General Electric CF34-8C5B1 Turbofan Parameters

Estimation using Artificial Neural Networks and Flight

Tests

Rojo Princy Andrianantara1, Georges Ghazi2, Ruxandra Mihaela Botez3

Univeristy of Québec, École de Technologie Supérieure,
Laboratory of Applied Research in Active Controls, Avionics and AeroServoElasticity LARCASE,
Montreal Quebec, H3C 1K3

I. INTRODUCTION
Mathematical modeling of engine systems is especially useful to monitor the engine health, to detect

the engine faults and run full diagnostics. In the best scenario, a full recovery is expected based on an

accurate engine model. Therefore, several techniques have been developed since the occurrence of first gas

turbines in order to evaluate their performances and degradation over the years.

The mathematical modeling problem is solved for an engine based on system identification techniques.

In this context, the modeler tries to identify a mathematical model reflecting the characteristics of the system

of interest based on input and output data. In the field of turbomachinery, one tends to model the

thermodynamic cycle of gas turbines, also known as the Brayton cycle on Figure 1. For this cycle, the well-

known Component Level Modeling (CLM) technique is usually applied, which consists of modeling each

engine component and parameters (local temperature, local pressure, local airflow) by applying

thermodynamic relationships [1]. However, the CLM technique requires component maps, which are often

classified as confidential by manufacturers. Therefore, user opts for engine modeling software, which can

model engine systems based on public engine data [1]. Thus, the system identification problem becomes a

1
Master Student, LARCASE, ETS, 1100 Notre Dame West, Montreal, QC, H3C-1K3, Canada.
2
Postdoctoral Researcher, LARCASE, ETS, 1100 Notre Dame West, Montreal, QC, H3C-1K3, Canada.
3
Full Professor, AIAA Associate Fellow, Canada Research Chair Holder Level 1 in Aircraft Modeling and
Simulation, Head of the LARCASE, ETS, 1100 Notre Dame West, Montreal, QC, H3C-1K3, Canada.
parameter estimation problem with its aim to fit the generic engine model to the actual engine model. The

estimated parameters are generally the engine component efficiencies.

One solution for system identification consists in using Artificial Intelligence (AI). Thanks to

development of AI methods and to their extensive applications in aerospace, AI could be used to solve

complex problems in the turbomachinery field. Indeed, since early 2000s researchers have applied Artificial

Neural Networks (ANN) to model gas turbines cycle [2 – 5]. Recently, more sophisticated Neural Networks

have been designed for faults detection, engine diagnostics and decision-making process [6 – 8]. In addition,

several works on engine performance modeling were carried out at the LARCASE based on empirical

methods [9], CLM techniques [10 – 12] and Neural Networks [13 – 14].

Figure 1 Brayton cycle for a turbofan engines

The main objective of this paper consists in identifying and validating an engine performance model of

the General Electric CF34-8C5B1 high bypass Turbofan represented in Figure 2, based on Artificial Neural

Networks and flight tests. The flight tests were realized on a Virtual Research Simulator (VRESIM) of the

CRJ-700 (Figure 3). The VRESIM is qualified Level-D, which is the highest qualification given to flight

simulators for flight dynamics and engines, according to the Federal Aviation Administration.
 (a) (b)

Figure 2 (a) Diagram of the General Electric CF34-8C5B1 Turbofan (b) Numbering of the engine
stations

A total of 451 flight tests were realized, covering the aircraft flight envelope as well as the engine

operation envelope for four major flight phases including Takeoff, Climb, Cruise and Descent. Then, the

data were extracted from the VRESIM and implemented into Matlab for system identification. The

collected data include pressure altitude, Mach number, fan speed, core speed, component temperatures,

component pressures, local airflows, temperature deviations, fuel flow and thrust.

Figure 3 Flight simulator of the CRJ700 (VRESIM)

 II. METHODOLOGY
The engine neural network modeling consists of two steps approach, as shown in Figure 4. First, the

engine parameters were modeled using a Multi-Input Multi-Output Feedforward Neural Network (MIMO

FFNN). The network inputs were the altitude, Mach number, Thrust Lever Angle (TLA), International

Standard Atmosphere data (ISA) such as static temperature, static pressure and air density, and the

temperature deviation from ISA. The modeled engine parameters were the Engine Pressure Ratio (EPR),

Fan Pressure Ratio (FPR), Overall Pressure Ratio (OPR), Inlet Turbine Temperature (ITT), fan speed (N 1)

and core speed (N2).

