JOURNAL OF AIRCRAFT

New Robust Method to Study Flight Flutter Testing

Soledad Le Clainche∗
Universidad Politécnica de Madrid, E-28040 Madrid, Spain
Rubén Moreno-Ramos†
Altran Innovación, E-28022 Madrid, Spain
Paul Taylor‡
Gulfstream Aerospace, Savannah, Georgia 31402
and
José M. Vega§
Universidad Politécnica de Madrid, E-28040 Madrid, Spain
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

DOI: 10.2514/1.C034863
This paper presents the application of a new method, called higher-order dynamic mode decomposition (HODMD),
to extract aircraft frequencies and damping in experimental data obtained from a flight flutter test. The method is an
extension of standard dynamic mode decomposition, which is a method typically used to extract flow patterns and
frequencies from unsteady fluid dynamics measurements or numerical simulations. In the fluid dynamic field,
HODMD has proven to be a very efficient, robust, and accurate method to extract modes, frequencies, and damping
from very noisy and large signals, with reduced manual interaction. In a similar way, this paper shows the good
performance of this method when it is applied to the noisy data of this complex experiment. This fact sheds light on the
new possibilities for the aerospace industry to incorporate HODMD to perform robust analyses.

I. Introduction Since the 1980s, the community has been paying attention to
developing new methods in order to improve the efficiency in the
F LIGHT flutter testing is a requirement for civil transport-
category aircraft certified in the United States under the Code of
Federal Regulations, Part 25. Flight flutter testing often involves
prediction of flutter instability [1–10]. Some of these methods assume
data extrapolation, in the sense that their approaches are based on
taking the aircraft to parts of the flight envelope it has not been to extrapolating damping trends from available aeroelastic damping data
before and performing stability checks at those points. [11]. Thus, the prediction of flutter is approached from subcritical
Normal flight clearance typically involves flying very fast at a broad conditions. Some authors [12–15] presented a comparison between the
range of altitudes, including multiple fuel and payload configurations, most effective and famous flutter methodologies. In most of the cases,
meaning that the testing can be very expensive in terms of both fuel and the flight-test methodologies consider an artificial excitation force
time. This places an emphasis on efficient testing techniques, including (a known input signal) and measure the system response. So, it is
test point sequencing, piloting techniques, and near-real-time data possible to identify the dynamics of the system (frequencies and
reduction and modal parameter extraction techniques. damping) from an input/output relation. However, in some cases, the
The execution of the flight flutter testing consists of mounting on input signal is the atmospheric turbulence, so the methodologies
the aircraft an array of accelerometers that measures the structural compared include systems of output-only data [16,17].
dynamics response to several inputs coming from the control surfaces The classical flutter prediction methodologies include the moving-
or any mounted flutter excitation system (FES). Accelerometer block approach [18], the least-squares curve-fitting method [19],
responses and some other types of data extracted from the flight, such the nonlinear autoregressive moving-average exogenous [20,21], the
as control surface displacements for pulses or control surface sweeps, autoregressive moving-average (ARMA) method [22,23], some
force measurements from the FES and flight parameters, are stored advanced signal processing techniques [24], etc. McNamara and
with the aim of analyzing them. The main goal is to obtain the aircraft Friedman [25] have shown the ARMA method to be highly efficient
damping rates ζ m and frequencies ωm that are related to the flutter and effective, leading to the reduction of computational costs in both
phenomenon. The computational resources required within this data memory and time. This aspect makes the ARMA method suitable for
acquisition process and analysis are directly linked to both the performing a damping/frequency analysis in flight flutter testing. The
quantity of the data stored (memory) and the time employed in the method, as presented by McNamara and Friedman [25], assumes that
damping/frequency calculations (computational time). These two there is a linear relationship between the state X~ k (at time instant tk )
issues are highly dependent on the numerical method used for the and the subsequent state X~ k
1 (at time instant tk
1 ). Such a linear
analysis. relationship is independent of k, and it may be represented by a matrix
for which the eigenvalues determine the frequencies and damping
Received 5 December 2017; revision received 5 July 2018; accepted for rates of the system.
publication 6 July 2018; published online 5 October 2018. Copyright © 2018 There is a tight link between the ARMA method and dynamic
by the American Institute of Aeronautics and Astronautics, Inc. All rights mode decomposition (DMD) [26], which is a method commonly
reserved. All requests for copying and permission to reprint should be used in fluid dynamics to study the frequencies and damping rates
submitted to CCC at www.copyright.com; employ the ISSN 0021-8669 (also known as growth rates) of a flow system. The standard DMD
(print) or 1533-3868 (online) to initiate your request. See also AIAA Rights
algorithm is based on the Koopman assumption [27], which relates
and Permissions www.aiaa.org/randp.
*Lecturer in Applied Mathematics, School of Aeronautics, E.T.S.I. each snapshot (state vector) with the subsequent snapshot via the
Aeronáutica y del Espacio, Pza. Cardenal Cisneros; soledad.leclainche@ linear operator R as
upm.es (Corresponding Author).
†
Specialist, Loads and Dynamics, Aeronautics Space and Defense Xk
1 ≃ RXk for k  1; : : : ; K − 1 (1)
Division, 1 C/Campezo. Senior Member AIAA.
‡
Staff Scientist, Dynamics, P.O. Box 2206, M/S R-04. Associate Fellow When the snapshots are organized in snapshot matrices of the form
AIAA.
§
Professor in Applied Mathematics, School of Aeronautics, E.T.S.I.
Aeronáutica y del Espacio, Pza. Cardenal Cisneros. p  xp ; xp
1 ; : : : ; xk ; xk
1 ; : : : ; xQ−1 ; xQ 
XQ (2)
Article in Advance / 1
 2 Article in Advance / LE CLAINCHE ET AL.

