JOURNAL OF AIRCRAFT

Vol. 56, No. 1, January–February 2019

Engineering Notes
Decision Tree– and Random Forest– High angle of attack and nonlinear aerodynamics modeling have
been studied for several decades, and numerous reports have been
Based Novel Unsteady Aerodynamics published in this study area. Goman and Khrabrov devised a state
Modeling Using Flight Data space approach to model the unsteady aerodynamics [6]. Nelson and
Pelletier have extensively reviewed the experimental information
on flow structure over delta wings and also on complete aircraft
Ajit Kumar∗ and Ajoy Kanti Ghosh† configurations [7]. Fischenberg and Jategaonkar have presented the
Indian Institute of Technology of Kanpur, Kanpur 208 016, study of unsteady aerodynamics using quasi-steady stall model on
C-160 transport aircraft [8]. Ghoreyshi and Cummings came up with
Uttar Pradesh, India a novel approach based on the time-dependent surrogate unsteady
DOI: 10.2514/1.C035034 aerodynamics model performed under different maneuvers [9].
Peyada and Ghosh developed a neural-network-based novel
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

method, Neural Gauss–Newton (NGN), to estimate aerodynamic

parameters [10]. Kumar and Ghosh used NGN later on to study
nonlinear aerodynamics [11]. Faller and Schreck developed a
I. Introduction recurrent multilayer perceptron neural network (MLP-NN)–based
method parameter identification [12]. Marques and Anderson
U NSTEADY aerodynamics has been a subject of interest for
many researchers in recent decades. Unsteady aerodynamics is
the study of the phenomena at the high angle of attack flying
demonstrated a multilayer-based temporal neural network under
transonic flow to estimate the parameters [13]. Voitcu and Wong
condition where flow separation takes place and affects the presented the dynamic behavior of an aeroelastic system using
aerodynamic forces and moments drastically. Under the stationary neural network [14]. Zhang et al. and also Winter and Breitsamter
attached flow conditions, aerodynamics effects can be adequately presented a radial basis function neural network (RBF-NN) on
modeled using linear parametric models. Tools are well developed for unsteady aerodynamics modeling under transonic flow [15,16].
a linear model. However, at higher angles of attack, aircraft undergo A successful combination of Zhang’s RBF-based nonlinear system
stall, and models become highly nonlinear and create an impasse in identification approach with the POD has been developed by
developing an effective mathematical model [1]. Lindhorst et al. [17]. A surrogated modeling approach on a similar
Unsteady aerodynamics phenomena generally have been line to the work of Lindhorst et al. is presented by Winter and
investigated using computational fluid dynamics (CFD) methods, Breitsamter [18]. Interestingly, NASA has for many years done
wind tunnel tests, and semi-empirical methods. These methods have research into unsteady aerodynamic models using an indicia-
certain limitations regarding demands of a large aerodynamic response model with output error in the frequency domain. The
database; coverage of wide ranges of the flight envelope; data frequency-domain-based method also has the advantage of
generation at high resolution and prominent data acquisition, which faster response than a time-domain-based method [19]. Unsteady
further leads to the demand of time; computational resources; and aerodynamic modeling using indicial functions are the research in
money. Also, wind-tunnel-extracted models need to undergo for the field of multivariate functions.
Neural-network-based methods have been broadly used in
boundary-layer, Reynolds-number, and scaling-effect corrections,
developing effective aerodynamic models. In this approach, the
which makes modeling efforts cumbersome. These database-based
hidden layer of the neural network makes it computationally
approaches also have challenges in gaining the physical insights of
intensive and sometimes causes instability in numerical solutions.
phenomena by simple inspection [1–3].
Another issue with it was that the neural network models are valid
Aerodynamic modeling using observed flight test data is one of the
only under certain conditions [20–24].
prominent approaches. Measured flight data generally capture the In the preceding paragraphs it is inferred that the unsteady
complete phenomena of the performance and can be used in model aerodynamics model can be presented either as a parametric or in a
postulation. The postulated model can be parametric or non- nonparametric form. In general, parametric models give the physical
parametric. The parametric model comes under the domain of aircraft insights of the problem, and all the associated parameters have some
parameter estimation area. Parameter estimation is a subsection of physical significance that is useful in characterizing the performance
system identification and is a process of estimating the aerodynamic evaluation and identification of the aircraft dynamics. However, there
parameters from the observed data. Aerodynamic parameters are are certain flight envelopes where the physics-based parametric
usually estimated using output error approach where the error is models are not accurate enough in capturing the complete dynamics.
minimized generally using the maximum likelihood approach. For example, in the analysis of the aircraft dynamics at a high angle of
Unsteady stall parametric model is used in parameter estimation from attack, near ground, or ground effect modeling, getting an accurate
the output error method. Aerodynamic modeling at a high angle of physics-based model is still an open area of research. With the access
attack is generally cumbersome, and the parametric approach makes to an enormous amount of flight data and progressing intelligent
it inefficient to capture total nonlinearity using traditional parametric techniques based on artificial intelligence and machine learning,
estimation methods [3–5]. these problems can be solved with intelligent techniques. Usually,
models obtained from these methods are training and logic based and
also mostly nonparametric in nature. Nonparametric models lack
Received 28 March 2018; revision received 17 July 2018; accepted for the physical understanding, but predicted models are found to have
publication 24 July 2018; published online 12 October 2018. Copyright better accuracy.
© 2018 by the American Institute of Aeronautics and Astronautics, Inc. All The current paper is an effort in the direction of nonparametric
rights reserved. All requests for copying and permission to reprint should be
submitted to CCC at www.copyright.com; employ the ISSN 0021-8669
modeling and presents a classification and regression tree (CART)–
(print) or 1533-3868 (online) to initiate your request. See also AIAA Rights and random forest (RF)–based novel approach to predict the unsteady
and Permissions www.aiaa.org/randp. aerodynamics model using quasi-stall flight test data. One noticeable
*Ph.D. Student, Department of Aerospace Engineering. benefit of decision tree–based methods is that these models are
†
Professor, Department of Aerospace Engineering. scalable to large problems and can handle smaller data set than NN
403
 404 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES

