1 s2.0 S1876610218304028 Main
1 s2.0 S1876610218304028 Main
1 s2.0 S1876610218304028 Main
com
Available online at www.sciencedirect.com
ScienceDirect
ScienceDirect
Availableonline
Available onlineatatwww.sciencedirect.com
www.sciencedirect.com
Energy Procedia 00 (2018) 000–000
ScienceDirect
ScienceDirect
Energy Procedia 00 (2018) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Energy
EnergyProcedia
Procedia148 (2018) 000–000
00 (2017) 860–867
www.elsevier.com/locate/procedia
73rd Conference of the Italian Thermal Machines Engineering Association (ATI 2018),
73rd Conference of the Italian
12–14Thermal Machines
September 2018, Engineering
Pisa, Italy Association (ATI 2018),
12–14 September 2018, Pisa, Italy
A diagnostics tool for aero-engines health monitoring using machine
A diagnostics
Thetool for aero-engines
15th International Symposiumhealth on District monitoring
Heating and Coolingusing machine
learning technique
learning technique
Assessing the feasibility
Maria Grazia De Giorgiaa*, Stefano of using the heat
Campilongo a demand-outdoor
, Antonio Ficarellaaa
a
temperature function
Maria Grazia a
De Giorgifor a*, long-term
Stefano Campilongo district, Antonio
heat demandFicarella forecast
University of Salento, Department of Engineering for Innovation, via per Moonteroni, 73100 LECCE (IT)
a
University of Salento, Department of Engineering for Innovation, via per Moonteroni, 73100 LECCE (IT)
I. Andrića,b,c*, A. Pinaa, P. Ferrãoa, J. Fournierb., B. Lacarrièrec, O. Le Correc
Abstract
Abstract
a
IN+ Center for Innovation, Technology and Policy Research - Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisbon, Portugal
b
In this work an integrated heath Veolia monitoring
Recherche & platform
Innovation,is291 Avenue Dreyfous
proposed Daniel, 78520
and developed Limay, Franceanalysis and degradation
for performance
c
diagnostics of Département
gas turbine Systèmes
engines. Énergétiques
The aim is et
to Environnement
link engine - IMT
measurable Atlantique,
data
In this work an integrated heath monitoring platform is proposed and developed for performance to 4
itsrue Alfred
health Kastler, 44300analysis
status. Nantes, France
and degradation
A numericaloftool
diagnostics gashas beenengines.
turbine implemented
The aimin order to calculate
is to link engine performance
engine measurable data to its in design
health condition and to create a database of
status.
expected
A numerical vales.
toolThen
has different degradation
been implemented in levels
order tohave been introduced
calculate in the two in
engine performance main components,
design conditioncompressor anda turbine
and to create databaseofofa
single spool
expected turbojet
vales. Thenand the diagnostics
different degradation instruments
levels have have
beenbeen trained to
introduced indetect
the two themain
component fault. compressor and turbine of a
components,
Abstract
In order
single to turbojet
spool evaluateand thetheperformance
diagnostics prediction
instrumentstwo havedifferent machine
been trained learning
to detect based techniques,
the component fault. namely, artificial neural
network
In order (ANN)
to evaluateand support vector machine
the performance (SVM)two
prediction havedifferent
been compared.
machine learning based techniques, namely, artificial neural
District heating networks are commonly addressed in the literature as one of the most effective solutions for decreasing the
Synthetic(ANN)
network data generation
and supporthasvector
been machine
carried out to show
(SVM) havehow
beenthe degradation effects can affect the engine performance. The two
compared.
greenhouse gas emissions from the building sector. These systems require high investments which are returned through the heat
main degradation
Synthetic causes considered
data generation are the out
has been carried compressor
to show fouling
how theand turbine erosion.
degradation effects can affect the engine performance. The two
sales. Due to the changed climate conditions and building renovation policies, heat demand in the future could decrease,
The machine
main degradation learning
causestechniques
considered were applied
are the with two
compressor aims: and
fouling aero-engine performance prediction and health diagnostics. The
turbine erosion.