Equation (1) can also be written as accurate results with a shorter-length signal than the DMD (see
figure 5.1 in [32]). In this paper, this new method is applied to a set of
2 ≃ RX1
XK K−1 ≃ XK−1 C (3) experimental flight flutter testing data that is obtained during a flutter
1
flight-test campaign, and its good performance is put in evidence.
where C is a companion matrix. The dynamics of the system Three different flight-test points re analyzed. The high performance of
(damping rates and frequencies) are determined from the eigenvalues the HODMD as compared to standard DMD (and, consequently, to the
of either R or C, as in the ARMA method. When a DMD analysis is ARMA method) provides a method that can be potentially used in
performed via the linear operator R instead of the companion matrix, flight flutter testing, reducing the costs in both memory and time.
the results are improved because, as will be explained in Sec. II, it is
possible to remove spatial redundancies (too many samples) or to
II. Higher-Order Dynamic Mode Decomposition
clean their noise.
The similarity between the ARMA method and the DMD The main goal of the HODMD is to express a set of instantaneous
algorithm is remarkable because both algorithms relate subsequent (spatiotemporal) data as an expansion of modes, as shown in Eq. (6);
vector states by a companion matrix. The differences found between thus, it is possible to study the main frequencies and growth rates
these two methods lie in the generation of the vector state that, in each composing a signal. For convenience, a set of K time equispaced
case, are ordered in different ways; consequently, the coefficients and snapshots (with Δt) obtained in a flight test (the signal of an
the structure of the companion matrix are also different. The accelerometer) are collected in the following matrix:
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

companion matrix is defined in the ARMA method as

1  v1 ; v2 ; : : : ; vk ; vk
1 ; : : : ; vK−1 ; vK 
VK (7)
2 3
−a1 1 0 ::: 0
6 −a where each vector vk is composed by the signal at time instant tk
6 2 0 1 ::: 0 77 collected in each one of the accelerometers of the flight test.
6 7
C~ ≡ 6
6 : : : ::: ::: ::: ::: 7
7 (4) Consequently, the dimension of this matrix is J × K, where J is the
6 7 number of accelerometers. Then, the HODMD algorithm proceeds in
4 −aK−1 0 0 ::: 1 5
two main steps:
−aK 0 0 ::: 0 1) Step 1 is the dimension reduction. In a first step, the spatial
dimension J (number of accelerometers) of the original dataset of
whereas in the standard DMD, it is defined as snapshots is reduced to a set of linearly dependent vectors of
2 3 dimension N, reducing the noise of the signal. In this way, singular
0 0 ::: 0 c1 value decomposition (SVD) [33] is applied to the snapshot matrix as
6 1 0 0 c2 7
6 ::: 7
6 7 1 ≃ WΣT
VK (8)
T
C≡6
6 0 1 ::: 0 c3 77 (5)
6 7 where W T W  T T T  the N × N unit matrix, and the diagonal of
4 ::: ::: ::: ::: ::: 5
matrix Σ contains the singular values σ 1 ; : : : ; σ K . The number of
0 0 ::: 1 cK retained SVD modes N is calculated through the standard SVD error
estimated for a certain tolerance ε (set by the user) as
where a1 ; : : : ; aK and c1 ; : : : ; cK are the coefficients calculated for
each companion matrix. Because the state vector contains relevant σ N
1 ∕σ 1 ≥ ε (9)
temporal information, the essence of these two algorithms is, however,
the same. Thus, different distributions of each one of the components Then, the reduced snapshot matrix of dimension N × K is
of the state vector will vary the coefficients of the companion matrix written as
but will provide the same results. It is important to mention that, to
obtain consistent results in both the ARMA method and the standard V^ K
1  ΣT ;
T with V K ^K
1  WV1 (10)
DMD, it is necessary to maintain the same order of the state vector
components at each time instant (columns of the state vector).
2) Step 2 is the DMD-d algorithm. The following higher-order
In temporal signals, it is possible to decompose the original data xk
Koopman assumption (written in matrix form)
(xk represents each one of the state variables, which are equispaced in
time with Δt) as an expansion of M Fourier-like modes in the K−d
R V K−d−1
; : : : ;
R V K−1
following way: d
1 ≃ R1 V 1
VK 2 2 d d (11)

