterioration revealed that the degradation profile corresponding
1224 to each gas-path faults is not consistent. This may lead to an incorrect representation of
1225 components deterioration, and thereby unreliable fault diagnostic results. Hence, there should
1226 be more investigations in this regard.
1227  Most of the devised techniques for simultaneous fault analysis were restricted to qualitative
1228 solutions (i.e., detecting and isolating without estimating the level of the fault, which is the very
1229 important step in the maintenance decision). Moreover, the accuracy of the available limited
1230 quantitative approaches requires improvement for multiple fault scenarios. Development of an
1231 effective gas-path diagnostic system that can perform both qualitative and quantitative
1232 diagnostics of both single and multiple fault scenarios thus needs further investigation.
1233  Development of efficient hybrid methods. Most of the available gas-path diagnostic methods are
1234 single-technique-based and it is difficult to find single-technique which can address all gas-path
1235 diagnostic related challenges along with providing accurate diagnostic results. It is
1236 recommended to combine two or more methods based on their merits.
1237  Development of integrated platforms. Although a large number of diagnostic methods have been
1238 devised so far, the majority of those methods considered different platforms with different levels
1239 of complexity and applied for different engine system monitoring (such as sensor, component,
1240 vibration, controller, and fuel and oil systems). Integration of verity of methods into a diagnostic
1241 tool being capable of addressing the entire GT system problems is required.
1242  Establishment of a practical approach to verification and validation. Engine users need practical
1243 tools to objectively assess the effectiveness (i.e., the technical and economic feasibility) of newly
1244 proposed solutions and determine its advantages over the existing maintenance practices before
1245 incorporating into their plants. However, yet there are no internationally accepted standards or
1246 generic software tools that can be applied for this purpose. Hence, the establishment of a practical
1247 verification and validation approaches requires attention in this field.
1248  Development of a user-friendly gas-path diagnostic software. Regarding engine performance
1249 simulation, there are some powerful commercial software available. Conversely, other than the
1250 traditional techniques, there are no advanced software tools based on AI methods. Therefore, a
1251 user-friendly gas-path diagnostic software that can acquire, preprocess and validate performance
1252 data, assess the condition of the engine and suggest the appropriate maintenance action is
1253 required. Writing-review and editing.
1254
1255 Author contributions: ADF wrote the manuscript and ATB, SIG and KGK supervised and made
1256 some critical revisions. All authors have read and approved the final manuscript.
1257
1258 Conflict of Interest: The authors declare no conflict of interest.
1259
1260 Acknowledgment: The authors would like to acknowledge Universiti Teknologi PETRONAS (UTP)
1261 for supporting this research financially as well as for all the facilities provided.
1262
1263 Reference
1264 [1] R. Singh, "Advances and opportunities in gas path diagnostics," 15th ISABE, Paper No. ISABE-2003-1008,
1265 2003.
1266 [2] G. Bechini, "Performance diagnostics and measurement selection for on-line monitoring of gas turbine
1267 engines," PhD Thesis, Cranfield University, 2007.
1268 [3] I. A. T. Association, "Airline Maintenance Cost: Executive Commentary," IATA.
1269 Web.<http://www.iata.org/workgroups/Documents/MCTF/AMC_ExecComment_FY09. pdf, 2011.
1270 [4] M. McFadden and D. S. Worrells, "Global outsourcing of aircraft maintenance," Journal of Aviation
1271 Technology and Engineering, vol. 1, p. 4, 2012.
1272 [5] S. O. T.-o. Ogaji, "Advanced Gas-path Fault Diagnostics for Stationary Gas Turbines," Ph.D. Thesis,
1273 Cranfield University, UK, 2003.
1274 [6] M. P. Boyce, Gas turbine engineering handbook: Elsevier, 2011.
1275 [7] C. Factbook, "The world factbook," https://www. cia. gov/library/publications/the-world-factbook, 2010.
1276 [8] G. J. Lengyel, "Department of Defense Energy Strategy: Teaching an old dog new tricks," DTIC
1277 Document2007.
1278 [9] L. Marinai, D. Probert, and R. Singh, "Prospects for aero gas-turbine diagnostics: A review," Applied Energy,
1279 vol. 79, pp. 109-126, 2004.
1280 [10] A. G. Stamatis, "Engine condition monitoring and diagnostics," in Progress in gas turbine performance, E.
1281 Benini, Ed., ed Rijeka, Croatia: InTech, 2013, pp. 187–212.
1282 [11] M. Tahan, E. Tsoutsanis, M. Muhammad, and Z. A. Karim, "Performance-based health monitoring,
1283 diagnostics and prognostics for condition-based maintenance of gas turbines: A review," Applied Energy,
1284 vol. 198, pp. 122-144, 2017.
1285 [12] C. Romesis, N. Aretakis, I. Roumeliotis, A. Alexiou, A. Tsalavoutas, A. Stamatis, et al., "Experiance with
1286 condition-based maintenance related methods and tools for gas turbines," in The future of Gas Turbine
1287 Technology, 7th International Gas Turbine Conference, Brussels, Belgium, 2014.
1288 [13] C. B. Meher-Homji, M. Chaker, and H. Motiwalla, "Gas turbine performance deterioration," in Proceedings
1289 of the 30th Turbomachinery Symposium, 2001, pp. 17-20.
1290 [14] C. B. Meher-Homji, T. Matthews, A. Pelagotti, and H. Weyermann, "Gas Turbines and Turbocompressors
1291 for LNG service," in Proceedings of the 36th Turbomachinery symposium, 2007.
1292 [15] L. Marinai, R. Singh, B. Curnock, and D. Probert, "Detection and prediction of the performance
1293 deterioration of a turbofan engine," in Proceedings of the International Gas Turbine Congress, 2003, pp. 2-7.