prolonging the investment return period.
studymachine
The was carried out based
learning on three
techniques weresamples
appliedflights, whose
with two dataaero-engine
aims: were used for the training
performance and testing
prediction andprocess
health of the prediction
diagnostics. The
The main scope of this paper is to assess the feasibility of using the heat demand – outdoor temperature function for heat demand
and diagnostics
study was carried tools.
out based on three samples flights, whose data were used for the training and testing process of the prediction
forecast. The district of Alvalade, located in Lisbon (Portugal), was used as a case study. The district is consisted of 665
The diagnostics
and knowledge tools. and the continuous monitoring of the engine health status can be crucial for maintenance and fleet management
buildings that vary in both construction period and typology. Three weather scenarios (low, medium, high) and three district
operations.
The knowledge and the continuous monitoring of the engine health status can be crucial for maintenance and fleet management
renovation scenarios were developed (shallow, intermediate, deep). To estimate the error, obtained heat demand values were
operations.
compared with results from a dynamic heat demand model, previously developed and validated by the authors.
© 2018 The Authors. Published by Elsevier Ltd.
© The
2018 results
The showed that
Authors. when only
Published weatherLtd.
by Elsevier
Elsevier change is considered, the margin of error could be acceptable for some applications
This
© is an
2018 Theopen accessPublished
Authors. article underby the CC BY-NC-ND
Ltd. license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
(theiserror
This in annual
an open accessdemand was lower
article under the CCthan 20% for license
BY-NC-ND all weather scenarios considered). However, after introducing renovation
(https://creativecommons.org/licenses/by-nc-nd/4.0/)
Selection and peer-review under responsibility
This is an open access article under the CC BY-NC-ND of thelicense
scientific committee of the 73rd Conference of the Italian Thermal
(https://creativecommons.org/licenses/by-nc-nd/4.0/)
Selection
scenarios, and
thepeer-review
error valueunder responsibility
increased up to 59.5%of the(depending
scientific committee of the and
on the weather 73rdrenovation
Conferencescenarios
of the Italian Thermal considered).
combination Machines
Machines
Selection
Engineering Engineering
andAssociation Association
peer-review (ATI under (ATI 2018). of the scientific committee of the 73rd Conference of the Italian Thermal
responsibility
2018).
The value of slope coefficient increased on average within the range of 3.8% up to 8% per decade, that corresponds to the
Machines Engineering Association (ATI 2018).
decrease in the number of heating hours of 22-139h during the heating season (depending on the combination of weather and
Keywords: Health Monitoring, Neural Network, diagnostics analysis.
renovation scenarios considered). On the other hand, function intercept increased for 7.8-12.7% per decade (depending on the
Keywords: Health Monitoring, Neural Network, diagnostics analysis.
coupled scenarios). The values suggested could be used to modify the function parameters for the scenarios considered, and
improve the accuracy of heat demand estimations.
1. Introduction
Several factors could degrade propulsion and flight performance of aero-engines. The most common degradation
mechanisms are associated to variations in blade surfaces due to erosion, fouling, debris deposit and the effect on the
blade aerodynamics.
Fouling is the most common reason of compressor deterioration and is recognized as the source of about 70–85% of
performance degradation in gas turbine engines. It occurs in the first 5-6 stages decreasing from the front to the back
end of the compressor [1]. Fouling deterioration of gas turbine depends on two factors [2]: the exposure of the
engine to compressor fouling based on its design parameters and the sensitivity of the gas turbine to a certain degree
of imposed fouling. Source for fouling phenomena are dust, dirt, sand, ash and carbon particles and soft particles
such as oil, unburned hydrocarbons, soot, etc. [3]. Fouling belongs to the group of recoverable deterioration and can
be reduced or even eliminated by cleaning to restore the gas turbine engine flow path surfaces to near initial
conditions.
