sensors

Article
Flight State Identification of a Self-Sensing Wing via
an Improved Feature Selection Method and Machine
Learning Approaches
Xi Chen 1, *, Fotis Kopsaftopoulos 2 ID
, Qi Wu 3 , He Ren 1 and Fu-Kuo Chang 4
1 Shanghai Engineering Research Center of Civil Aircraft Health Monitoring, Shanghai Aircraft Customer
Service Co., Ltd., Shanghai 200241, China; [email protected]
2 Department of Mechanical, Aerospace, and Nuclear Engineering, Rensselaer Polytechnic Institute,
Troy, NY 12180, USA; [email protected]
3 School of Electronic, Information and Electrical Engineering, Shanghai Jiao Tong University,
Shanghai 200240, China; [email protected]
4 Department of Aeronautics and Astronautics, Stanford University, Stanford, CA 94305, USA;
[email protected]
* Correspondence: [email protected] or [email protected]; Tel.: +86-21-20875861




Received: 31 March 2018; Accepted: 24 April 2018; Published: 29 April 2018


Abstract: In this work, a data-driven approach for identifying the flight state of a self-sensing
wing structure with an embedded multi-functional sensing network is proposed. The flight state is
characterized by the structural vibration signals recorded from a series of wind tunnel experiments
under varying angles of attack and airspeeds. A large feature pool is created by extracting potential
features from the signals covering the time domain, the frequency domain as well as the information
domain. Special emphasis is given to feature selection in which a novel filter method is developed
based on the combination of a modified distance evaluation algorithm and a variance inflation
factor. Machine learning algorithms are then employed to establish the mapping relationship from
the feature space to the practical state space. Results from two case studies demonstrate the high
identification accuracy and the effectiveness of the model complexity reduction via the proposed
method, thus providing new perspectives of self-awareness towards the next generation of intelligent
air vehicles.

Keywords: self-sensing wing; feature extraction; feature selection; flight state identification;
machine learning

1. Introduction
The current state sensing and awareness of flight vehicles relies on traditional sensors and
detection devices mounted on different locations of the vehicle, e.g., Pitot tubes installed in front of the
nose for airspeed measurement, transducers located on each side of the fuselage for angle of attack
detection. Inspired by the unsurpassed flight capabilities of birds, a novel “fly-by-feel” (FBF) concept
has been recently proposed for the development of the next generation of intelligent air vehicles that
can “feel”, “think”, and “react” [1,2]. Such bio-inspired systems will not only be able to sense the
environment (temperature, pressure, aerodynamic forces, etc.), but also be able to think in real-time and
be aware of their current flight state and structural health condition. Further, such systems will react
intelligently under various situations and achieve superior performance and agility. Compared with
the traditional approaches, this FBF concept has the following advantages: (1) structural complexity
reduction by integrated structures with self-sensing ability, (2) structural health on-line monitoring
through embedded multi-functional materials, (3) autonomous flight control and decision-making

Sensors 2018, 18, 1379; doi:10.3390/s18051379 www.mdpi.com/journal/sensors

Sensors 2018, 18, 1379 2 of 21

Sensors 2018, 18, x 2 of 21

based on self-awareness [2]. Towards this end, great challenges have been posed to the current
structural design andbased
decision-making dataonprocessing methods
self-awareness with a departure
[2]. Towards from
this end, great the existing
challenges havetechnologies.
been posed to
the current structural design and data processing methods with a
Recent years have seen the development of different sensing network architectures departure from the existing and
technologies.
simulations [3–6], among which, an expandable network made of polymer-based substrates was
designed Recent years haveand
by the Structure seen the development
Composites Lab (SACL)of different sensing
at Stanford network This
University. architectures
network and contains
simulations [3–6], among which, an expandable network made of polymer-based substrates was
many micro-nodes which have the potential to integrate micro-sensors, actuators and electronics for
designed by the Structure and Composites Lab (SACL) at Stanford University. This network contains
different applications [7]. Based on the development of integration and fabrication techniques [8–10],
many micro-nodes which have the potential to integrate micro-sensors, actuators and electronics for
a smart structure
different with the[7].
applications sensor
Basednetwork monolithically
on the development embedded
of integration andin the layuptechniques
fabrication of a composite
[8–10], UAV
wingawas successfully fabricated [11]. This smart wing consists of four sensor networks
smart structure with the sensor network monolithically embedded in the layup of a composite UAV and each
network
wing was successfully fabricated [11]. This smart wing consists of four sensor networks and each lead
is integrated with strain gauges, resistive temperature detectors (RTD) and piezoelectric
zirconate titanate
network (PZT) transducers.
is integrated Specifically,
with strain gauges, resistive the strain gauge
temperature is used
detectors (RTD)toand
measure the wing
piezoelectric leadstrain
zirconateand
distribution titanate (PZT)any
identify transducers. Specifically,
potentially dangerous the areas.
strain gauge is used to
RTD detects measure
the the wing
temperature strain
distribution
distribution
in order to provide and the
identify any potentially
temperature dangerous[12].
compensation areas.PZTRTDtransducers
detects the temperature
can be useddistribution
for both active
and passive measurements. In the active mode, they can be used for damage detection andactive
in order to provide the temperature compensation [12]. PZT transducers can be used for both structural
and passive measurements. In the active mode, they can be used for damage detection and structural
health monitoring while in passive mode, the wing structural vibration during flying can be captured
health monitoring while in passive mode, the wing structural vibration during flying can be captured
to reflect the air dynamic characteristics [11]. The wing configuration is shown in Figure 1.
to reflect the air dynamic characteristics [11]. The wing configuration is shown in Figure 1.