1294 [16] L. A. Urban, "Parameter selection for multiple fault diagnostics of gas turbine engines," ASME J. Eng. Power,
1295 vol. 97, pp. 225-230, 1975.
1296 [17] C. B. Meher-Homji, M. Chaker, and A. F. Bromley, "The fouling of axial flow compressors: causes, effects,
1297 susceptibility, and sensitivity," in ASME Turbo Expo 2009: Power for Land, Sea, and Air, 2009, pp. 571-590.
1298 [18] I. S. Diakunchak, "Performance deterioration in industrial gas turbines," Journal of Engineering for Gas
1299 Turbines and Power, vol. 114, pp. 161-168, 1992.
1300 [19] M. Morini, M. Pinelli, P. Spina, and M. Venturini, "Influence of blade deterioration on compressor and
1301 turbine performance," Journal of engineering for gas turbines and power, vol. 132, p. 032401, 2010.
1302 [20] F. Melino, M. Morini, A. Peretto, M. Pinelli, and P. R. Spina, "Compressor fouling modeling: relationship
1303 between computational roughness and gas turbine operation time," Journal of Engineering for Gas Turbines
1304 and Power, vol. 134, p. 052401, 2012.
1305 [21] H. I. H. Saravanamuttoo and A. N. Lakshminarasimha, "A preliminary assessment of compressor fouling,"
1306 1985.
1307 [22] N. Aretakis, I. Roumeliotis, and K. Mathioudakis, "Performance model “zooming” for in-depth component
1308 fault diagnosis," Journal of engineering for gas turbines and power, vol. 133, p. 031602, 2011.
1309 [23] A. Lakshminarasimha, M. Boyce, and C. Meher-Homji, "Modelling and analysis of gas turbine performance
1310 deterioration," in ASME 1992 International Gas Turbine and Aeroengine Congress and Exposition, 1992, pp.
1311 V004T10A022-V004T10A022.
1312 [24] C. B. Meher-Homji and A. Bromley, "Gas turbine axial compressor fouling and washing," in 33rd
1313 Turbomachinery Symposium, Houston, TX, September, 2004, pp. 20-23.
1314 [25] A. Zwebek and P. Pilidis, "Degradation effects on combined cycle power plant performance - Part I: Gas
1315 turbine cycle component degradation effects," Journal of Engineering for Gas Turbines and Power, vol. 125, pp.
1316 651-657, 2003.
1317 [26] A. N. Lakshminarasimha, M. P. Boyce, and C. B. Meher-Homji, "Modelling and analysis of gas turbine
1318 performance deterioration," in ASME 1992 International Gas Turbine and Aeroengine Congress and Exposition,
1319 GT 1992, 1992.
1320 [27] R. Kurz and K. Brun, "Gas Turbine Performance—What Makes the Map," in 29th Texas A&M
1321 Turbomachinery Symposium, Houston, TX, September, 2000, pp. 18-21.
1322 [28] A. P. Tarabrin, V. A. Schurovsky, A. I. Bodrov, and J. P. Stalder, "An analysis of axial compressor fouling
1323 and a blade cleaning method," Journal of Turbomachinery, vol. 120, pp. 256-261, 1998.
1324 [29] E. A. Ogiriki, Y. G. Li, T. Nikolaidis, T. E. Isaiah, and G. Sule, "Effect of fouling, thermal barrier coating
1325 degradation and film cooling holes blockage on gas turbine engine creep life," Procedia CIRP, vol. 38, pp.
1326 228-233, 2015.
1327 [30] G. F. Aker and H. I. H. Saravanamuttoo, "Predicting gas turbine performance degradation due to
1328 compressor fouling using computer simulation techniques," Journal of Engineering for Gas Turbines and
1329 Power, vol. 111, pp. 343-350, 1989.
1330 [31] H. Yang and H. Xu, "The new performance calculation method of fouled axial flow compressor," The
1331 Scientific World Journal, vol. 2014, 2014.
1332 [32] A. Tarabrin, V. Schurovsky, A. Bodrov, and J.-P. Stalder, "Influence of Axial Compressor Fouling on Gas
1333 Turbine Unit Perfomance Based on Different Schemes and With Different Initial Parameters," in ASME
1334 1998 International Gas Turbine and Aeroengine Congress and Exhibition, 1998, pp. V004T11A006-V004T11A006.
1335 [33] A. D. Mezheritsky and A. V. Sudarev, "The Mechanism of Fouling and the Cleaning Technique in
1336 Application to Flow Parts of the Power Generation Plant Compressors," p. V004T11A003, 1990.
1337 [34] F. Seddigh and H. Saravanamuttoo, "A proposed method for assessing the susceptibility of axial
1338 compressors to fouling," in ASME 1990 International Gas Turbine and Aeroengine Congress and Exposition,
1339 1990, pp. V003T07A012-V003T07A012.
1340 [35] M. P. Boyce and F. Gonzalez, "A study of on-line and off-line turbine washing to optimize the operation of
1341 a gas turbine," Journal of engineering for gas turbines and power, vol. 129, pp. 114-122, 2007.
1342 [36] J. P. Stalder, "Gas Turbine Compressor Washing State of the Art: Field Experiences," Journal of Engineering
1343 for Gas Turbines and Power, vol. 123, p. 363, 2001.
1344 [37] N. Casari, M. Pinelli, A. Suman, L. di Mare, and F. Montomoli, "An Energy-Based Fouling Model for Gas
1345 Turbines: EBFOG," Journal of Turbomachinery, vol. 139, p. 021002, 2017.
1346 [38] B. Maiwada, N. I. Mu'az, S. Ibrahim, and S. M. Musa, "Impacts of Compressor Fouling On the Performance
1347 of Gas Turbine," International Journal of Engineering Science, vol. 2118, 2016.
1348 [39] N. Aretakis, I. Roumeliotis, G. Doumouras, and K. Mathioudakis, "Compressor washing economic analysis
1349 and optimization for power generation," Applied energy, vol. 95, pp. 77-86, 2012.
1350 [40] I. S. Diakunchak, "Performance Improvement in Industrial Gas Turbines," p. V001T01A005, 1993.
1351 [41] M. Wilcox, R. Baldwin, A. Garcia-Hernandez, and K. Brun, "Guideline for gas turbine inlet air filtration
1352 systems," Gas Machinery Research Council, Dallas, TX, 2010.
1353 [42] M. Singh Grewal, "Gas turbine engine performance deterioration modelling and analysis," 1988.
1354 [43] A. Salar, S. M. Hosseini, B. R. Zangmolk, and A. K. Sedigh, "Improving model-based gas turbine fault
1355 diagnosis using multi-operating point method," in Computer Modeling and Simulation (EMS), 2010 Fourth
1356 UKSim European Symposium on, 2010, pp. 240-247.
1357 [44] Y. Qingcai, S. Li, Y. Cao, and N. Zhao, "Full and Part-Load Performance Deterioration Analysis of Industrial
1358 Three-Shaft Gas Turbine Based on Genetic Algorithm," in ASME Turbo Expo 2016: Turbomachinery Technical