X
M is applied to the reduced snapshot matrix. Premultiplying Eq. (11) by
xk ≃ xapprox:
k ≡ am um eζm
iωm k−1Δt ; k  1; : : : ; K (6) W T , where the reduced Koopman operator yields R^ k  W T Rk W, one
m1 leads to the following expression:

where the number of terms M can be referred to as the spectral V^ K ^ ^ K−d
R^ 2 V^ K−d
1
d
1 ≃ R1 V 1 2
; : : : ;
R^ d V^ K−1
d (12)
complexity, K ≤ n is the temporal dimension, um are the spatial
coefficients or modes (unitary vector), and am are their corresponding The main system dynamics (frequencies, damping rates, and
amplitudes. The dimension of the subspace generated by the M DMD modes) are contained in all these linear operators that can be
modes is the spatial complexity N. encompassed in the modified Koopman matrix (that contains the
Using the standard DMD algorithm, not only the damping rates and dynamics of the system) of dimension Nd × Nd, which is defined as
frequencies are easily obtained but also the associated spatial modes
and their corresponding amplitudes. The benefits of the DMD have 2 3
0 I 0 ::: 0 0
been put in evidence several times in the literature [28–30]. However, 6 0
the method does not always give the expected results when the signal 6 0 I ::: 0 0 7 7
6 7
(or flow) studied is noisy and complex [31] or when the spatial R~ ≡ 6
6 ::: ::: ::: ::: ::: ::: 77 (13)
complexity N is smaller than the spectral complexity M. Therefore, Le 6 7
4 0 0 0 ::: I 0 5
Clainche and Vega [32] proposed an extended algorithm, called
higher-order dynamic mode decomposition (HODMD), in order to R^ 1 R^ 2 R^ 3 ::: R^ d−1 R^ d
address such problems. The method, which is similar to the ARMA
method or the standard DMD, is highly effective in terms of Using a modified reduced snapshot matrix, it is possible to write a
computational cost. Additionally, it has been shown to provide more more general higher-order Koopman assumption as
 Article in Advance / LE CLAINCHE ET AL. 3

V~ 2K−d
1  R~ V~ K−d
1 (14) As in the standard DMD, matrix R is obtained by using the
pseudoinverse of matrix T K−d
1 , which is done via plain (no truncation)
where V~ K−d
1 and V~ 2K−d
1 are the submatrices formed with the first SVD as
K − d columns and the last K − d columns of the modified snapshot
matrix, which are defined as T K−d
1  UΛV T (19)
2 3 where the diagonal of Λ contains the singular values, and UUT 
V^ K−d
1
6 1
7 U T U  V T V  the N 0 × N 0 unit matrix. Substituting Eq. (19) into
6 V^ K−d
2 7 Eq. (18) and postmultiplying the resulting equation by VΛ−1 UT
6 2 7
6 7
V~ 1K−d
1 ≡ 6
6 : : : 7
7 (15) yields
6 K−1 7
6 V^ d−1 7 R  T 2K−d
1 VΛ−1 UT (20)
4 5
V^ K
Solving the eigenvalue problem of R provides the eigenvectors qm
d

and the associated eigenvalues μm . The following transformation

This matrix is expected to also exhibit redundancies that are
eliminated by a new dimension reduction via truncated SVD as
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

ζ m
iωm  logμm ∕Δt (21)

V~ 1K−d
1 ≃ U~ Σ~ T~ T ≡ UT
~ K−d
1
1 with T 1K−d
1  Σ~ T~ T (16) provides the frequencies ωm and the damping rates ζm of the DMD
expansion [Eq. (6)].
where U~ T U~  V~ T V~  the N 0 × N 0 unit matrix, and the diagonal of Retaining only the first N components of the vectors q~m  Uq ~ m,
matrix Σ~ contains the singular values σ~ 1 ; : : : ; σ~ N 0 . The number of for which the size is Nd, it is possible to construct the reduced DMD
retained SVD modes N 0 is again calculated as modes u^ m . The DMD modes um of expansion (6) are then calculated
as um  W u^ m . The amplitudes am are obtained via the least-square
σ~ N 0
1 ∕σ~ 1 ≥ ε (17) fitting of the approximation.
Finally, the number of M modes retained in the expansion [Eq. (6)]
where the tunable threshold ε coincides with that in Eq. (9). The step are calculated by applying a second tolerance ε1 set by the user as
is completed via premultiplication of Eq. (14) by U~ T that, by
am ∕a1 ≥ ε1 (22)
invoking Eq. (16), yields
More details can be found in [32].
 K−d
T 2K−d
1  RT (18)
1 It is remarkable that, if d  1 [from Eq. (11)], the algorithm is
similar to a standard DMD algorithm, which is formulated in Eq. (1)
where the new modified N 0 × N 0 Koopman matrix is defined (similar to the ARMA method [22,23], as previously explained in the
~
as R  U~ T R~ U. Introduction [Sec. I] of this paper). Thus, when the standard DMD