Another cause of degradation is erosion phenomenon on turbine stages due to hard abrasive particles suspended in
the air or gas stream. Erosion causes performance losses due to the decrease in surface roughness, an increase in
blade tip clearance, blunting of the blade leading edge, thinning of the trailing edge and shortening of the blade
chord [4-6]. In the turbine section, damage due to erosive particles that enter with the fuel is particularly severe.
The identification and prediction of the degradation phenomena mentioned above by detection of engine status is the
challenge of recent research studies [7-8]. Diagnostics method consists in the implementation of an efficient health
monitoring which investigates engine performance degradation observing component parameters compared to the
same parameters of an “healthy” engine.
The key approaches useful for these kinds of problems are based on the use of gas path analysis (GPA), which aims
to delivering early warning information about ongoing degradations through measurable changes in engine
performance parameters. An interesting analysis of the measurements validation before they are used in GPA tools
has been carried out in [9]. Several parameters should be taken into account in the implementation of prediction
methods, these can be collected information from in-flight measurements or from engine model to estimate health
indicators, which cannot be directly measured. In [10] the habitual problem related to the lack of data from real data
acquisition system is overcome using multiple operating point minimization technique. Sallee et al. [11] have
described how the two main engine-parameters, namely flow capacity and efficiency, are affected by the engine
configuration and deterioration.
Regarding the machine learning techniques, neural networks (ANN) are most widespread because provides a viable
tool for dealing with nonlinear problems and modelling complex and nonlinear dynamic systems with great
flexibility and capability [12-17]. Barad et al. [18] proved that ANN-based performance healthy-monitoring tool is
sufficiently robust and delivers an early warning compared with the mechanical parameters. Neural network can be
trained to perform a particular function by adjusting the values of the connections between elements leading to a
specific target output. Several kinds of ANN are used in the application jet engine fault diagnosis: dynamic neural
networks [12], feed-forward neural network [19], Radial Basis Network (RBN) [20].
However, even if ANN can be only applied to the specified aero-engine type, often ANNs do not accurately describe
the complex system to monitor aero-engine condition. While traditional artificial neural networks are constructed
around the empirical risk minimization principle, which limits their generalization capability, support vector
machines (SVM) indicates its superior ability in solving nonlinear regression problems. The SVM model has a
similar functional form to ANN but has a better generalization performance, and a good ability to model complex
problems in presence of data sets with several variables and with a limited set of available data for training.
Furthermore despite the good prediction performances of ANNs, they present disadvantages such as the tendency to
over-fit, and although the training data may be very well fitted, the resulting function hasn’t got a general value.
Moreover, the ANN needs large computational resources for training. Recently the SVM algorithm was successfully
used as a novel powerful learning tool machine used for forecasting in several fields [21-22].
Therefore, in this paper we developed a data driven method which employs both artificial neural networks and
support vector machines to build aero-engine monitoring model using a set of training data and testing data. The
numerical results used for the models test show that the approach can provide an effective monitoring in aero-engine
performance and it could identify the component (compressor or turbine) that presents degradation.
862 Maria Grazia De Giorgi et al. / Energy Procedia 148 (2018) 860–867
Maria Grazia De Giorgi/ Energy Procedia 00 (2018) 000–000 3
Nomenclature
INPUT PARAMETERS
VECTOR SET
Xtrain (Alt Mach %N)FLIGHT 1-2
Ytrain (mf EGT)FLIGHT 1-2
Xtest (Alt Mach %N)FLIGHT 3
Ytest (mf EGT)FLIGHT 3
ANN or SVM model takes into account engine operating condition and gives as output engine measurable data. This
first step can be considered an engine model able to monitor the engine performance. The data coming from the
machine learning model are compared with the data coming from real engine acquisition system.
In a second phase, if the two data disagree, the ANN or SVM diagnostics model works to detect the degraded
component, taking into account flight data and data from engine acquisition system and giving as output diagnostics
information about components health state and helping about maintenance decisions.
3. Machine learning tools for gas turbine performance and health status estimation
In the first part of this section a brief description of the artificial neural network and support vector machine has
been presented; then the test cases and relative scenarios are shown.