1359 Conference and Exposition, Harbin, China, 2016, pp. V006T05A016-V006T05A016.
1360 [45] R. Kurz and K. Brun, "Degradation in Gas Turbine Systems," Journal of Engineering for Gas Turbines and
1361 Power, vol. 123, pp. 70-77, 2000.
1362 [46] R. Kurz and K. Brun, "Gas turbine tutorial—maintenance and operating practices effects on degradation
1363 and life," in 36th Turbomachinery Symposium, Houston, TX, September, 2007, pp. 10-13.
1364 [47] A. Gannan, "Cascade testing and CFD applied to gas turbine performance improvement with compressor
1365 cleaning," 2010.
1366 [48] M. Wilcox, R. Kurz, and K. Brun, "Successful Selection and Operation of Gas Turbine Inlet Filtration
1367 Systems," in Proceedings of the 40th Turbomachinery Symposium, Houston, TX, Sept, 2011, pp. 12-15.
1368 [49] R. Kurz and K. Brun, "Degradation of gas turbine performance in natural gas service," Journal of Natural Gas
1369 Science and Engineering, vol. 1, pp. 95-102, 2009.
1370 [50] R. K. Mishra, "Fouling and Corrosion in an Aero Gas Turbine Compressor," Journal of Failure Analysis and
1371 Prevention, 2015.
1372 [51] A. G. Varelis, "Technoeconomic study of engine deterioration and compressor washing for military gas
1373 turbine engines," 2008.
1374 [52] A. Zwebek, "Combined cycle performance deterioration analysis," 2002.
1375 [53] R. Kurz, K. Brun, and M. Wollie, "Degradation Effects on Industrial Gas Turbines," Journal of Engineering
1376 for Gas Turbines and Power, vol. 131, pp. 062401-062401, 2009.
1377 [54] I. S. Diakunchak, "Performance deterioration in industrial gas turbines," in ASME 1991 International Gas
1378 Turbine and Aeroengine Congress and Exposition, 1991, pp. V004T10A012-V004T10A012.
1379 [55] A. Giebmanns, J. Backhaus, C. Frey, and R. Schnell, "Compressor Leading Edge Sensitivities and Analysis
1380 With an Adjoint Flow Solver," p. V06AT35A009, 2013.
1381 [56] F. Radtke and G. Dibelius, "Reynoldszahleinfluss bei hochbelasteten axialen Turbinenbeschaufelungen,"
1382 VDI-Berichte, 1980.
1383 [57] B. MacIsaac, "Engine performance and health monitoring models using steady state and transient
1384 prediction methods," DTIC Document1992.
1385 [58] T. Giampaolo, The gas turbine handbook: Principles and practices: The Fairmont Press, Inc., 2008.
1386 [59] E. Ntantis, "Capability expansion of non-linear gas path analysis," PhD Thesis, Cranfield University, 2008.
1387 [60] Y. Ying, Y. Cao, S. Li, J. Li, and J. Guo, "Study on gas turbine engine fault diagnostic approach with a hybrid
1388 of gray relation theory and gas-path analysis," Advances in Mechanical Engineering, vol. 8, p.
1389 1687814015627769, 2016.
1390 [61] S. O. T. Ogaji, Y. G. Li, S. Sampath, and R. Singh, "Gas Path Fault Diagnosis of a Turbofan Engine From
1391 Transient Data Using Artificial Neural Networks," pp. 405-414, 2003.
1392 [62] Z. N. Sadough Vanini, K. Khorasani, and N. Meskin, "Fault detection and isolation of a dual spool gas
1393 turbine engine using dynamic neural networks and multiple model approach," Information Sciences, vol.
1394 259, pp. 234-251, 2014.
1395 [63] S. Sina Tayarani-Bathaie and K. Khorasani, "Fault detection and isolation of gas turbine engines using a
1396 bank of neural networks," Journal of Process Control, vol. 36, pp. 22-41, 12// 2015.
1397 [64] Y. Zhang, C. Bingham, M. Garlick, and M. Gallimore, "Applied fault detection and diagnosis for industrial
1398 gas turbine systems," International Journal of Automation and Computing, vol. 14, pp. 463-473, 2017.
1399 [65] W. Wang, Z. Xu, R. Tang, S. Li, and W. Wu, "Fault detection and diagnosis for gas turbines based on a
1400 kernelized information entropy model," The Scientific World Journal, vol. 2014, 2014.
1401 [66] B. Pourbabaee, N. Meskin, and K. Khorasani, "Sensor fault detection, isolation, and identification using
1402 multiple-model-based hybrid Kalman filter for gas turbine engines," IEEE Transactions on Control Systems
1403 Technology, vol. 24, pp. 1184-1200, 2016.
1404 [67] J. H. Taylor and M. Omana, "Fault detection, isolation and accommodation using the generalized parity
1405 vector technique," IFAC Proceedings Volumes, vol. 41, pp. 1914-1921, 2008.
1406 [68] S. Y. Vural and C. Hacızade, "SENSOR/ACTUATOR FAULT DETECTION, ISOLATION AND
1407 ACCOMMODATION APPLIED TO UAV MODEL," Journal of Aeronautics and Space Technologies, pp. 1-
1408 12%V 9, 2016-07-25 2016.
1409 [69] J. Yang, Y. Chen, and L. Zhang, "An Efficient Approach for Fault Detection, Isolation, and Data Recovery
1410 of Self-Validating Multifunctional Sensors," IEEE Transactions on Instrumentation and Measurement, vol. 66,
1411 pp. 543-558, 2017.
1412 [70] S. Simani, "Identification and fault diagnosis of a simulated model of an industrial gas turbine," IEEE
1413 Transactions on Industrial Informatics, vol. 1, pp. 202-216, 2005.
1414 [71] G. Vachtsevanos, F. Lewis, M. Roemer, A. Hess, and B. Wu, "Fault Diagnosis," in Intelligent Fault Diagnosis
1415 and Prognosis for Engineering Systems, ed: John Wiley & Sons, Inc., 2007, pp. 172-283.
1416 [72] D. Simon, "Challenges in aircraft engine gas path health management," Tutorial on Aircraft Engine Control
1417 and Gas Path Health Management Presented at, 2012.
1418 [73] H. Khorasgani, D. E. Jung, G. Biswas, E. Frisk, and M. Krysander, "Robust residual selection for fault
1419 detection," in Decision and Control (CDC), 2014 IEEE 53rd Annual Conference on, 2014, pp. 5764-5769.
1420 [74] S. O. T. Ogaji, R. Singh, and S. D. Probert, "Multiple-sensor fault-diagnoses for a 2-shaft stationary gas-
1421 turbine," Applied Energy, vol. 71, pp. 321-339, 4// 2002.
1422 [75] Z. N. Sadough Vanini, N. Meskin, and K. Khorasani, "Multiple-Model Sensor and Components Fault
1423 Diagnosis in Gas Turbine Engines Using Autoassociative Neural Networks," Journal of Engineering for Gas
1424 Turbines and Power, vol. 136, p. 091603, 2014.
1425 [76] D. Zhou, H. Zhang, and S. Weng, "A new gas path fault diagnostic method of gas turbine based on support