3 1.5

1
2

0.5
Amplitude

Amplitude

1
0
0
−0.5

−1
−1

−2 −1.5
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time 4 Time 4
x 10 x 10

0.6 0.3

0.4 0.2

0.2 0.1
Amplitude

Amplitude

0 0

−0.2 −0.1

−0.4 −0.2

−0.6 −0.3

−0.8 −0.4
0 0.5 1 1.5 2 0.5 1 01.5 2
Time
x 10
4 Time x 10
4

Fig. 1 Temporal evolution of the signal captured in four different accelerometers in the real flight-test experiments (the variable of “Time” uses
nondimensional units).
 4 Article in Advance / LE CLAINCHE ET AL.

succeeds, the HODMD provides the same results (for values of d≥1) Eq. (6). Thus, in cases in which the results are very noisy and complex
and, when the standard DMD fails (noisy or transient flows), the (a signal defined by a large number of frequencies), it is possible to
HODMD improves its performance. improve the performance of the HODMD by using the method
The parameter d identified in Eq. (11) is set by the user after some iteratively [31]. Once the HODMD is applied to the original data, it is
calibration, looking for the robustness of the results. This process is reconstructed as in Eq. (6). Then, it is possible to apply the HODMD
very simple but needs to be carried out carefully, especially in the to the reconstructed data, obtaining a new DMD expansion, and so
analysis of data coming from the flight test. The method is applied on. The iterative HODMD method is the one that is used for the
several times using various parameters of d, ε, and ε1 . The calibration present analysis.
process is satisfied when the parameters selected provide similar Because the data analyzed are very noisy and complex, the data
results; in other words, when the frequencies and damping rates presented in the snapshot matrix [Eq. (7)] are transformed using the
calculated are similar in all the test cases carried out [31]. autocorrelation matrix [24]. Then, the HODMD algorithm is applied
Nevertheless, in real data coming from the flight test, the complexity to the part corresponding to the positive lags of the correlation matrix.
is large due to the level of random noise included in the signal. Thus, In this way, the data are cleaner; consequently, the dominant
in the calibration process, it is not possible to look for exactly the frequencies can be calculated in a more accurate way. However, the
same results in the different test cases, but it is necessary to fix some calculated damping rates are modified. The results (frequencies)
references, for example, in the value of the highest amplitude obtained using the autocorrelation matrix are also shown in this
frequencies, as will be seen in the following section. paper, and they are used to estimate the value of the frequency in the
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

After the HODMD calculation has been performed, it is possible to analysis using the raw data. This task facilitates the calibration
reconstruct the original results using the general DMD expansion of process carried out in the HODMD.

0.4

0.3
SNR

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90
Sensor
Fig. 2 Signal-to-noise ratio of the data collected in each one of the accelerometers in the flight test.

100
40

50 20
Amplitude

Amplitude

0 0

−20
−50

−40
−100
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time 4 Time 4
x 10 x 10

40 20

20 10
Amplitude

Amplitude

0 0

−20 −10

−40 −20
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time x 10
4 Time x 10
4

Fig. 3 Counterpart of Fig. 1 but using the autocorrelated data [24].

 Article in Advance / LE CLAINCHE ET AL. 5

III. HODMD to Analyze Flight-Test Data

HODMD has been applied to a set of data coming from real aircraft
flutter flight testing with the aim of calculating the most relevant
frequencies, damping rates, and their associated mode shapes. Three
different flight tests have been carried out using 92 accelerometers,
which represented the spatial complexity (J  N) of the system. The
number of snapshots used for this analysis was K  4301 (of a total
number of 21,508 samples), equispaced in time Δt. On the one hand,
as mentioned before, classical methods failed in the damping/
frequency calculations if N < M: for example, if the signal was too
complex (a large number of frequencies and damping rates; M ≫ N)
or if the number of accelerometers was too low (N ≪ M). Thus,
HODMD seemed a suitable choice to perform this analysis. On the
other hand, a DMD algorithm using a companion matrix (or ARMA)
would make profit of the spatial redundancies of the data (the
companion matrix was based on the linear dependence of the data)
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