Figure 4 Engine Neural Network Model architecture

The resulting Neural Network NET-1 aims to mimic the engine control system known as Full Authority

Digital Engine Controller (FADEC), and calculates the engine thrust ratings shown in Figure 5, such as

Maximum Takeoff Thrust (MTO), Maximum Climb Thrust (MCLB) and Idle Thrust (IDLE). Each thrust

rating corresponds to a specific position of the Thrust Lever Angle, respectively for Takeoff, Climb and

Descent operations. The thrust during cruise is calculated in the first step by assuming a constant variation

between the IDLE and MCLB positions.

The second step consists of designing the second neural network NET-2, which calculates the engine

performances, including the fuel flow and net thrust by using a MIMO FFNN. The NET-2 inputs consists

of the engine parameters NET-1 outputs and the Mach number.

 Figure 5 Thrust Lever Angle and Engine Thrust ratings

The Neural Networks were modeled in Matlab using its integrated Neural Network Toolbox. The

Bayesian regularization algorithm was applied to train both neural networks NET-1 and NET-2, and has a

very good generalisation ability. The number of hidden layers was limited at one in order to estimate the

number of neurons while the hyperbolic tangent sigmoid function was applied as activation function. The

Mean Square Error (MSE) was considered as cost function to evaluate the neural network performance for

both networks. In order to avoid overfitting, the number of neurons for each neural network has been set at

the minimum MSE, which corresponds to 5% of its maximum value, as seen on Figure 6.

Figure 6 MSE vs Number of neurons

 The 451 flight profiles were gathered and resampled into flight points, where one point represents one

flight condition (i.e., altitude, Mach number, and TLA). Using a time-spacing of 20 to 30 seconds, a total

number of 3,400 flight points have been gathered for the engine database (see Figure 7), of which 500

points (15%) were used for engine model identification, and 2,900 points (85%) were used for engine model

validation.

Figure 7 Flight tests and Engine database generation

III. MODEL VALIDATION AND SIMULATION

The validation process consists of a comparison study between the engine model outputs and the

VRESIM outputs, which are the reference aircraft outputs. Results on Table 1 show that the engine model

is able to accurately predict the engine parameters (𝐸𝑃𝑅, 𝐹𝑃𝑅, 𝑂𝑃𝑅, 𝐼𝑇𝑇, 𝑁 and 𝑁 ) and the engine

performances (Fuel flow 𝑊 and Thrust 𝐹 ) with a mean relative error smaller than 3%, and a standard

deviation error smaller than 5%. An exception however occurs during thrust prediction for the descent

phase, where the standard deviation is about 8.74%. This high percentage may be explained by the fact that

the thrust is relatively small during descent and may change its sign. Nevertheless, the thrust residual error
is around 50 lbf as seen on Figure 8, which is negligible and represents less than 0.5% of the engine nominal

thrust (13,500 lbf).

Table 1 Mean relative errors 𝝁, standard deviations 𝝈 and maximum error in [%] of
the engine model outputs for Maximum Takeoff (MTO), Maximum Climb (MCLB),
Cruise regime and Idle Thrust (IDLE)

MTO MCLB Cruise Regime IDLE

Engine
𝜇 𝜎 max. 𝜇 𝜎 max. 𝜇 𝜎 max. 𝜇 𝜎 max.
Parameter
𝑬𝑷𝑹 0.00 0.21 0.85 0.08 0.48 1.81 0.02 1.09 3.12 0.03 1.18 4.37