1426 vector machine," Journal of Engineering for Gas Turbines and Power, vol. 137, p. 102605, 2015.
1427 [77] L. C. Jaw and Y.-J. Lee, "Engine diagnostics in the eyes of machine learning," in ASME Turbo Expo 2014:
1428 Turbine Technical Conference and Exposition, Düsseldorf, Germany, 2014, pp. V006T06A029-V006T06A029.
1429 [78] R. Abernathy, B. D. Powell, D. L. Colbert, D. G. Sanders, and J. Thompson Jr, "Handbook, Uncertainty in
1430 Gas Turbine Measurements," DTIC Document1973.
1431 [79] R. Abernethy and J. Thompson Jr, "UNCERTAINTYIN GAS TURBINE MEASUREMENTS."
1432 [80] M. S. Jasmani, Y.-G. Li, and Z. Ariffin, "Measurement selections for multicomponent gas path diagnostics
1433 using analytical approach and measurement subset concept," Journal of Engineering for Gas Turbines and
1434 Power, vol. 133, p. 111701, 2011.
1435 [81] A. J. Volponi, "Gas Turbine Engine Health Management: Past, Present, and Future Trends," Journal of
1436 Engineering for Gas Turbines and Power, vol. 136, p. 051201, 2014.
1437 [82] D. L. Simon, S. Garg, G. W. Hunter, T.-H. Guo, and K. J. Semega, "Sensor needs for control and health
1438 management of intelligent aircraft engines," ASME Paper No. GT2004-54324, 2004.
1439 [83] S. O. T. Ogaji and R. Singh, "Advanced engine diagnostics using artificial neural networks," Applied Soft
1440 Computing Journal, vol. 3, pp. 259-271, 2003.
1441 [84] J. Kurzke, "Model Based Gas Turbine Parameter Corrections," pp. 91-99, 2003.
1442 [85] L. C. Jaw, "Recent advancements in aircraft engine health management (EHM) technologies and
1443 recommendations for the next step," in ASME turbo expo 2005: power for land, sea, and air, 2005, pp. 683-695.
1444 [86] S. O. T. Ogaji, S. Sampath, R. Singh, and S. D. Probert, "Parameter selection for diagnosing a gas-turbine's
1445 performance-deterioration," Applied Energy, vol. 73, pp. 25-46, 2002.
1446 [87] L. Qiu, P. Bahl, L. Zhou, and A. R. Rao, "Fault detection and diagnosis," ed: Google Patents, 2009.
1447 [88] S. Dash and V. Venkatasubramanian, "Challenges in the industrial applications of fault diagnostic systems,"
1448 Computers & chemical engineering, vol. 24, pp. 785-791, 2000.
1449 [89] V. Venkatasubramanian, R. Rengaswamy, K. Yin, and S. N. Kavuri, "A review of process fault detection
1450 and diagnosis: Part I: Quantitative model-based methods," Computers & chemical engineering, vol. 27, pp.
1451 293-311, 2003.
1452 [90] M. Fast, M. Assadi, and S. De, "Development and multi-utility of an ANN model for an industrial gas
1453 turbine," Applied Energy, vol. 86, pp. 9-17, 1// 2009.
1454 [91] R. Verma, N. Roy, and R. Ganguli, "Gas turbine diagnostics using a soft computing approach," Applied
1455 Mathematics and Computation, vol. 172, pp. 1342-1363, 1/15/ 2006.
1456 [92] Y. G. Li, "Performance-analysis-based gas turbine diagnostics: A review," Proceedings of the Institution of
1457 Mechanical Engineers, Part A: Journal of Power and Energy, vol. 216, pp. 363-377, 2002.
1458 [93] A. K. S. Jardine, D. Lin, and D. Banjevic, "A review on machinery diagnostics and prognostics implementing
1459 condition-based maintenance," Mechanical Systems and Signal Processing, vol. 20, pp. 1483-1510, 2006.
1460 [94] C. Kong, "Review on Advanced Health Monitoring Methods for Aero Gas Turbines using Model Based
1461 Methods and Artificial Intelligent Methods," International Journal of Aeronautical and Space Sciences, vol. 15,
1462 pp. 123-137, 2014.
1463 [95] N. Zhao, X. Wen, and S. Li, "A Review on Gas Turbine Anomaly Detection for Implementing Health
1464 Management," in ASME Turbo Expo 2016: Turbomachinery Technical Conference and Exposition, Harbin, China,
1465 2016, pp. V001T22A009-V001T22A009.
1466 [96] M. Zedda, "Gas turbine engine and sensor fault diagnosis," 1999.
1467 [97] L. A. Urban, "Gas Path Analysis Applied to Turbine Engine Condition Monitoring," Journal of Aircraft, vol.
1468 10, pp. 400-406, 1973.
1469 [98] A. J. Volponi, "Sensor error compensation in engine performance diagnostics," in ASME 1994 International
1470 Gas Turbine and Aeroengine Congress and Exposition, 1994, pp. V005T15A008-V005T15A008.
1471 [99] D. L. Doel, "TEMPER: a gas-path analysis tool for commercial jet engines," in ASME 1992 International Gas
1472 Turbine and Aeroengine Congress and Exposition, 1992, pp. V005T15A013-V005T15A013.
1473 [100] M. Provost, "The use of optimal estimation techniques in the analysis of gas turbines," 1994.
1474 [101] P. Escher, "Pythia: An object-orientated gas path analysis computer program for general applications,"
1475 1995.
1476 [102] M. Provost and R. Singh, "Gas-path analysis: preparing for success," ROLLS ROYCE PLC-REPORT-PNR,
1477 1995.
1478 [103] A. Gulati, "An optimization tool for gas turbine engine diagnostics," PhD Thesis, Cranfield University, 2001.
1479 [104] Y. G. Li and P. Pilidis, "GA-based design-point performance adaptation and its comparison with ICM-
1480 based approach," Applied Energy, vol. 87, pp. 340-348, 2010.
1481 [105] P. C. Escher, "Gas turbine data validation using gas path analysis," in ASME Turbo Expo 2002: Power for
1482 Land, Sea, and Air, 2002, pp. 45-51.
1483 [106] Y. Li, "Gas turbine performance and health status estimation using adaptive gas path analysis," Journal of
1485 [107] A. G. Stamatis, "Evaluation of gas path analysis methods for gas turbine diagnosis," Journal of Mechanical
1486 Science and Technology, vol. 25, pp. 469-477, 2011.
1487 [108] E. Larsson, "Model Based Diagnosis and Supervision of Industrial Gas Turbines PDF," 2014.
1488 [109] M. Chen, L. Quan Hu, and H. Tang, "An Approach for Optimal Measurements Selection on Gas Turbine
1489 Engine Fault Diagnosis," Journal of Engineering for Gas Turbines and Power, vol. 137, 2015.
1490 [110] R. E. Kalman, "A new approach to linear filtering and prediction problems," Journal of basic Engineering, vol.
1491 82, pp. 35-45, 1960.
1492 [111] R. E. Kalman and R. S. Bucy, "New results in linear filtering and prediction theory," Journal of basic