models. In these methods, the algorithm learns information directly A. Classification and Regression Tree
from the data, and so they are also known as machine learning CART is a classification and regression method developed by
techniques; they use observed flight data to construct the decision tree Breiman et al. [26] in the early 1980s, which constructs the decision
and set of decision trees [25–29] to predict the unsteady aerodynamic tree using observed data. Decision trees are usually presented by a set
force and moment coefficients. of questions that then split the learning sample into smaller parts. The
The efficacy of the proposed methods is examined on ATTAS aim of this method is not only to find the models that produce accurate
aircraft data by estimating the lift force coefficient (CL ), drag force predictions, but also to extract knowledge intelligently. Tree-based
coefficient (CD ), and pitching moment coefficient (Cm ) using the methods stand as one of the most effective and useful methods
measured flight data. Open-accessible flight test data of the research capable of producing both reliable and understandable results. There
aircraft DLR-ATTAS were used to estimate aircraft longitudinal are following features that make this approach quite attractive in
force and moment coefficients [1] and demonstrated the proposed practice:
algorithms for unsteady aerodynamics modeling from flight data. 1) Nonparametric and can model any complex relations between
Flight test data were segregated as input data set (angle of attack, inputs and outputs, without any a priori assumption
elevator deflection, pitch rate, and relative airspeed) and as output 2) Handles ordered or categorical variables or a mix of both
dataset (measured force and moment coefficients). These methods 3) Robust to noise, outliers, or errors in labels
were trained and validated with real flight data: 70% of the data were 4) Easy to interpret, understand, and visualize
used in training, and the remaining 30% were used for validation of the In CART, the regression tree is used when output variables are
model. The statistical modeling technique was used in quantifying the continuous, and the classification tree is used when output variables
uncertainty in the model. Estimated nonlinear aerodynamic models are categorical. In the case of regression tree, values obtained by
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