𝑭𝑷𝑹 0.21 1.23 3.86 1.01 2.26 6.48 2.31 4.56 6.11 0.19 2.11 7.05

𝑰𝑻𝑻 0.16 1.43 2.97 0.16 0.94 3.55 0.84 4.10 9.86 0.07 2.88 8.60

𝑶𝑷𝑹 0.12 0.37 1.60 0.15 0.73 2.79 0.48 2.74 6.48 0.03 1.07 4.02

𝑵𝟏 0.02 0.60 0.87 0.20 0.67 2.18 0.37 1.82 4.02 0.25 3.18 7.15

𝑵𝟐 0.09 0.39 0.77 0.11 0.52 1.88 0.29 1.01 1.95 0.14 1.01 2.27

𝑾𝑭 0.00 0.21 0.85 0.08 0.48 1.81 0.02 1.09 3.12 0.03 1.18 4.37

𝑭𝑵 0.07 0.18 0.52 0.04 0.45 1.75 0.01 0.99 2.59 1.46 8.74 36.5

100 100

80 80

60 60

40 40

20 20

0 0
-50 0 50 100 -100 -50 0 50 100

60 120

50 100

40 80

30 60

20 40

10 20

0 0
-200 0 200 -100 -50 0 50 100

Figure 8 Thrust residual error distributions for

Takeoff, Climb, Cruise and Descent phases
 The engine model was simulated for various flight profiles including takeoff, climb and descent phases,

and for the cruise phase during an engine degradation situation. The simulation results in Figure 9 toFigure

11 show that the model is perfectly predicting all engine parameters within the 5% tolerance limit, and it is

able to detect engine failure, as shown in

Figure 12. Indeed, an engine model such as the one developed in this paper, is useful to identify engine

failures, as it includes a large number of essential parameters to monitor the health of the engine.

CONCLUSION

A new methodology has been developed and validated in this paper in order to identify an engine

performance model using neural network techniques and flight tests. In the first step, the engine parameters

were estimated, including 𝐸𝑃𝑅, 𝐹𝑃𝑅, 𝑂𝑃𝑅, 𝐼𝑇𝑇, 𝑁 and 𝑁 . One should notice that this first step consisted

in modeling the engine thrust ratings and the component maps. Then, the second step consisted in predicting

the engine performances including fuel flow and thrust. Consequently, the developed methodology resulted

in a combination of CLM and NN techniques. Although the engine model was static, the results showed

that the model had an excellent ability of prediction with a relative error less than 5%. Further works can

be derived from this research, as developing an engine dynamic model capable of predicting its transient