1493 engineering, vol. 83, pp. 95-108, 1961.
1494 [112] R. Ganguli, Gas Turbine Diagnostics: Signal Processing and Fault Isolation: CRC press, 2012.
1495 [113] G. Bishop and G. Welch, "An introduction to the kalman filter," Proc of SIGGRAPH, Course, vol. 8, p. 41,
1496 2001.
1497 [114] A. J. Volponi, H. DePold, R. Ganguli, and C. Daguang, "The Use of Kalman Filter and Neural Network
1498 Methodologies in Gas Turbine Performance Diagnostics: A Comparative Study," Journal of Engineering for
1499 Gas Turbines and Power, vol. 125, pp. 917-924, 2003.
1500 [115] W. S. Cooper, "Use of optimal estimation theory, in particular the Kalman filter, in data analysis and signal
1501 processing," Review of scientific instruments, vol. 57, pp. 2862-2869, 1986.
1502 [116] P. Dewallef, C. Romessis, O. Léonard, and K. Mathioudakis, "Combining Classification Techniques With
1503 Kalman Filters for Aircraft Engine Diagnostics," Journal of Engineering for Gas Turbines and Power, vol. 128,
1504 p. 281, 2006.
1505 [117] A. MZWINSKI, "Stochastic process and filtering theory'," ed: Academic Press, 1970.
1506 [118] G. Welch and G. Bishop, "An introduction to the Kalman filter," 2006.
1507 [119] M. Provost, "The use of optimal estimation techniques in the analysis of gas turbines," PhD Thesis,
1508 Cranfield University, 1994.
1509 [120] M. Zedda and R. Singh, "Gas Turbine Engine and Sensor Fault Diagnosis Using Optimization Techniques,"
1510 Journal of Propulsion and Power, vol. 18, pp. 1019-1025, 2002.
1511 [121] T. Kobayashi and D. L. Simon, "Application of a bank of Kalman filters for aircraft engine fault diagnostics,"
1512 in ASME Turbo Expo 2003, collocated with the 2003 International Joint Power Generation Conference, Atlanta,
1513 Georgia, USA,, 2003, pp. 461-470.
1514 [122] D. L. Simon and A. W. Rinehart, "Sensor Selection for Aircraft Engine Performance Estimation and Gas
1515 Path Fault Diagnostics," Journal of Engineering for Gas Turbines and Power, vol. 138, p. 071201, 2016.
1516 [123] S. Borguet, P. Dewallef, and O. Léonard, "A way to deal with model-plant mismatch for a reliable diagnosis
1517 in transient operation," Journal of Engineering for Gas Turbines and Power, vol. 130, p. 031601, 2008.
1518 [124] N. Meskin, E. Naderi, and K. Khorasani, "A multiple model-based approach for fault diagnosis of jet
1519 engines," IEEE Transactions on Control Systems Technology, vol. 21, pp. 254-262, 2013.
1520 [125] F. Lu, H. Ju, and J. Huang, "An improved extended Kalman filter with inequality constraints for gas turbine
1521 engine health monitoring," Aerospace Science and Technology, vol. 58, pp. 36-47, 2016.
1522 [126] Y. Li, "Performance-analysis-based gas turbine diagnostics: A review," Proceedings of the Institution of
1523 Mechanical Engineers, Part A: Journal of Power and Energy, vol. 216, pp. 363-377, 2002.
1524 [127] L. Marinai, "Gas-path diagnostics and prognostics for aero-engines using fuzzy logic and time series
1525 analysis," PhD Thesis, Cranfield University, 2004.
1526 [128] S. Sampath, "Fault diagnostics for advanced cycle marine gas turbine using genetic algorithm," PhD Thesis,
1527 Cranfield University, 2003.
1528 [129] M. Zedda, "Gas turbine engine and sensor fault diagnosis," PhD Thesis, Cranfield University, 1999.
1529 [130] Q. Yang, S. Li, and Y. Cao, "A strong tracking filter based multiple model approach for gas turbine fault
1530 diagnosis," Journal of Mechanical Science and Technology, vol. 32, pp. 465-479, 2018.
1531 [131] A. Konar, "Artificial intelligence and soft computing: behavioral and cognitive modeling of the human
1532 brain," 2000.
1533 [132] D. L. Simon, "Propulsion Diagnostic Method Evaluation Strategy (ProDiMES) User's Guide," 2010.
1534 [133] O. Awodele and O. Jegede, "Neural networks and its application in engineering."
1535 [134] H. Asgari, X. Chen, M. B. Menhaj, and R. Sainudiin, "Artificial Neural Network–Based System
1536 Identification for a Single-Shaft Gas Turbine," Journal of Engineering for Gas Turbines and Power, vol. 135, p.
1537 092601, 2013.
1538 [135] H. Demuth, M. Beale, and M. Hagan, "Neural Network Toolbox for Use with MATLAB User’s Guide,
1539 version 4.0. 6, ninth printing revised for version 4.0. 6 (release 14SP3)," The MathWorks: Natick, MA, 2015a.
1540 [136] C. d. Kong, J. y. Ki, M. c. Kang, and S. h. Kho, "Intelligent performance diagnostics of a gas turbine engine
1541 using user-friendly interface neural networks," Aircraft Engineering and Aerospace Technology, vol. 76, pp.
1542 391-397, 2004.
1543 [137] G. R. Matuck, J. o. R. Barbosa, C. Bringhenti, and I. Lima, "Multiple Faults Detection of Gas Turbine by MLP
1544 Neural Network," pp. 697-703, 2009.
1545 [138] K. Patan, M. Witczak, and J. Korbicz, "Towards robustness in neural network based fault diagnosis,"
1546 International Journal of Applied Mathematics and Computer Science, vol. 18, pp. 443-454, 2008.
1547 [139] S. S. Tayarani-Bathaie, Z. S. Vanini, and K. Khorasani, "Dynamic neural network-based fault diagnosis of
1548 gas turbine engines," Neurocomputing, vol. 125, pp. 153-165, 2014.
1549 [140] R. Mohammadi, E. Naderi, K. Khorasani, and S. Hashtrudi-Zad, "Fault diagnosis of gas turbine engines by
1550 using dynamic neural networks," in 2011 IEEE International Conference on Quality and Reliability, ICQR 2011,
1551 2011, pp. 25-30.
1552 [141] S. Kiakojoori and K. Khorasani, "Dynamie neural networks for jet engine degradation prediction and
1553 prognosis," in Neural Networks (IJCNN), 2014 International Joint Conference on, Beijing, China, 2014, pp. 2531-
1554 2538.
1555 [142] M. Amozegar and K. Khorasani, "An ensemble of dynamic neural network identifiers for fault detection
1556 and isolation of gas turbine engines," Neural Networks, vol. 76, pp. 106-121, 4// 2016.
1557 [143] M. Tahan, M. Muhammad, and Z. A. Karim, "A multi-nets ANN model for real-time performance-based