were compared with maximum likelihood estimation (MLE)– terminal nodes in the training of the data are the average response of
predicted models. MLE methods have been a popular classical method the observation falling in the region and this process continues until
for several decades. Estimated results from both the methods were the stopping criterion is met. The splitting process results in fully
found to be in close agreement with each other. CART- and RF- growing trees until the stopping criterion is reached. But the fully
estimated models of CL and CD are fairly close to MLE-estimated grown trees are likely to overfit data, leading to poor accuracy, and so
results. CART- and RF- estimated model of Cm has better correlation pruning is done to avoid it. The simple example of regression-tree-
than MLE-based estimated Cm . Statistical analysis of the presented based modeling is explained in [26] using the Boston housing
results deduces that these machine-learning-based methods can be a classical dataset. The CART algorithm is demonstrated on quasi-stall
viable alternative to this problem. Between CARTand RF, RF found to modeling of DLR-ATTAS aircraft using flight test data.
be slightly better for obvious reasons. Further, they do not require to The CART algorithm can be understood by the following steps:
solve the equation of motion; this advantage further motivates 1) Label the data: Identify the target variable and input variables.
promising directions for future research on nonlinear modeling and In the current application, target variables are coefficient of drag force
identification. (CD ), coefficient of lift force (CL ), and coefficient of the pitching
This paper has the following contributions: moment (Cm ), and input variables are angle of attack (α), elevator
1) Development of a CART- and RF-based model for lift force, deflection (δe), pitch rate (q), and relative airspeed (V).
drag force, and pitching moment coefficient 2) Best split: Identify the best split for each of the input variables.
2) Statistical analysis of the estimated results and comparison with A variable value that produces the greatest separation in the target
conventional method MLE variable called split is selected. Separation in regression trees is
The paper is structured as follows. Section II explains the CART defined using the sum of the squared errors.
and RF methodology and implementation on the problem. Minimum square error (MSE) is used as the best split criterion in this
Section III presents the estimation results and discussion using application.
the CART, RF, and MLE methods in detail, followed by 3) Repeat the process until meeting the stopping criterion.
conclusions. Error tolerance is set as 10−6
4) Pruning of the tree
MSE has been used to avoid the overfitting issue. However, CART
algorithm is prone to advantages like an unstable tree and split by one
variable. These issues are generally handled by incorporating many
II. Methodology: Nonparametric Nonlinear decision trees together. RF is one of such an attempt toward this goal
Aerodynamic Modeling and described in the next section.
CART- and RF-based machine learning methods are applied to
model the quasi-steady stall phenomena of ATTAS research B. Random Forest
aircraft. These methods use tree-based nonparametric approach of The RF proposed by Breiman in 2001 is a statistical estimation-
modeling using flight data. Nonparametric models are structureless based nonparametric tool consisting of many decision trees. As the
models that do not involve parametric equation or any physical name suggests, it creates a forest by generating many decision trees.
parameters such as parametric model. Parametric models are In general, when more trees are in the forest, more robust is the
defined as structured equations with parameters. Parameters prediction and thus higher accuracy. RF generates a forest of the
associated in the model have physical significance and contribute to CART by sampling the data and variables randomly and iteratively.
the analysis of aerodynamic characterization. Although parametric The RF predicts based on the averaging prediction made by each tree
models are very useful and give enough physical insight, they lack in the ensemble using input data. The performance measure is
in completely capturing the dynamics of the phenomena. Any checked by the following two measures [30–35]:
structured or parametrized model will have some amount of 1) A measure of accuracy, by mean squared error
inaccuracy depending on the severity and nonlinearly of the 2) A measure of node impurity by Gini index
dynamics, which reduces the efficiency in modeling. On the other The Gini index, which measures node impurity, is obtained by
hand, CART and RF are data-driven modeling approaches where subtracting the sum of the squared probabilities of each class
they capture the complete dynamics by understanding the gathered from one. P
flight data. These methods lack the physical understanding but It can be written as Gini
1 − ki
1 P2i , where k represents class.
increase the modeling accuracy, which is helpful in designing There are two main stages in the RF algorithm: first is the creation
control and guidance law, and fault detection algorithm. The next of an RF; second is to make a prediction from the classifier created.
section describes how these methods are applied to estimate RF algorithm:
longitudinal force and moment model under the quasi-stall 1) Randomly select “K” features from “m” total features
condition using flight data gathered on ATTAS aircraft. where K ≪ m
 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES 405

2) Among the “K” features, calculate the node “d” using the best 3) Out of bag prediction to compute predicted class probabilities
split point 4) Minimum number of observations per tree leaf: 5
3) Split the node into daughter nodes using the best split 5) Split criterion: MSE
4) Repeat first three steps until “l” number of nodes has been Independent variables in both the algorithms considered were the
reached angle of attack, elevator deflection, pitch rate, and relative airspeed.
5) Build forest by repeating steps 1 to 4 Target variables considered were CL , CD , and Cm , which have been
Advantages of RF are that 1) it avoids the overfitting problem; 2) for obtained by following relations:
both classification and regression tasks, the same RF algorithm can be
used; and 3) it can be used for identifying the most important features CL
−CZ cos α  CX sin α
from the training dataset, in other words, feature engineering.
Sets of regression tree were constructed using target and CD
−CX cos α − CZ sin α
independent variables. Target and independent variables are the same I Y q_ − Feltx sin σ t  ltz cos σ t   rqI X − I Z   I XZ p2 − r2 
as the CART algorithm. Split criteria were by a mean square error. Cm

0.5ρV 2
(1)
III. Estimation Results and Discussion
where
Machine-learning-based CART and RF methods have been
proposed to estimate the nonlinear aerodynamic coefficient model of max − Fe cos σ t
the aircraft. These methods were examined on openly accessible flight CX

Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

0.5ρV 2 S
test data of the research aircraft DLR-ATTAS, Germany [1], to estimate
maz  Fe sin σ t
the lift force, drag force, and pitching moment coefficients. The CZ
(2)
MATLAB [35] software package has been used to implement these 0.5ρV 2 S
algorithms. Further, estimated results of force and moment coefficients
have been compared with the conventional method based on MLE. where
I Y
moment of inertia along the y axis, kg ⋅ m2
q_
pitch acceleration, rad∕s2
A. Nonlinear Aerodynamic Modeling Using CART- and RF-Based
Novel Algorithm
Fe
thrust force, N
ltx , ltz
location of the engine relative to the c.g.
CART- and RF-based methods have been examined on DLR- σ t
thrust inclination angle, deg
ATTAS quasi-steady stall data. Force and moment coefficients, CL , I X
moment of inertia along the x axis, kg ⋅ m2
CD , and Cm , of the aircraft have been estimated using these data- I Z
moment of inertia along the z axis, kg ⋅ m2
driven developed model. Of the flight data, 70% have been used in I XZ
cross moment of inertia in the plane xz, kg ⋅ m2
training the model, and the remaining 30% of the data have been used p
roll rate, rad∕s
in validating the force and moment coefficient models. r
yaw rate, rad∕s
The CART algorithm has been implemented with the following ρ
density, kg∕m3
configurations: V
air relative speed, m∕s
1) Pruning criteria: MSE m
mass, kg
2) Quadratic error tolerance 10e-6 S
reference area m2
3) Split criterion: MSE ax
axial acceleration, m∕s2
The RF algorithm has been implemented with the following az
normal acceleration, m∕s2
configurations:
1) Ensemble with 100 trees Figure 1 shows the time history of the data used in both of these
2) Tree method: Regression algorithms. The figure shows that the change in the elevator

5
q[dps]

0
-5
0 20 40 60 80 100 120 140 160 180 200
20
q [deg]

10
0
0 20 40 60 80 100 120 140 160 180 200
20
a [deg]

0
-20
0 20 40 60 80 100 120 140 160 180 200
100
V[m/s]

80
60
0 20 40 60 80 100 120 140 160 180 200
a [m/s 2 ]

0
-10
-20
z

0 20 40 60 80 100 120 140 160 180 200

a [m/s 2 ]

2
1
0
x

0 20 40 60 80 100 120 140 160 180 200

10
d e [deg]

0
-10
0 20 40 60 80 100 120 140 160 180 200
time [s]
Fig. 1 Time history plot for ATTAS quasi-stall data.
 406 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES

2 Measured Flight data gathering on dynamic stall are more risky and tedious.
CART Model
1.5 RF Model Quasi-steady stall enables hysteresis time constant τ2 only. The
CL

1
transient effect is neglected by setting τ1 as zero, and this model is
called quasi-steady stall model. In this model, a1 , τ2 , and α are
0.5
0 20 40 60 80 100 120 140
generally adequate to capture the stall hysteresis. The following θstall
vector is a set of aerodynamic and stall parameters that are predicted
0.4
by minimizing the output error in MLE framework.
0.3
CD

0.2
0.1
θstall

0 CD0 eCL0 CLα CLαMα Cm0 Cmα Cmq Cmδe a1 α τ2 CDX CmX T
0 20 40 60 80 100 120 140
(5)
0.2
0.1
θstall vector estimated as: CD0
0.04351.00 , e
0.8390.82,
Cm

0
CL0
0.1582.08, CLα
3.2981.14, CLαMα
9.072.20,
-0.1
Cm0
0.053.50, Cmα
−0.1763, Cmq
−6.1464.47, Cmδe

-0.2
0 20 40 60 80 100 120 140 −0.3914.08, a1
23.7163.41, α
0.3090.35, τ2

time [s] 
24.025c∕V1.46, CDX
0.07923.82, CmX
−0.12613.98
Fig. 2 CART- and RF-based developed CL , CD , and Cm model for with acceptable Cramer–Rao bounds. Numbers inside  represent
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