[34]; it might capture the most relevant frequency but with smaller
accuracy, and some of the results also might be misunderstood [31].
HODMD cleaned the spectra from noise and captured more than a Fig. 4 Spectrum of frequencies and amplitudes obtained using
single relevant frequency. Figure 1 shows an example of the temporal HODMD with different tolerances (ε and ε1 ) and d in the analysis of
evolution of the signal captured in four different accelerometers. As flight-test data.
seen, the data analyzed considered five different frequency impulses
carried out at different locations of the airplane. As expected, in all the
cases, the frequencies decayed (negative damping rate) but, signal, and that the method is not able to clean. Nevertheless, these
depending on the location of the sensor, the signal amplitude and the results show that the method is able to clean the noise and to capture
damping rate of the frequency varied. This fact complicated the frequencies in datasets completely mixed up with noise. Table 2
analysis of the data, giving high relevance to the calibration process shows the values of the frequency and the damping rate of the highest-
carried out in HODMD in order to look for the robustness of the amplitude mode retained in each one of the vertical lines shown in
results obtained. The signal-to-noise ratio (SNR) has been calculated Fig. 4. The data presented come from test cases 3, 4, and 5 (not using
in each one of the accelerometers, defined as SNR  μ∕σ, where μ the autocorrelated data; see Table 1). As can be seen, the results are
and σ are the data mean and standard deviation. Figure 2 shows the consistent and robust. Similar frequencies and growth rates are
SNR of the signal captured by each one of the accelerometers. The obtained in the three cases. However, there is a slight inconsistency in
number of accelerometers was reduced from 92 to 87 because the test cases T2 and T1 in the calculations of the damping rates related to
data collected in the five remaining sensors contained spurious ω5 and ω6 , respectively. As previously mentioned, these are
information, defined as not a number by the own device. These five lower-amplitude frequencies mixed up with noise. Thus, to calculate
signals have been removed from the remaining analyses. As seen, the these modes (to improve the data cleaning), it will be necessary to
SNR varied from values of 30% (0.3) and 1% (0.01) of noise, use smaller tolerances (ε  ε1 < 5 ⋅ 10−2 ) or larger values of d
depending on the sensor; and the mean SNR was 10% (0.1). Finally, (d > 1000), as in T3, that always provide accurate results.
Fig. 3 shows the evolution of the signal in the four different sensors Nevertheless, this method captures frequencies and damping rates
presented in Fig. 1, but using the autocorrelation. As seen, the level of from the flight test automatically (without need of any previous
noise was reduced in all the cases and the five impulses region were knowledge or assumption, or additional signal manipulation). The
better identified, making it easier to capture the most relevant results show that HODMD perfectly captures four frequencies and
frequencies. damping rates. The two remaining frequencies, ω5 and ω6 , are also
As explained in Sec. II, HODMD is based on the robustness of the accurately calculated; and only in one of the three tests carried out
results; thus, if the method is applied to the same dataset at different
time instants and using different time intervals of d or ε and ε1 , similar
results should be obtained. To illustrate the method, HODMD has Table 2 Frequencies and damping rates
been applied to the output of the 87 accelerometers of the first flight obtained using HODMD in flight-test data with
test of a flutter flight-test campaign using either the original raw data three different tolerances (ε and ε1 ) and da
or the positive lags of the autocorrelation matrix and different test
Test 2fΔt ζ
parameters of d, ε, and ε1 , shown in Table 1.
Figure 4 shows the results obtained for this analysis. As can be ω1 T1 0.0169 0.02
T2 0.0169 0.02
seen, in all cases, it is possible to clearly identify six peaks related to
T3 0.0170 0.03
different frequencies. The amplitudes of four of these peaks are larger ω2 T1 0.0255 0.075
than 0.1 (mean value for the SNR), and they are very well defined in T2 0.0248 0.08
the spectrum (straight vertical line that defines ω1 , ω2 , ω3 , and ω4 ). T3 0.0253 0.06
However, the amplitude of the frequency related to the two remaining ω3 T1 0.0113 0.05
peaks (ω5 and ω6 ) is smaller than 0.1. Because the calculated mean T2 0.0122 0.044
SNR is 0.1, it is possible to conclude that these frequencies are mixed T3 0.0116 0.05
with some spurious modes, which are related to the noise of the ω4 T1 0.0704 0.02
T2 0.0705 0.03
T3 0.0703 0.025
Table 1 Parameters used in the ω5 T1 0.0270 0.08
analysis of flight-test data using HODMD T2 0.0266 0.027
T3 0.0267 0.03
Test Autocorrelation d ε  ε1 ω6 T1 0.0077 0.015
1 Yes 300 5 ⋅ 10−2 T2 0.0077 0.08
2 Yes 500 5 ⋅ 10−3 T3 0.0077 0.017
3 No 500 10−2
5 ⋅ 10−2
a
4 No 1000 The results correspond to the highest-amplitude frequency
5 No 1000 10−3 from each frequency group shown in Fig. 4. T1, T2, and T3
correspond to tests 3, 4, and 5 in the figure.
 6 Article in Advance / LE CLAINCHE ET AL.