1558 automatic fault diagnosis of industrial gas turbine engines," Journal of the Brazilian Society of Mechanical
1559 Sciences and Engineering, vol. 39, pp. 2865-2876, 2017.
1560 [144] Z. M. Zin, "Using auto-associative neural networks to compress and visualize multidimensional data," in
1561 2014 11th International Conference on Ubiquitous Robots and Ambient Intelligence, URAI 2014, Kuala Lumpur,
1562 Malaysia, 2014, pp. 408-412.
1563 [145] M. A. Kramer, "Nonlinear Principal Component Analysis Using Autoassociative Neural Networks," AIChE
1564 Journal, vol. 37, p. 233, 1991.
1565 [146] M. A. Kramer, "Neutral network applications in chemical engineeringAutoassociative neural networks,"
1566 Computers & Chemical Engineering, vol. 16, pp. 313-328, 1992/04/01 1992.
1567 [147] T.-H. Guo, J. Saus, C.-F. Lin, and J.-H. Ge, "Sensor validation for turbofan engines using an autoassociative
1568 neural network," in Guidance, Navigation, and Control Conference, San Diego,CA,U.S.A., 1996, p. 3926.
1569 [148] P. J. Lu and T. C. Hsu, "Application of autoassociative neural network on gas-path sensor data validation,"
1571 [149] P. J. Lu, M. C. Zhang, T. C. Hsu, and J. Zhang, "An evaluation of engine faults diagnostics using artificial
1572 neural networks," Journal of Engineering for Gas Turbines and Power, vol. 123, pp. 340-346, 2001.
1573 [150] T. A. Lemma and F. M. Hashim, "Wavelet analysis and auto-associative neural network based fault
1574 detection and diagnosis in an industrial gas turbine," in Business Engineering and Industrial Applications
1575 Colloquium (BEIAC), 2012 IEEE, 2012, pp. 103-108.
1576 [151] D. F. Specht, "Probabilistic neural networks," Neural networks, vol. 3, pp. 109-118, 1990.
1577 [152] A. K. Jain, J. Mao, and K. M. Mohiuddin, "Artificial neural networks: A tutorial," Computer, vol. 29, pp. 31-
1578 44, 1996.
1579 [153] R. Eustace and G. Merrington, "A Probabilistic Neural Network Approach to Jet Engine Fault Diagnosis,"
1580 in IEA/AIE, 1995, pp. 67-76.
1581 [154] C. Romesis and K. Mathioudakis, "Setting Up of a Probabilistic Neural Network for Sensor Fault Detection
1582 Including Operation With Component Faults," Journal of Engineering for Gas Turbines and Power, vol. 125,
1583 pp. 634-641, 2003.
1584 [155] K. Mathioudakis and C. Romessis, "Probabilistic neural networks for validation of on-board jet engine
1585 data," Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, vol. 218,
1586 pp. 59-72, 2004.
1587 [156] I. Loboda, Y. Feldshteyn, and V. Ponomaryov, "Neural networks for gas turbine fault identification:
1588 Multilayer perceptron or radial basis network?," in Proceedings of the ASME Turbo Expo, 2011, pp. 465-475.
1589 [157] I. Loboda and M. A. Olivares Robles, "Gas turbine fault diagnosis using probabilistic neural networks,"
1590 International Journal of Turbo & Jet-Engines, vol. 32, pp. 175-191, 2015.
1591 [158] R. Schaback, "A practical guide to radial basis functions," Electronic Resource, 2007.
1592 [159] S. Simani and C. Fantuzzi, "Fault diagnosis in power plant using neural networks," Information Sciences, vol.
1593 127, pp. 125-136, 2000.
1594 [160] J. A. Leonard, M. A. Kramer, and L. H. Ungar, "Using radial basis functions to approximate a function and
1595 its error bounds," IEEE Transactions on Neural networks, vol. 3, pp. 624-627, 1992.
1596 [161] N. Roy and R. Ganguli, "Filter design using radial basis function neural network and genetic algorithm for
1597 improved operational health monitoring," Applied Soft Computing, vol. 6, pp. 154-169, 2006.
1598 [162] I. Loboda, U. E. Rios, and C. E. Sanchez, "Probabilistic neural networks for gas turbine fault recognition,"
1599 Авиационно-космическая техника и технология, pp. 53–58-53–58, 2012.
1600 [163] X. Wen, X. Zhang, and Y. Zhu, "Design of Fault Detection Observer Based on Hyper Basis Function,"
1601 Tsinghua Science and Technology, vol. 20, pp. 200-204, 2016.
1602 [164] R. Ganguli, "Application of Fuzzy Logic for Fault Isolation of Jet Engines," Journal of Engineering for Gas
1603 Turbines and Power, vol. 125, pp. 617-623, 2003.
1604 [165] a. A. M. J. Fuster P. Ligeza A. , "Adductive diagnostic procedure based on an AND/OR/NOT graph for
1605 expected behaviour: application to a gas turbine," in 10th International Congress and Exhibition on Condition
1606 Monitoring and Diagnostic Engineering Management, Finland, 1997, pp. pp.51l-520.
1607 [166] L. Marinai and R. Singh, "A fuzzy logic approach to gas path diagnostics in Aero-engines," in Computational
1608 Intelligence in Fault Diagnosis, ed: Springer, 2006, pp. 37-79.
1609 [167] R. Ganguli, "Data Rectification and Detection of Trend Shifts in Jet Engine Path Measurements Using
1610 Median Filters and Fuzzy Logic," 2002.
1611 [168] S. O. T. Ogaji, L. Marinai, S. Sampath, R. Singh, and S. D. Prober, "Gas-turbine fault diagnostics: a fuzzy-
1612 logic approach," Applied Energy, vol. 82, pp. 81-89, 9// 2005.
1613 [169] A. Kyriazis and K. Mathioudakis, "Gas Turbine Fault Diagnosis Using Fuzzy-Based Decision Fusion,"
1615 [170] S. Sampath, S. Ogaji, R. Singh, and D. Probert, "Engine-fault diagnostics: an optimisation procedure,"
1616 Applied Energy, vol. 73, pp. 47-70, 2002.
1617 [171] D. E. Golberg, "Genetic algorithms in search, optimization, and machine learning. Addion Wesley,"
1618 Reading, 1989.
1619 [172] A. Chipperfield, P. Fleming, H. Pohlheim, and C. Fonseca, "Genetic algorithm toolbox for use with
1620 MATLAB," 1994.
1621 [173] A. Gulati, M. Zedda, and R. Singh, "Gas turbine engine and sensor multiple operating point analysis using
1622 optimization techniques," in 36th AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, Las Vegas,
1623 NV, U.S.A., 2000, p. 3716.
1624 [174] I. Loboda, S. Yepifanov, and Y. Feldshteyn, "A Generalized Fault Classification for Gas Turbine Diagnostics
1625 at Steady States and Transients," Journal of Engineering for Gas Turbines and Power, vol. 129, p. 977, 2007.