ATTAS research aircraft. the relative standard deviation. These data are obtained by running
the MATLAB code given in Ref. [1]. Interested readers can get the
access to these details from p. 421 or p. 475 of the newer edition
deflection affects all the other presented corresponding flight of Ref. [1].
variables in a similar variation pattern. The presented data were Comparison of the CART-, RF-, and MLE-estimated force and
acquired at the sampling rate of 25 Hz. Gathered flight data have been moment coefficients CL , CD , and Cm are presented in Figs. 3–5.
segregated as a target and independent variables to use in the
proposed algorithm. Elevator deflection observed is around 8 deg in
2
both the directions. Axial and normal accelerations follow the change
Measured
in the elevator deflections. Variations of 18 to −18 deg have been CART predicted
recorded. Change in pitch angle is 15 deg approximately, and −4 to 1.8 RF predicted
MLE predicted
4 dps change have been observed in pitch rate data.
Training data for the developed model of CL , CD , and Cm using 1.6
CARTand RF are presented in Fig. 2. These models were designed by
judicious selection of training and validation data. Training data of
1.4
70% capture all the dynamic and include most of the variation. So
CL

training of these two models was achieved using 70% of the flight
data gathered on ATTAS aircraft. Further, these models are validated 1.2
with the remaining 30% of the flight data. CART- and RF-estimated
models are validated with the measured coefficients and also with the 1
derived model from MLE-estimated parameters and input variables.
Validation data deduce that the models obtained from the CART and 0.8
RF are well trained.

B. Nonlinear Modeling Using MLE 0 10 20 30 40 50 60 70

MLE is a broadly used estimation tool for parameter estimation time [s]
and aerodynamic modeling from decades. This method uses the Fig. 3 Validation of lift force coefficient (CL ) model using a CART, RF,
approach of the minimization of output error between the measured and MLE.
and model response. The force and moment coefficients of the
ATTAS aircraft were estimated using quasi-steady model described
1.6
in Eqs. (3) and (4) [1].
CART
 p 1.5 RF
1 X 2 MLE
CL α; X
CL0  CLα α 1.4
2
1 1.3
CD
CD0  C2 α; X  CDX 1 − X
eπAR L
CL predicted

1.2
qc
Cm
Cm0  Cmα α  Cδe δe  Cmq  CmX 1 − X (3) 1.1
2V
1
The flow separation point X in the above equation is estimated
using the following relation in Eq. (4): 0.9

0.8
dX 1
τ1  X
f1 − tanha1 α − τ2 α_ − α g (4) 0.7
dt 2

where a1 is airfoil static stall characteristics 0.8 1 1.2 1.4 1.6 1.8 2
τ1 , transient time constants, s C L measured
τ2 is hysteresis time constant, s Fig. 4 Prediction correlation of lift force coefficient (CL ) among CART,
α is the break point RF, and MLE.
 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES 407

0.3 0.26

Measured 0.24
CART predicted
RF predicted
0.25 MLE predicted 0.22

0.2

CD predicted
0.2 0.18
CD

0.16

0.15 0.14

0.12

0.1 0.1
CART
0.08 RF
MLE
0.05 0.06
0 10 20 30 40 50 60 70 0.05 0.1 0.15 0.2 0.25 0.3
time [s] CD measured
Fig. 5 Validation of drag force coefficient (CD ) model using a CART, Fig. 6 Prediction correlation of drag force coefficient (Cm ) among
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

RF, and MLE. CART, RF, and MLE.

Figure 3 shows the validation of the lift force coefficient (CL ) using 0.1
proposed methods CART and RF from measured output data. The
plot also shows the MLE-estimated CL that is derived from the
predicted aerodynamic parameters and input variables. Estimated 0.05
and measured CL curves are fitting well on each other.
Figure 4 shows a prediction correlation of measured lift force
coefficient (CL ) with the CART, RF, and MLE. The correlation 0
coefficient of all these three methods with measured data is close to
Cm

one. The correlation coefficient of measure CL with CART, RF, and

MLE is calculated to be 0.9908, 0.9930, and 0.9898, respectively,
-0.05
which shows confidence in the estimates. These results are also Measured
CART predicted
tabulated in Table 1. Estimated CL from CART and RF are on par with RF predicted
MLE-estimated CL . MLE predicted
Figure 5 shows the validation of the estimated drag force -0.1