3.1 Artificial Neural Network
The neural network consists of a number of interconnected processing nodes called neurons, which are organized in
a sequence of layers, including an input layer, a single intermediate hidden layer, and an output layer.
The input signals – of varying intensity and strength – feed through the neuron and combine to form a net input into
another neuron. The output layer, equal to the number of dependent variables, is calculated by weight and bias
associated with connections among neurons. The intermediate layer is connected to the input and output layers. Each
neuron in the hidden and output layers receives the signals from all the neurons in a layer above it and then performs
a weighted summation and transfer function of the inputs.
In particular a two-layer feed-forward network with sigmoid hidden neurons and linear output neurons was
implemented. The network was trained with Bayesian regularization training algorithm function. The optimal
number of hidden neurons and of the number of epochs was obtained carrying out a sensibility analysis based on the
percent error between the target value and output values obtained by using neural networks with different numbers
of hidden neurons and epochs. The best performance of the neural network was achieved with 5 hidden neurons and
2000 epochs for every simulates test case. For both the models the data of three flights need to be divided into
training, validation, and test sets. The validation set was used to ensure that there was no overfitting in the final
results. In detail for training and validation dataset have been used data from Flight#1 and Flight#2; the network
testing has been performed on the data from Flight#3. In this manner we are able to quantify the tool efficiency
using data “never seen before” by the network.
Table 2. Engine degradation level.
Synthetic health
CONFIGURATION ηcomp ηturb
state
HEALTHY 0 0.90 0.89
0.85
COMPRESSOR
1 0.80 0.89
DEGRADATION
0.77
0.84
TURBINE
2 0.90 0.79
DEGRADATION
0.76
COMPRESSOR & 0.85 0.84
TURBINE 3 0.85 0.79
DEGRADATION 0.80 0.84
864 Maria Grazia De Giorgi et al. / Energy Procedia 148 (2018) 860–867
Maria Grazia De Giorgi/ Energy Procedia 00 (2018) 000–000 5
Table 3. Different data set for ANN and SVM engine health status diagnostics models.
INPUT PARAMETERS
VECTOR
SET
Xtrain (Mach mf EGT)FLIGHT 1-2
Ytrain (0 1 2 3)FLIGHT 1-2
Xtest (Mach mf EGT)FLIGHT 3
Ytest (0 1 2 3)FLIGHT 3
can be expected to perform on data not used to train it. As said before Flight #1 and Flight #2 have been used for
training. The accuracy level has been measured on the data from Flight #3, comparing fuel flow and EGT values
coming from ANN and SVM output with the expected values. Fig. 1 shows the trend of the two physical quantities
and reports the synthetic value of the associated RMSE. It’s evident that both the models are able to predict well the
engine performance. It’s notable that the SVM model has a bit larger RMSE value than ANN model, particularly for
the output values near the minimum, 0.25 for normalized fuel flow and 0.65 for normalized EGT. The trained and
tested models can be used to create a database of expected values otherwise it can be run in real time to give the
reference value for a specific set of altitude, Mach value and engine rotational speed.
(a)
(b)
Fig. 1. . Performance prediction results for ANN model (a) and SVM (b).
866 Maria Grazia De Giorgi et al. / Energy Procedia 148 (2018) 860–867
Maria Grazia De Giorgi/ Energy Procedia 00 (2018) 000–000 7
(a)
(b)
Fig. 2. Engine synthetic health status prediction for ANN (a) and SVM (b).
We tested the performance of the dataset listed in Table 3, where we chose engine performance parameters as input
and the previously described synthetic health state as output. The used dataset is a compromise between
performance parameters, such as fuel flow rate and exhaust gas temperature, and flight parameters, such as Mach
number. FigureFig. 2 shows the trend of the models output respect to the expected values and the efficiency of this
forecast is depicted with histograms; in this case the tool efficiency is calculated counting the number of right values
respect to the total values. In the efficiency histograms the four zones with a different health state code have been
highlighted and separated in order to evaluate the model performance in every condition. In Fig. 2a the results for
the ANN model and in Fig. 2b for SVM model are depicted.