1626 [175] Y. G. Li, M. F. Abdul Ghafir, L. Wang, R. Singh, K. Huang, X. Feng, et al., "Improved Multiple Point
1627 Nonlinear Genetic Algorithm Based Performance Adaptation Using Least Square Method," Journal of
1629 [176] C. Kong, M. Kang, and G. Park, "Study on Condition Monitoring of 2-Spool Turbofan Engine Using Non-
1630 Linear Gas Path Analysis Method and Genetic Algorithms," International Journal of Materials, Mechanics and
1631 Manufacturing, pp. 214-220, 2013.
1632 [177] J. S. Breese, E. J. Horvitz, M. A. Peot, R. Gay, and G. H. Quentin, "Automated decision-analytic diagnosis
1633 of thermal performance in gas turbines," in ASME 1992 International Gas Turbine and Aeroengine Congress
1634 and Exposition, 1992, pp. V005T15A015-V005T15A015.
1635 [178] C. A. Palmer, "Combining Bayesian belief networks with gas path analysis for test cell diagnostics and
1636 overhaul," in ASME 1998 International Gas Turbine and Aeroengine Congress and Exhibition, 1998, pp.
1637 V005T15A015-V005T15A015.
1638 [179] A. Kadamb, "Bayesian belief network for aero gas-turbine module and system-fault isolation," MS.c.,
1639 Cranfield University, UK, 2003.
1640 [180] C. Romessis, A. Stamatis, and K. Mathioudakis, "Setting up a belief network for turbofan diagnosis with
1641 the aid of an engine performance model," ISABE paper, vol. 1032, pp. 19-26, 2001.
1642 [181] K. Mathioudakis, P. Kamboukos, and A. Stamatis, "Turbofan Performance Deterioration Tracking Using
1643 Nonlinear Models and Optimization Techniques," Journal of Turbomachinery, vol. 124, p. 580, 2002.
1644 [182] C. Romessis and K. Mathioudakis, "Bayesian network approach for gas path fault diagnosis," Journal of
1645 engineering for gas turbines and power, vol. 128, pp. 64-72, 2006.
1646 [183] Y. K. Lee, D. N. Mavris, V. V. Volovoi, M. Yuan, and T. Fisher, "A fault diagnosis method for industrial gas
1647 turbines using Bayesian data analysis," Journal of engineering for gas turbines and power, vol. 132, p. 041602,
1648 2010.
1649 [184] D. T. Pham and G. A. Ruz, "Unsupervised training of Bayesian networks for data clustering," Proceedings of
1650 the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 465, pp. 2927-2948, 2009.
1651 [185] J. I. Alonso-Barba, L. delaOssa, J. A. Gámez, and J. M. Puerta, "Scaling up the Greedy Equivalence Search
1652 algorithm by constraining the search space of equivalence classes," International Journal of Approximate
1653 Reasoning, vol. 54, pp. 429-451, 2013/06/01/ 2013.
1654 [186] H. A. Nozari, M. A. Shoorehdeli, S. Simani, and H. D. Banadaki, "Model-based robust fault detection and
1655 isolation of an industrial gas turbine prototype using soft computing techniques," Neurocomputing, vol. 91,
1656 pp. 29-47, 2012.
1657 [187] R. Bettocchi, M. Pinelli, P. R. Spina, and M. Venturini, "Artificial Intelligence for the Diagnostics of Gas
1658 Turbines—Part I: Neural Network Approach," Journal of Engineering for Gas Turbines and Power, vol. 129,
1659 pp. 711-719, 2006.
1660 [188] Y. G. Li, M. F. A. Ghafir, L. Wang, R. Singh, K. Huang, and X. Feng, "Nonlinear Multiple Points Gas Turbine
1661 Off-Design Performance Adaptation Using a Genetic Algorithm," Journal of Engineering for Gas Turbines and
1662 Power, vol. 133, p. 071701, 2011.
1663 [189] C. A. Palmer, "Combining Bayesian belief networks with gas path analysis for test cell diagnostics and
1664 overhaul," in American Society of Mechanical Engineers (Paper), 1998.
1665 [190] J. Pearl, Probabilistic reasoning in intelligent systems: networks of plausible inference: Elsevier, 2014.
1666 [191] E. Tsoutsanis, "Performance adaptation of gas turbines for power generation applications," PhD, Cranfield
1667 University, UK, 2010.
1668 [192] G. Denney, "F16 jet engine trending and diagnostics with neural networks," in Optical Engineering and
1669 Photonics in Aerospace Sensing, Orlando, FL, United States, 1993, pp. 419-422.
1670 [193] S. Sampath and R. Singh, "An Integrated Fault Diagnostics Model Using Genetic Algorithm and Neural
1671 Networks," Journal of Engineering for Gas Turbines and Power, vol. 128, p. 49, 2006.
1672 [194] A. Green, D. Allen, A. Green, and D. Allen, "Artificial intelligence for real time diagnostics of gas turbine
1673 engines," in 33rd Joint Propulsion Conference and Exhibit, 1997, p. 2899.
1674 [195] T. Kobayashi and D. L. Simon, "Hybrid neural-network genetic-algorithm technique for aircraft engine
1675 performance diagnostics," Journal of Propulsion and Power, vol. 21, pp. 751-758, 2005.
1676 [196] R. F. Orsagh, M. J. Roemer, C. J. Savage, and M. Lebold, "Development of performance and effectiveness
1677 metrics for gas turbine diagnostic technologies," in Aerospace Conference Proceedings, 2002. IEEE, 2002, pp.
1678 6-6.
1679 [197] V. George, L. Frank, R. Michael, H. Andrew, and W. Biqing, "Intelligent fault diagnosis and prognosis for
1680 engineering systems," ed: John Wiley & Sons, Inc, 2006.
1681 [198] D. L. Simon, S. Borguet, O. Léonard, and X. F. Zhang, "Aircraft engine gas path diagnostic methods: public
1682 benchmarking results," Journal of Engineering for Gas Turbines and Power, vol. 136, p. 041201, 2014.
1683 [199] P. Kamboukos and K. Mathioudakis, "Comparison of Linear and Nonlinear Gas Turbine Performance
1684 Diagnostics," Journal of Engineering for Gas Turbines and Power, vol. 127, p. 49, 2005.
1685 [200] C. Romesis and Y.-G. Li, "CONDITION-BASED MAINTENANCE FOR GAS TURBINES PLANTS."
1686 [201] B. Curnock, "Obidicote project—word package 4: Steady-state test cases," Rolls-Royce Plc, Technical Report
1687 DNS62433, 2000.
1688 [202] D. L. Simon, J. Bird, C. Davison, A. Volponi, and R. E. Iverson, "Benchmarking gas path diagnostic methods:
1689 a public approach," in ASME Turbo Expo 2008: Power for Land, Sea, and Air, 2008, pp. 325-336.
1690
© 2018 by the authors. Submitted for possible open access publication under the terms
and conditions of the Creative Commons Attribution (CC BY) license
(http://creativecommons.org/licenses/by/4.0/).
1691