Figure
Figure 1.1.The
Theself-sensing
self-sensing composite
compositewing
wingdesign [2].[2].
design

After realizing sensing ability through multi-functional structures development, the next step is
After
to equiprealizing
the smartsensing
wing ability throughand
with thinking multi-functional
judging capability,structures
i.e., thedevelopment, the nexttostep
structure is expected be is to
equipaware
the smart wing with thinking
of surroundings and identifyand its
judging
currentcapability,
flying state.i.e.,There
the structure
have been is expected to be aware
studies devoted to of
surroundings
addressingand the identify its current flying
related identification problemstate. There
based on have
eitherbeen
strainstudies devoted
or vibration to addressing
signals obtained the
from
related experiments. problem
identification Huang et al. studied
based the active
on either flutter
strain or control
vibration andsignals
closed-loop flutterfrom
obtained identification
experiments.
and a fast-recursive subspace method was applied in high-dimensional aero-servo-elastic
Huang et al. studied the active flutter control and closed-loop flutter identification and a fast-recursive system.
subspace method was applied in high-dimensional aero-servo-elastic system. The wind flutter
The wind tunnel test showed that the natural frequency and modal damping ratios of the tunnel test
modes can be precisely tracked [13]. Pang and Cesnik employed non-linear least squares fit and
showed that the natural frequency and modal damping ratios of the flutter modes can be precisely
Kalman filtering to obtain wing shape information and rigid body attitude. Results revealed that the
tracked [13]. Pang and Cesnik employed non-linear least squares fit and Kalman filtering to obtain
Kalman filter has good performance in the presence of sensor noise [14]. For elastic deformation,
wingSodja
shapeet information
al. conductedand rigid body
a dynamic attitude.
aeroelastic windResults
tunnelrevealed
experiment thatunder
the Kalman
harmonic filter has good
pitching
performance
excitations, inexperimental
the presence of including
data sensor noise [14]. For
the bending andelastic
torsiondeformation,
deformation were Sodja et al. conducted
consistent with
a dynamic aeroelastic
the elastic wind tunnel
analysis model developed experiment
by the Delft under harmonic
University pitching excitations,
of Technology [15]. For moreexperimental
general
data flight
including
states,the bending andand
Kopsaftopoulos torsion
Chang deformation
established awere consistent
stochastic global with the elastic
identification analysis
method usingmodel
PZT signals from both time and frequency domain based on developed
developed by the Delft University of Technology [15]. For more general flight states, Kopsaftopoulos Vector-dependent
Functionally
and Chang Pooleda(VFP)
established model
stochastic [2,16,17].
global A large range
identification methodof airspeeds
using PZT andsignals
angles from
of attack
bothwere
time and
considered in the VFP-based identification framework and the structural
frequency domain based on developed Vector-dependent Functionally Pooled (VFP) model [2,16,17]. dynamics of the composite
wing could be captured and predicted.
A large range of airspeeds and angles of attack were considered in the VFP-based identification
Overall, the above data processing approaches mainly belong to state space methods and
framework and the structural dynamics of the composite wing could be captured and predicted.
improved time series analysis. Based on the previous study yet from another perspective, if we can
Overall, the above data processing approaches mainly belong to state space methods and
improved time series analysis. Based on the previous study yet from another perspective, if we can
Sensors 2018, 18, 1379 3 of 21