Dividing the results in zones with different degraded scenarios it’s detectable the component with major troubles but
it’s impossible to quantify its deterioration level giving us only an alert about the engine health condition.
The SVM models have been trained with different optimization function: linear, quadratic and cubic. Table 4 reports
the performance for each available function. The performance are measured in terms of RMSE value for the model
that predicts engine performance and in terms of accuracy for the model that classifies engine health status.
The quadratic learning function which shows the best results has been chosen for the SVM models. The high
accuracy level (99.4%) is evident also in the graphs of Figure 4b. The SVM model with quadratic learning function
has the best performance as classifier of the engine health status. The ANN model has very good performance for
the healthy engine condition but there are some misleading prediction for the output conditions 1, 2 and 3.
Conclusions
This work aims to create a tool for monitoring aero-engine health status. Different machine learning techniques have
been compared. Both ANN and SVM based models have shown very good results, in terms of engine performance
parameters calculated from only few operational data; the very small computational time for this tools allows their
use also in real-time condition. The engine model based on ANN is slightly better than the SVM based model
probably because of the linearity between data used as input and output. The diagnostics models needed more time
Maria Grazia De Giorgi et al. / Energy Procedia 148 (2018) 860–867 867
8 Maria Grazia De Giorgi / Energy Procedia 00 (2018) 000–000
and more attempts to find the right compromise between performance and operating parameters to use as input. For
this purpose, SVM model has shown excellent performance, outperforming ANN as classifier tool.
As further development we want to create a tool able to calculate the analytic value of components efficiency.
The knowledge of the real engine health status can be crucial for maintenance and fleet management decisions. The
long-term objective of the research will be a data-driven maintenance, which can be carried out with an innovative
methodological analysis of the big data cluster from aero-engine system.
References
[1] Tarabrin, A. P., Schurovsky, V. A., Bodrov, A. I., and Stalder, J. P., 1998, “Influence of Axial Compressor Fouling on Gas Turbine Unit
Performance Based on Different Schemes and With Different Initial Parameters,” ASME Report No. 98-GT-416.
[2] Meher-Homji C.B., Chaker M., Bromley A.F. (2009). The fouling of axial flow compressors: causes, effects, susceptibility, and sensitivity.
Proceedings of ASME Turbo Expo 2009: Power for Land, Sea and Air vol. 4: 571-590.
[3] Kurz R. and Brun, K. (2001). Degradation in gas turbine systems. Journal of Engineering for Gas Turbines and Power vol.123: 70-77.
[4] Hamed A. A., Tabakoff W., Rivir R. B., Das K. and Arora P. (2005). Turbine blade surface deterioration by erosion. Journal of
turbomachinery vol. 127: 445–452.
[5] De Giorgi M.G., Campilongo S., Ficarella, A., Coltelli, M., Pfister, V. and Sepe, F. (2013). “Experimental and numerical study of particle
ingestion in aircraft engine”. In: vari. Proceedings of the ASME Turbo Expo 2013 GT2013-95662 June 3-7, 2013, San Antonio, Texas, USA.
[6] De Giorgi M.G., Campilongo S., Ficarella, A. “Predictions of Operational Degradation of the Fan Stage of an Aircraft Engine Due to
Particulate Ingestion”. J. of Eng. for gas turbines and power, vol. 137, p. 052603-1-052603-15, ISSN: 0742-4795, doi: 10.1115/1.4028769.
[7] Litt, J. S., Simon, D. L. and S. Garg. “A survey of intelligent control and health management technologies for aircraft propulsion systems,”
Tech. Rep. NASA-TM-2005-213622, 2005.
[8] Melcher, K. J., Maul, W. A. and Garg, S. “Propulsion health management system development for affordable and reliable operation of space
exploration systems,” AIAA 2007-6237, 2007.