A Review On Gas Turbine Gas-Path Diagnostics: State-Of-The-Art Methods, Challenges and Opportunities

Uploaded by

Copyright:

Available Formats

A Review On Gas Turbine Gas-Path Diagnostics: State-Of-The-Art Methods, Challenges and Opportunities

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Review On Gas Turbine Gas-Path Diagnostics: State-Of-The-Art Methods, Challenges and Opportunities

Uploaded by

Copyright:

Available Formats

1 Review

2 A Review on Gas Turbine Gas-path Diagnostics:

29 Keywords: Gas turbine performance; gas-path diagnostics; condition-based maintenance; fault

82 2. Gas Turbine Performance Degradation

104 2.1. Fouling

Compressor fouling consequences Ref.

156 2.2. Erosion

169 2.3. Corrosion

181 2.4. Foreign Object Damage (FOD)/Domestic Object damage (DOD)

194 2.5. Increase in Blade tip clearance

Physical Contaminant/Cause Exposed Effect Performance Results References

215 3. Fault Diagnostics

276 3.1. Challenges of a Successful GT Fault Diagnostics

307 severity of components’ performance deterioration. A careful measurement selection is

367 3.2. Desirable Attributes of a Fault Diagnostic System

411 4. State-of-the-art: GT Gas-path Diagnostic Methods

421 Table 3. Categorization of fault diagnostic methods presented in literature.

Author Ref. Year Classification Categories

422 4.1. Model-based Diagnostic Methods

429 4.1.1. Gas-path Analysis

444 4.1.1.1. Linear GPA (LGPA)

494 4.1.1.2. Non-Linear GPA (NLGPA)

561 4.1.2. The Kalman Filter

576 The problem is defined mathematically as follows:

579 o Measurement equation: Zk  Hk Xk  vk (11)

592 Ewk   0 (14)

593 Evk   0 (15)

603 Pk1/ k  Fk Pk / k FkT  Gkuk (17)

607 1. State estimate extrapolation:

626 - k : System error covariance at time k

682 4.2. Advantages and Limitations of MB Methods

694 Table 4. Advantages and limitations of GPA and KF related to GT diagnostics.

Method Advantages Limitations

699 4.3. AI based Methods

717 4.3.1. Artificial Neural Networks

733 4.3.1.1. Multilayer Perceptron

767 4.3.1.2. Autoassociative Neural Networks

781 Figure 9. General architecture of AANN.

808 4.3.1.3. Probabilistic Neural Network

847 4.3.1.4. Radial Basis Function Networks

914 4.3.2. Fuzzy Logic methods

945 4.3.3. Genetic Algorithm

1023 4.3.4. Bayesian Belief Network

1066 4.3.5. Expert Systems

1101 4.4. Advantages and Limitation of AI Methods

1111 Table 5. Advantages and limitations of AI-based methods regarding GT diagnostics.

Method Advantages Limitations

1112 Table 5. (Continued…).

Method Advantages Limitations

1114 4.5. Hybrid Methods

1150 5. GT Diagnostic Methods Validation Techniques

1196 6. Conclusions and Future Research Directions

You might also like