(for each case) do the damping rates diverge. Thus, to ensure sensors, located at some specific places, it will be possible to obtain
convergence of this method, it is necessary to compare the results of at the same solution.
least three test cases. The results obtained using HODMD have been compared with the
Figure 5 shows the modes related to the four highest-amplitude least-squares curve-fitting (LSCF) method [24], which is accepted in
frequencies ω1 , ω2 , ω3 , and ω4 . The real and imaginary parts of the the community as suitable for the analysis of flutter instability and is
DMD modes alternate every semiperiod (of the frequency captured) used in the industry (i.e., the second and fourth authors of this paper
in time π. The module of the DMD modes provides an idea of where have used this method in an industrial application). Table 3 compares
the main activity (sensor activity) is focused. The modes with the the frequencies and damping obtained using the HODMD and LSCF
highest values represent the sensor that is relevant for capturing such methods. As seen, both methods agree in the calculations of the
frequencies. Figure 6 shows the DMD modes for which the absolute frequencies ω1 , ω2 , ω3 , and ω5 . The signals consist of control surface
value is larger than one. As can be seen, only a few sensors (∼30 from pulses, which are not the most adequate excitation to accurately
87) are related to these frequencies; thus, with a smaller number of capture damping values. However, both methods provide values of

3 3 4

2 2
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

3
1 1

DMD mode
DMD mode

DMD mode

0 0 2

-1 -1
1
-2 -2

-3 -3 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Sensor Sensor Sensor
4 1.5 3.5

3 1 3
0.5
2 2.5
DMD mode

DMD mode
DMD mode

0
2
1 -0.5
1.5
0 -1
1
-1.5
-1
-2 0.5

-2 -2.5 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Sensor Sensor Sensor
3 3
3
2 2.5
2
1 2
DMD mode

1
DMD mode
DMD mode

0 0 1.5

-1 -1 1

-2 -2 0.5

-3 -3 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Sensor Sensor Sensor
4 1 3.5

3
3
0.5
2.5
DMD mode

DMD mode
DMD mode

2
2
0
1 1.5

1
-0.5
0
0.5

-1 -1 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Sensor Sensor Sensor
Fig. 5 DMD modes in each one of the 87 sensors used in the flight-test analysis: real and imaginary parts and the module of the mode (left to right); and
DMD modes of ω1 , ω2 , ω3 , and ω4 (top to bottom).
 Article in Advance / LE CLAINCHE ET AL. 7

4 4
Mode 1 Mode 1 4
3.5 Mode 2 3.5 Mode 2
Mode 3 Mode 3 3.5
Mode 4 Mode 4
3 3
3
DMD mode

DMD mode

DMD mode
2.5 2.5 2.5

2 2 2
Mode 1
1.5 1.5 Mode 2
1.5
Mode 3
Mode 4
1 1 1
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Sensor Sensor Sensor
Fig. 6 Four highest-amplitude DMD modes with absolute value larger than one for the real and imaginary parts (left and middle), and with absolute
values larger than two for the module (right).
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

Table 3 Frequencies and damping rates IV. Conclusions

obtained using HODMD in the first fight test in
T3 and using the LSCF method [24] The application of the recently developed higher-order dynamic
mode decomposition method to the extraction of aircraft modal
Method 2fΔt ζ characteristics from flutter flight-testing data has been presented in
ω1 HODMD 0.0170 0.03 this paper. The method originates from the standard dynamic mode
LSCF 0.0171 0.09 decomposition (which is similar to the autoregressive moving-
ω2 HODMD 0.0253 0.06 average method) and improves it to be able to deal with higher spatial
LSCF 0.0248 0.06 dimension, lower signal durations and a higher signal-to-noise ratio.
ω3 HODMD 0.0116 0.05 The method has been able to extract, in a robust and efficient
LSCF 0.0116 0.05 manner, the most relevant aircraft modes from test data in an
ω5 HODMD 0.0267 0.03
LSCF 0.0272 0.04
operational sense: that is, without making use of any known excitation.
Additionally, the results have been compared and validated with data
extracted using industry standard methods.
The results are very promising, and this first version of the method
seems very robust. Nevertheless, in future works, the suitability of the
the same order of magnitude, pending confirmation of the damping method needs to be corroborated with the analysis of a higher number
values from additional test information (like frequency sweeps or of test points.
dwells), in which damping trends are key; however, this is out of the
scope of this paper and remains an open topic for future research. The
main advantage of HODMD is that it is an automatic technique (it Acknowledgment
does not require any manual interaction to obtain the results) and
robust. The LSCF method is able to additionally capture other This work was partially supported by the Spanish Ministry of
frequencies, including all the ones captured by HODMD; however, Economy and Competitiveness, under grant TRA2016-75075-R.
the stability diagram does not show sufficient convergence of
the poles.
Following the same methodology, HODMD has been performed References
for the two remaining flight tests. Figure 7 shows the evolution of [1] Schwanz, R. C., and Wells, W., “Identification of Aeroelastic
each one of the three highest-amplitude frequencies and damping Parameters Using a Recursive Sequential Least Squares Method,”
rates as a function of the test case. These results are in good agreement Proceedings of the 6th Atmospheric Flight Mechanics Conference,
with the changes in flight conditions. Increasing the Mach number AIAA Paper 1980-1634, 1980.
implies an increase in the damping rate and the frequency. Although [2] Bennett, R. M., and Abel, I., “Flight Flutter Test and Data Analysis
the data presented are confidential, from the previous comparison of Techniques Applied to a Drone Aircraft,” Journal of Aircraft, Vol. 19,
No. 7, 1982, pp. 589–595.
the HODMD method with the LSCF method, it is possible to
doi:10.2514/3.57433
conclude that the HODMD method is an accurate and robust tool that [3] Nissim, E., and Gilyard, G., “Method for Experimental Determination
could be potentially used to predict the flutter speed, with the of Flutter Speed by Parameter Identification,” Proceedings of the 30th
additional advantage that the method is automatic (does not require Structures, Structural Dynamics and Materials Conference, AIAA
any manual interaction). Paper 1989-1324, 1989.