coefficient (CD ) using the CART and RF algorithms from measured

CD . MLE-derived CD also plotted like Fig. 3 to compare the results.
CD estimated using CART and RF are fitting well with the measured -0.15
0 10 20 30 40 50 60 70
and MLE-derived CL .
Figure 6 shows a prediction correlation of measured CD with the Time [s]
CART-, RF-, and MLE-estimated CD . The correlation coefficient of Fig. 7 Validation of pitching moment coefficient (Cm ) model using
all these three methods with measured data is found close to one. The CART, RF, and MLE.
correlation coefficient of measure CD with CART-, RF-, and MLE-
estimated CD is calculated to be 0.9402, 0.9559, and 0.9486, outperform the MLE method by keeping correlation coefficient as
respectively, which shows confidence in the estimates. These results comparison parameter.
also are presented in Table 1. Statistical analysis shows that estimated Figure 9 presents a hysteresis plot between the lift coefficient and
CL from CART and RF are on par with the MLE-derived CD . angle of attack. Predicted CL using all the three methods, CART, RF,
Figure 7 shows the validation of the estimated pitching moment and MLE, are plotted in the figure. All these curves are following
coefficient (Cm ) using the CART and RF algorithms from measured hysteresis and fitting well on each other.
Cm . MLE-derived Cm also are plotted in the figure to compare the Table 1 presents the prediction correlation and RMSE calculation
results. Cm estimated using CART and RF are seen to be fitting well for the CART, RF, and MLE methods. It has been observed that
with the measured and MLE-derived Cm . CART- and RF-estimated coefficients are on par with the MLE-
Figure 8 shows a prediction correlation of measured Cm with the estimated coefficients. Between these two methods, RF has better
CART-, RF-, and MLE-based estimated Cm . The correlation estimate than CART, which also justifies the thesis that the forest
coefficient of all these three methods with the measured data is found modeling approach is superior to the decision tree modeling
to be close to one. The correlation coefficient of measure Cm with the approach. Also, it is observed that the estimation of CART- and
CART-, RF-, and MLE-based estimated Cm is calculated to be RF-estimated Cm has better correlation than MLE-estimated Cm .
0.8961, 0.9332, and 0.4670, respectively. The estimated Cm using Statistical data presented in Table 1 conclude that the proposed
CART and RF has better correlation than MLE-based estimated Cm . machine-learning-based methods, CART and RF, can be a viable
It is interesting to see in this case that CART and RF methods solution to this problem.

Table 1 Prediction correlation and RMSE calculation for CART, RF, and MLE methods
CART Random forest MLE
Force/moment coefficient Prediction correlation RMSE Prediction correlation RMSE Prediction correlation RMSE
CL 0.9908 0.0278 0.9930 0.0244 0.989818 0.029438
CD 0.9402 0.0100 0.9559 0.0086 0.948576 0.009290
Cm 0.8961 0.0058 0.9332 0.0047 0.466906 0.011350
 408 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES

0.1
comparison. CART- and RF-estimated models for CL and CD were
fairly close to the ML-estimated results and the estimated Cm was
higher in correlation than the MLE estimate.
0.05 From the statistical analysis, it can be concluded that the RF is
slightly superior to CARTand that these two methods can be seen as a
viable alternative to MLE, which is a parametric approach to model
Cm predicted

0 aerodynamic coefficients. Further, these proposed methods do not

require to solve the equation of motion and deal with fewer
complexities by comparing with other machine learning methods
-0.05 such as Artificial neural networks (ANNs). The advantages of these
methods motivate the promising directions for future research and
can easily be generalized to many other applications of nonlinear
-0.1
modeling.
AR predicted
RF predicted

-0.15
MLE Predicted References
-0.12 -0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 [1] Jategaonkar, R. V., Flight Vehicle System Identification: A Time Domain
Cm measured Methodology, AIAA Education Series, AIAA, Reston, VA, 2006,
Fig. 8 Prediction correlation of pitching moment coefficient (Cm ) Chaps. 2, 12.
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

among CART, RF, and ML. [2] Grauer, J. A., and Morelli, E. A., “Generic Global Aerodynamic Model
for Aircraft,” Journal of Aircraft, Vol. 52, No. 1, 2014, pp. 13–20.
[3] Klein, V., and Morelli, E. A., Aircraft System Identification: Theory and
Practice, AIAA, Reston, VA, 2006, pp. 351–382.
2 [4] Fischenberg, D., “Identification of an Unsteady Aerodynamics Stall
Model from Flight Test Data,” RTO MP-11, DLR-Germany, Paper
1.8
No. 17, 1999.
[5] Mehra, R. K., Stepner, D. E., and Tyler, J. S., “Maximum Likelihood
Identification of Aircraft Stability and Control Derivatives,” Journal of
1.6 Aircraft, Vol. 11, No. 2, 1974, pp. 81–89.
doi:10.2514/3.60327
[6] Goman, M. G., and Khrabrov, A. N., “State-Space Representation of
1.4 Aerodynamic Characteristics of an Aircraft at High,” Journal of
CL

Aircraft, Vol. 31, No. 5, 1994, pp. 1109–1115.