states.
Figure 9 Example of simulation during climb phase from 25,000 ft to 40,000 ft
Figure 10 Example of simulation during takeoff phase from 0 ft to 1,500 ft
Figure 11 Example of simulation during descent phase from 40,000 ft to 1,500 ft
Figure 12 Example of simulation during cruise phase at 10,000 ft with applied degradation on
High Pressure Compressor at t=65 sec
 REFERENCES
[1] J. W. Chapman, T. M. Lavelle, and J. S. Litt, “Practical techniques for modeling gas turbine engine
performance,” in 52nd AIAA/SAE/ASEE joint propulsion conference, 2016, p. 4527.
[2] K. Kanelopoulos, A. Stamatis, and K. Mathioudakis, “Incorporating neural networks into gas turbine
performance diagnostics,” in Turbo Expo: Power for Land, Sea, and Air, 1997, vol. 78712, p.
V004T15A011.
[3] K. Mathioudakis and C. Romessis, “Probabilistic neural networks for validation of on-board jet engine
data,” Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering,
vol. 218, no. 1, pp. 59–72, 2004.
[4] T. Kobayashi and D. L. Simon, “Hybrid neural-network genetic-algorithm technique for aircraft engine
performance diagnostics,” Journal of Propulsion and Power, vol. 21, no. 4, pp. 751–758, 2005.
[5] A. Lazzaretto and A. Toffolo, “Analytical and neural network models for gas turbine design and off-
design simulation,” International Journal of Thermodynamics, vol. 4, no. 4, pp. 173–182, 2001.
[6] D. Zhou, Q. Yao, H. Wu, S. Ma, and H. Zhang, “Fault diagnosis of gas turbine based on partly
interpretable convolutional neural networks,” Energy, vol. 200, p. 117467, 2020.
[7] L. Zhao, C. Mo, T. Sun, and W. Huang, “Aero engine gas-path fault diagnose based on multimodal deep
neural networks,” Wireless Communications and Mobile Computing, vol. 2020, 2020.
[8] A. Gishvarov and J. Raherinjatovo, “Parametric Diagnostic State of Gas Turbine Power Plant,”
Procedia Engineering, vol. 206, pp. 1819–1824, 2017.
[9] G. Ghazi and R. M. Botez, “Identification and validation of an engine performance database model for
the flight management system,” Journal of Aerospace Information Systems, vol. 16, no. 8, pp. 307–326,
2019.
[10] P. A. Bardela, R. M. Botez, T. Bournisien, and R. Rusovici, “Identification and validation of the cessna
citation X turbofan modelling with flight tests,” in 2018 AIAA Modeling and Simulation Technologies
Conference, 2018, p. 1920.
[11] P. A. Bardela and R. M. Botez, “Identification and validation of the Cessna Citation X engine
component level modeling with flight tests,” in AIAA modeling and simulation technologies conference,
2017, p. 1942.
[12] R. M. Botez, P.-A. Bardela, and T. Bournisien, “Cessna Citation X simulation turbofan modelling:
identification and identified model validation using simulated flight tests,” The Aeronautical Journal, vol.
123, no. 1262, pp. 433–463, 2019.
[13] M. Zaag, R. M. Botez, and T. Wong, “Cessna citation X engine model identification using neural
networks and extended great deluge algorithms,” INCAS Bulletin, vol. 11, no. 2, pp. 195–207, 2019.
[14] M. Zaag and R. M. Botez, “Cessna citation X engine model identification and validation in the cruise
regime from flight tests based on neural networks combined with extended great deluge algorithm,” in AIAA
Modeling and Simulation Technologies Conference, 2017, p. 1941.
[15] H. Qiu, N. Eklund, W. Yan, P. Bonissone, F. Xue, and K. Goebel, “Estimating deterioration level of
aircraft engines,” in ASME Turbo Expo 2007: Power for Land, Sea, and Air, 2007, pp. 661–667.
[16] J. D. Mattingly, Elements of gas turbine propulsion, vol. 1. McGraw-Hill New York, 1996.
[17] J. W. Chapman and J. S. Litt, “Control design for an advanced geared turbofan engine,” in 53rd
AIAA/SAE/ASEE Joint Propulsion Conference, 2017, p. 4820.
[18] H. I. Saravanamuttoo, G. F. C. Rogers, and H. Cohen, Gas turbine theory. Pearson Education, 2001.
[19] P. Zhu and H. Saravanamuttoo, “Simulation of an advanced twin-spool industrial gas turbine,” 1992.
[20] R. A. Roberts and S. M. Eastbourn, “Modeling techniques for a computational efficient dynamic
turbofan engine model,” International Journal of Aerospace Engineering, vol. 2014, 2014.
[21] M. Hagan, H. Demuth, M. Beale, and O. De Jesus, “Neural Network Design, 2nd,” Ed., Lexington,
KY, 2016.
[22] A. Stamatis, K. Mathioudakis, and K. Papailiou, “Adaptive simulation of gas turbine performance,”
1990.
[23] A. Benyounes, A. Hafaifa, and M. Guemana, “Gas turbine modeling based on fuzzy clustering
algorithm using experimental data,” Applied Artificial Intelligence, vol. 30, no. 1, pp. 29–51, 2016.
[24] M. Segui, S. Bezin, and R. M. Botez, “Cessna Citation X performances improvement by an adaptive
winglet during the cruise flight,” International Journal of Aerospace and Mechanical Engineering, vol. 12,
no. 4, pp. 423–430, 2018.
[25] G. Ghazi, M. Bosne, Q. Sammartano, and R. M. Botez, “Cessna citation X stall characteristics
identification from flight data using neural networks,” in AIAA Atmospheric Flight Mechanics Conference,
2017, p. 0937.
[26] L. Ruichao, G. Yingqing, S. K. Nguang, and C. Yifeng, “Takagi-Sugeno fuzzy model identification
for turbofan aero-engines with guaranteed stability,” Chinese Journal of Aeronautics, vol. 31, no. 6, pp.
1206–1214, 2018.