Fig. 7 Frequency (left) and damping rate (right) variations in each one of the three flight tests performed.
 8 Article in Advance / LE CLAINCHE ET AL.

[4] Cooper, J., “Parameter Estimation Methods for Flight Flutter Testing,” Guidance, Control, and Dynamics, Vol. 29, No. 2, 2006, pp. 374–381.
Proceedings of the 80th AGARD Structures and Materials Panel, doi:10.2514/1.15178
AGARD CP-566, Neuilly-Sur-Seine, France, 1995, pp. 10.1–10.12. [21] Kukreja, S. L., and Brenner, M. J., “Nonlinear Black-Box Modeling of
[5] Brenner, M. J., Lind, R. C., and Voracek, D. F., “Overview of Recent Aeroelastic Systems Using Structure Detection: Application to F/A-18
Flight Flutter Testing Research at NASA Dryden,” NASA TM 4792, Data,” Journal of Guidance, Control, and Dynamics, Vol. 30, No. 2,
April 1997. 2007, pp. 557–564.
[6] Feron, E., Brenner, M., Paduano, J., and Turevskiy, A., “Time- Frequency doi:10.2514/1.20835
Analysis for Transfer Function Estimation and Application to Flutter [22] Onoda, J., “Estimation of Dynamic Characteristics of a Wing from the
Clearance,” Journal of Guidance, Control, and Dynamics, Vol. 21, No. 3, Random Response to Turbulence,” Journal of the Japan Society for
1998, pp. 375–382. Aeronautical and Space Sciences, Vol. 26, No. 299, 1978, pp. 649–656.
doi:10.2514/2.4269 doi:10.2322/jjsass1969.26.649
[7] Turevskiy, A., Feron, E., and Paduano, J., “Flutter Boundary Prediction [23] Matsuzaki, Y., and Ando, Y., “Estimation of Flutter Boundary from
Using Physical Models and Experimental Data,” Journal of Guidance, Random Responses due to Turbulence at Subcritical Speeds,” Journal of
Control, and Dynamics, Vol. 22, No. 1, 1999, pp. 168–171. Aircraft, Vol. 18, No. 10, 1981, pp. 862–868.
doi:10.2514/2.7621 doi:10.2514/3.44737
[8] Pickrel, C. R., and White, P. J., “Flight Flutter Testing of Transport [24] Taylor, P. M., Moreno Ramos, R., Banavara, N., Narisetti, R. K., and
Aircraft: In-Flight Modal Analysis,” Proceedings of the IMAC Morgan, L., “Flight Flutter Testing at Gulfstream Aerospace Using
International Modal Analysis Conference, Kissimmee, FL 2003. Advances Signal Processing Techniques,” Proceedings of the 58th
[9] Li, X., and Brenner, M., “Practical Aeroservoelasticity In-Flight AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Downloaded by UNIVERSITY OF NEW SOUTH WALES on October 15, 2018 | http://arc.aiaa.org | DOI: 10.2514/1.C034863