1.2 doi:10.2514/3.46618
[7] Nelson, R. C., and Pelletier, A., “The Unsteady Aerodynamics of
Slender Wings and Aircraft Undergoing Large Amplitude Maneuvers,”
1 Progress in Aerospace Sciences, Vol. 39, Nos. 2–3, 2003, pp. 185–248.
[8] Fischenberg, D., and Jategaonkar, R. V., “Identification of Aircraft Stall
Measured Behavior from Flight Test Data,” RTO MP-11, DLR-Germany, Paper
0.8 CART predicted No. 17, 1999.
RF predicted
MLE predicted [9] Ghoreyshi, M., and Cummings, R. M., “Unsteady Aerodynamics
Modeling for Aircraft Maneuvers: A New Approach Using Time-
4 6 8 10 12 14 16 18 Dependent Surrogate Modeling,” Aerospace Science and Technology,
a [deg] Vol. 39, Dec. 2014, pp. 222–242.
doi:10.1016/j.ast.2014.09.009
Fig. 9 Stall hysteresis comparison using CART, RF and ML prediction.
[10] Peyada, N. K., and Ghosh, A. K., “Aircraft Parameter Estimation Using
a New Filtering Technique Based Upon a Neural Network and
Gauss-Newton Method,” The Aeronautical Journal, Vol. 113, No. 1142,
However, we accept the limitation of these machine-learning- 2009, pp. 243–252.
based methods. They lack the physical insights of the problem. The doi:10.1017/S0001924000002918
current work is an effort in the direction of the modeling of the [11] Kumar, R., and Ghosh, A. K., “Nonlinear Aerodynamic Modeling of
phenomena where physics-based models are generally not sufficient Hansa-3 Aircraft Using Neural Gauss-Newton Method,” Journal of
to capture the whole dynamics. These methods are completely data Aerospace Science and Technology, Vol. 63, No. 3, 2011, pp. 194–204.
driven, and modeling performance depends on the quality of [12] Faller, W. E., and Schreck, S. J., “Unsteady Fluid Mechanics
available data. These methods also see a potential application in Applications of Neural Networks,” Journal of Aircraft, Vol. 34, No. 1,
ground effect modeling, which is still an open research. 1997, pp. 48–55.
doi:10.2514/2.2134
[13] Marques, F. D., and Anderson, J., “Identification and Prediction of
Unsteady Transonic Aerodynamic Loads by Multi-Layer Functionals,”
IV. Conclusions Journal of Fluids and Structures, Vol. 15, No. 1, 2001, pp. 83–106.
doi:10.1006/jfls.2000.0321
In this paper, the classification and regression tree (CART)- and
[14] Voitcu, O., and Wong, Y. S., “Neural Network Approach for Nonlinear
random forest (RF)-based novel algorithms were proposed to estimate Aeroelastic Analysis,” Journal of Guidance, Control, and Dynamics,
the nonlinear aerodynamic force and moment coefficients. The Vol. 26, No. 1, 2003, pp. 99–105.
efficacy of these machine-learning-based algorithms was examined on doi:10.2514/2.5019
the quasi-stall data of DLR-ATTAS aircraft data: 70% of the quasi-stall [15] Zhang, W., Wang, B., Ye, Z., and Quan, J., “Efficient Method for Limit
aerodynamics data were used in training, and the rest 30% were used Cycle Flutter Analysis Based on Nonlinear Aerodynamic Reduced-
for validation of the model. Order Models,” AIAA Journal, Vol. 50, No. 5, 2012, pp. 1019–1028.
Estimated lift force coefficient (CL ), drag force coefficient (CD ), doi:10.2514/1.J050581
and pitching moment coefficient (Cm ) of the aircraft were statistically [16] Winter, M., and Breitsamter, C., Reduced-Order Modeling of Unsteady
analyzed. The statistical modeling technique was used in quantifying Aerodynamic Loads Using Radial Basis Function Neural Networks,
Deutsche Gesellschaft für Luft-und Raumfahrt-Lilienthal-Oberth eV,
the correlation and RMSE of the results. The CART- and RF- Bonn, Germany, 2011.
estimated nonlinear aerodynamic coefficients were moderately [17] Lindhorst, K., Haupt, M. C., and Horst, P., “Efficient Surrogate
in high correlation with the measurement with minimal RMSE. Modelling of Nonlinear Aerodynamics in Aerostructural Coupling
Further, results were compared with maximum likelihood estimation Schemes,” AIAA Journal, Vol. 52, No. 9, 2014, pp. 1952–1966.
(MLE)-predicted force and moment coefficients and were on par in doi:10.2514/1.J052725
 J. AIRCRAFT, VOL. 56, NO. 1: ENGINEERING NOTES 409