[9] Bettocchi, R., Pinelli, M., Venturini, M., Spina, P.R., Sebastianelli, S., 2001, “A system for health state determination of natural gas
compression gas turbines”, ASME Paper 2001-GT-0223.
[10] Pinelli, M., Spina, P.R., Venturini, M., 2003, Optimized operating point selection for gas turbine health state analysis by using a multi-point
technique, ASME Paper GT2003-38191.
[11] Sallee G.P. (1978) Performance Deterioration Based on Existing (Historical) Data–JT9D Jet Engine Diagnostics Progr. NASA CR-135448.
[12] Tayarani-Bathaie, S.S., Sadough Vanini, Z.N. and Khorasani, K.. “Fault detection of gas turbine engines using dynamic neural networks”,
Conference Paper in Canadian Conference on Electrical and Computer Engineering, April 2012.
[13] Dragomir O.E., Gouriveau R., Dragomir F., Minca E. and Zerhouni N. (2009). “Review of prognostic problem in condition based
maintenance”. European control conference (ECC) IEEE: 1587–92.
[14] Bettocchi, R., Pinelli, M., Venturini, M., Spina, P.R., 2004, Gas turbine operating state determination by using neural network models, Atti
59° Congresso Nazionale ATI (Genova ), SGE, Padova.
[15] Bettocchi, R., Pinelli, M., Venturini, M., Spina, P.R., Burgio, M., 2004, Set up of a robust neural network for gas turbine simulation, ASME
Paper GT2004-53421.
[16] Bettocchi R., Pinelli M., Spina P. R., Venturini M., Zanetta, G.A., 2006, Assessment of the Robustness of Gas Turbine Diagnostics Tools
Based on Neural Networks, ASME Paper GT2006-90118.
[17] Bettocchi, R., Pinelli, M., Spina, P.R., Venturini, M., 2007, Artificial Intelligence for the Diagnostics of Gas Turbines — Part I: Neural
Network Approach, ASME J. Eng. Gas Turbines Power, 129, pp. 711-719.
[18] Barad S.G., Ramaiah P., Giridhar R. and Krishnaiah G. “Neural network approach for a combined performance and mechanical health
monitoring of a gas turbine engine”. Mechanical Systems and Signal Processing, 2012, 27: 729–742.
[19] Loboda I., Feldshteyn Y. and Ponomaryov V. “Neural Networks for Gas Turbine Fault Identification: Multilayer Perceptron or Radial Basis
Network” Proceedings of ASME Turbo Expo 2011.
[20] Sampath S. an Singh R. “An Integrated Fault DiagnosticS Model Using Genetic Algorithm and Neural Networks”. Journal of Engineering
for Gas Turbines and Power, 2012 vol.128: 49–56.
[21] Fu, X. Y., and Zhong, S. S. (2013, July). “Aero-engine turbine exhaust gas temperature prediction using process support vector machines”,
in International Symposium on Neural Networks (pp. 300-310). Springer, Berlin, Heidelberg.
[22] De Giorgi M.G., Campilongo S., Ficarella and Congedo P. “Comparison Between Wind Power Prediction Models Based on Wavelet
Decomposition with Least-Squares Support Vector Machine (LS-SVM) and Artificial Neural Network (ANN)”. ENERGIES, vol. 7, p. 5251-
5272, ISSN: 1996-1073, doi: 10.3390/en7085251.
[23] De Giorgi, M. G., De Carlo, L. and Ficarella, A. “Degradation prognostic in gas turbine engines using artificial neural networks”. 7th EASN
International Conference, 26-29 of September, in Warsaw, Poland.
[24] Mattingly, J. D., Heiser, W. H. and Pratt, D. T. Aircraft Engine Design, Second Edition by the American Institute of Aeronautics and
Astronautics, Inc, 2002.
[25] Suykens J.A.K., Van Gestel T., De Brabanter J., De Moor B., Vandewalle J., Least Squares Support Vector Machines, W. Sci. Singapore.