Identification and Adaptive Control,” Proceedings of AIAA Atmospheric Conference, AIAA Paper 2017-1823, 2017.
Flight Mechanics Conference, AIAA Paper 2010-7502, 2010. [25] McNamara, J. J., and Friedmann, P. P., “Flutter-Boundary Identification
[10] Matsuzaki, Y., and Torii, H., “Flutter-Boundary Prediction of a for Time-Domain Computational Aeroelasticity,” AIAA Journal,
Morphing Wing in the Process of Adaptation,” AIAA Journal, Vol. 50, Vol. 45, No. 7, 2007, pp. 1546–1555.
No. 6, 2012, pp. 1257–1264. doi:10.2514/1.26706
doi:10.2514/1.J051202 [26] Schmid, P., “Dynamic Mode Decomposition of Numerical and
[11] Kehoe, M. W., “A Historical Overview of Flight Flutter Testing,” NASA Experimental Data,” Journal of Fluid Mechanics, Vol. 656, Aug. 2010,
TM 4720, 1995. pp. 5–28.
[12] Ruhlin, C. L., Atson, J., Ricketts, R. H., and Doggett, R. V., “Evaluation doi:10.1017/S0022112010001217
of Four Subcritical Response Methods for On-Line Prediction of Flutter [27] Koopman, B., “Hamiltonian Systems and Transformations in Hilbert
Onset in Wind Tunnel Tests,” Journal of Aircraft, Vol. 20, No. 10, 1983, Space,” Proceedings of the National Academy of Science USA, Vol. 17,
pp. 835–840. 1931, pp. 315–318.
doi:10.2514/3.44951 [28] Le Clainche, S., Li, I. J., Theofilis, V., and Soria, J., “Flow Around a
[13] Dimitriadis, G., and Cooper, J. E., “Flutter Prediction from Flight Hemisphere-Cylinder at High Angle of Attack and Low Reynolds
Flutter Test Data,” Journal of Aircraft, Vol. 38, No. 2, 2001, Number Part I: Experimental and Numerical Investigation,” Aerospace
pp. 355–367. Sciences and Technology, Vol. 44, 2015, pp. 77–87.
doi:10.2514/2.2770 doi:10.1016/j.ast.2014.03.017
[14] Dimitriadis, G., and Cooper, J. E., “Comment on “Flutter Prediction [29] Le Clainche, S., Rodríguez, D., Theofilis, V., and Soria, J., “Flow Around
from Flight Flutter Test Data,” Journal of Aircraft, Vol. 43, No. 3, 2006, a Hemisphere-Cylinder at High Angle of Attack and Low Reynolds
pp. 862–863. Number. Part II: POD and DMD Applied to Reduced Domains,”
doi:10.2514/1.C9463TC Aerospace Sciences and Technology, Vol. 44, 2015, pp. 88–100.
[15] Lind, R., “Comment on ‘Flight-Test Evaluation of Flutter Prediction doi:10.1016/j.ast.2014.10.009
Methods’,” Journal of Aircraft, Vol. 40, No. 5, 2003, pp. 964–970. [30] Gómez, F., Le Clainche, S., Paredes, P., Hermanns, M., and Theofilis, V.,
doi:10.2514/2.6881 “Four Decades of Studying Global Linear Instability: Progress and
[16] Zeng, J., and Kukreja, S. L., “Flutter Prediction for Flight/Wind-Tunnel Challenges,” AIAA Journal, Vol. 50, No. 12, 2012, pp. 2731–2743.
Flutter Test Under Atmospheric Turbulence Excitation,” Journal of doi:10.2514/1.J051527
Aircraft, Vol. 50, No. 6, 2013, pp. 1696–1709. [31] Le Clainche, S., Vega, J. M., and Soria, J., “Higher Order Dynamic
doi:10.2514/1.C031710 Mode Decomposition for Noisy Experimental Data: Flow Structures on
[17] Follador, R., de Souza, C. E., da Silva, R. G. A., and Góes, L. C. S., a Zero-Net-Mass-Flux Jet,” Experimental Thermal and Fluid Science,
“Comparison of In-Flight Measured and Computed Aeroelastic Vol. 88, Nov. 2017, pp. 336–353.
Damping: Modal Identification Procedures and Modeling Approaches,” doi:10.1016/j.expthermflusci.2017.06.011
Journal of Aerospace Technology Management, Vol. 8, No. 2, 2016, [32] Le Clainche, S., and Vega, J. M., “Higher Order Dynamic Mode
pp. 163–177. Decomposition,” SIAM Journal on Applied Dynamical Systems, Vol. 16,
doi:10.5028/jatm.v8i2.558 No. 2, 2017, pp. 882–925.
[18] Hammond, C. E., and Dogget, R. V., Jr., “Determination of Subcritical doi:10.1137/15M1054924
Damping by Moving-Block/Randomdec Applications,” Flutter Testing [33] Sirovich, L., “Turbulence and the Dynamics of Coherent Structures.
Techniques, NASA Scientific and Technical Information Office, Parts I–III,” Quarterly of Applied Mathematics, Vol. 45, No. 3, 1987,
Washington, D.C., 1975, pp. 59–76. pp. 561–571.
[19] Benett, R. G., and Desmarais, R., “Curve-Fitting of Aeroelastic doi:10.1090/qam/1987-45-03
Transient Response Data with Exponential Functions,” Flutter Testing [34] Schmid, P., “Application of the Dynamic Mode Decomposition to
Techniques, NASA SP-415, 1976, pp. 43–58. Experimental Data,” Experiments in Fluids, Vol. 50, No. 4, 2011,
[20] Kukreja, S. L., and Brenner, M. J., “Nonlinear Aeroelastic System pp. 1123–1130.
Identification with Application to Experimental Data,” Journal of doi:10.1007/s00348-010-0911-3