[18] Winter, M., and Breitsamter, C., Unsteady Aerodynamic Modeling Technique in Estimating Peak Particle Velocity Caused by Blasting,”
Using Neuro-Fuzzy Approaches Combined with POD, Deutsche Engineering with Computers, Vol. 33, No. 1, 2017, pp. 45–53.
Gesellschaft für Luft-und Raumfahrt-Lilienthal-Oberth eV, Bonn, doi:10.1007/s00366-016-0455-0
Germany, 2015. [28] Breiman, L., Friedman, J. H., Olshen, R., and Stone, C. J., Classification
[19] Morelli, E. A., and Klein, V., “Application of System Identification to and Regression Tree, Wadsworth Brooks/Cole Advanced Books &
Aircraft at NASA Langley Research Center,” Journal of Aircraft, Software, Pacific, CA, 1984, Chap. 8.
Vol. 42, No. 1, 2005, pp. 12–25. [29] Razi, M. A., and Athappilly, K., “A Comparative Predictive Analysis of
doi:10.2514/1.3648 Neural Networks (NNs), Nonlinear Regression and Classification and
[20] De JesusMota, S., and Botez, R. M., “New Helicopter Model Regression Tree (CART) Models,” Expert Systems with Applications,
Identification Method Based on a Neural Network Optimization Vol. 29, No. 1, 2005, pp. 65–74.
Algorithm and on Flight Test Data,” The Aeronautical Journal, Vol. 115, doi:10.1016/j.eswa.2005.01.006
No. 1167, 2011, pp. 295–314. [30] Antipov, E. A., and Pokryshevskaya, E. B., “Mass Appraisal of
doi:10.1017/S0001924000005789 Residential Apartments: An Application of Random Forest for
[21] Boely, N., Botez, R. M., and Kouba, G., “Identification of a Nonlinear Valuation and a CART-Based Approach for Model Diagnostics,”
F/A-18 Model by Use of Fuzzy Logic and Neural Network Methods,” Expert Systems with Applications, Vol. 39, No. 2, 2012,
Proceedings of the Institution of Mechanical Engineers, Part G, Journal pp. 1772–1778.
of Aerospace Engineering, Vol. 225, No. 5, 2011, pp. 559–574. doi:10.1016/j.eswa.2011.08.077
[22] Kumar, A., and Ghosh, A. K., “Data-Driven Method Based Aerodynamic [31] Loh, W.-Y., “Classification and Regression Trees,” Wiley Interdisciplinary
Parameter Estimation from Flight Data,” 2018 AIAA Atmospheric Flight Reviews: Data Mining and Knowledge Discovery, Vol. 1, No. 1, 2011,
Mechanics Conference, AIAA Paper 2018-0768, 2018. pp. 14–23.
[23] Saderla, S., Dhayalan, R., and Ghosh, A. K., “Non-Linear Aerodynamic [32] Verikas, A., Gelzinis, A., and Bacauskiene, M., “Mining Data with
Downloaded by Indian Institute of Science on December 8, 2023 | http://arc.aiaa.org | DOI: 10.2514/1.C035034

Modelling of Unmanned Cropped Delta Configuration from Experimental Random Forests: A Survey and Results of New Tests,” Pattern
Data,” The Aeronautical Journal, Vol. 121, No. 1237, 2017, pp. 320–340. Recognition, Vol. 44, No. 2, 2011, pp. 330–349.
doi:10.1017/aer.2016.124 doi:10.1016/j.patcog.2010.08.011
[24] Nelles, O., Nonlinear System Identification from Classical approaches [33] Joshi, A., Monnier, C., Betke, M., and Sclaroff, S., “Comparing Random
to Neural Networks and Fuzzy Models, Springer–Verlag, Berlin, 2001, Forest Approaches to Segmenting and Classifying Gestures,” Image and
Chaps. 1, 12. Vision Computing, Vol. 58, Feb. 2017, pp. 86–95.
[25] Loh, W. Y., “Classification and Regression Trees,” Encyclopedia of doi:10.1016/j.imavis.2016.06.001
Statistics in Quality and Reliability, edited by F. Ruggeri, R. Kenett, and [34] Markham, I. S., Mathieu, R. G., and Wray, B. A., “Kanban Setting
F. W. Faltin, Wiley, Chichester, U.K., 2011, pp. 315–323. Through Artificial Intelligence: A Comparative Study of Artificial
[26] Timofeev, R., Classification and Regression Trees (CART) Theory and Neural Networks and Decision Trees,” Integrated Manufacturing
Applications, Humboldt Univ., Berlin, 2004, Chap. 5. Systems, Vol. 11, No. 4, 2000, pp. 239–246.
[27] Khandelwal, M., Armaghani, D. J., Faradonbeh, R. S., Yellishetty, M., doi:10.1108/09576060010326230
Majid, M. Z. A., and Monjezi, M., “Classification and Regression Tree [35] MATLAB, The MathWorks. Inc., Natick, MA, 2016, p. 488.