extract distinguished features from the continuous coupled structural aerodynamic behavior, it is
possible to identify the flight state directly using the limited features instead of detailed characterization
of the structural responses. Machine learning techniques can be employed to establish the mapping
relationship from the feature space to the practical state space.
Facing a series of signals generated from the embedded sensor network, one of the main challenges
is what kind of features should be extracted and whether these features are useful for classification.
A set of features without careful selection and evaluation may lead to poor results whatever superior
machine learning models are applied. Feature engineering is such a research field including feature
extraction and selection. For a period of time series signals with noise, various statistical features
can be calculated such as the mean value, standard deviation, peak value, kurtosis, etc. from both
time domain and frequency domain [18], a feature pool is then created with different number of
features depending on the characteristics of the signals [19–21]. More features are encouraged to
avoid missing important candidates with superior classification performance. The next step is feature
selection in which a limited subset is obtained by eliminating less effective features. It reduces
model dimension and computational time [22]. Generally, feature selection can be divided into three
categories as filter, wrapper and embedded. Filter methods rank the variables completely separate
to the model used for classification. The assignment of feature importance is based on information
generated by some statistical algorithms. Filter methods are computationally simple and fast without
the interaction with the classifier and feature dependencies [23]. Embedded solutions select salient
features as part of the learning process of the model, which can be linear regression, support vector
machine, decision tree, random forest, etc. These methods integrate the subset selection into the model
construction but are difficult to adjust for the optimal search [24]. The third category is wrapper,
in which features are selected based on the performance of a given model by searching the possible
subsets space and assessing the performance of the given model on each subset, models can be various
learning machines [25]. Although wrapper methods often achieve sound classification performance by
considering the feature dependencies, the frequent interactions between feature subset search and the
classifier cause high computational costs [26].
We have demonstrated the effectiveness of establishing the mapping relationship from the feature
space to the flight state space through neural networks modelling [27]. This paper significantly
improves the previous work by creating a much larger feature pool and considering the co-linearity
among various features. To sum up, the objective of this paper is the introduction and evaluation
of a novel feature selection method for accurate flight state identification of a self-sensing wing
structure based on experimental vibration data recorded by piezoelectric sensors under multiple flight
states. The developed method belongs to the filter family and is capable of obtaining a group of
most important features for classification with low mutual dependency. The framework of the data
acquisition, methodology development, evaluation and application is shown in Figure 2.
The rest of the paper is organized as follows: Section 2 presents the problem statement. Section 3
focuses on the feature extraction and feature selection in which the novel filter algorithm is introduced.
Two case studies including the general flight state identification and the stall detection and alerting are
conducted in Section 4 followed by their results and discussions in Section 5. Concluding marks are
made in the last section.
Sensors 2018, 18, x 3 of 21

extract distinguished features from the continuous coupled structural aerodynamic behavior, it is
possible to identify the flight state directly using the limited features instead of detailed
characterization of the structural responses. Machine learning techniques can be employed
Sensors 2018, 18, 1379 4 of 21 to
establish the mapping relationship from the feature space to the practical state space.

Data Acquisition

Self-sensing wing PZT signal preparation

Methodology Development

Feature pool Modified distance evaluation

1
 ue , j  u k , j
K
d (j b ) 
K ( K  1) k , e 1
Time domain
1
  k 1 d k , j
K
d (j w)
In dom

K
ain y
dom uenc
fo
rm ain
at

d (j b ) Variance inflation factor

q
io

j j
F re
n

1
d (j w) VIFj 
1  R 2j

Feature extraction Feature selection

Evaluation: Flight State Identification

3D visulizaiton Correlation Analysis Accuracy comparison

Application: Stall Alerting

Stall phenomenon Accuracy comparison

Figure 2. Framework of the proposed methodology.

Figure 2. Framework of the proposed methodology.

Facing a series of signals generated from the embedded sensor network, one of the main
2. Problem Statement
challenges is what kind of features should be extracted and whether these features are useful for
The problem
classification. A set statement
of featuresofwithout
this work is as follows:
careful selection based on signals may
and evaluation collected
lead from the results
to poor PZT
sensors embedded in the self-sensing wing through a series of experiments under varying
whatever superior machine learning models are applied. Feature engineering is such a research field flight
states, develop
including feature aextraction
feature selection method For
and selection. that aisperiod
capableofoftime
obtaining
series limited
signals useful features
with noise, for
various
flight state identification with high accuracy and low model complexity. Specifically, the coupled
Sensors 2018, 18, 1379 5 of 21

aerodynamic-mechanical responses represent different flight states, with each state characterized by
a specific angle of attack (AoA) and airspeed and kept constant during the data collection. The first
problem is that whether a few salient features can be extracted from a period of vibrational time series
(e.g., thousands of data points) as a representation of the corresponding flight state. In this way, we can
skip the investigation into the detailed aeroelastic behavior and use the limited features to identify
the specific flight state directly instead of using the entire lengthy signal. This would significantly
reduce the complexity of the flight state characterization. The second problem is how to guarantee the
effectiveness of selected features. If the selected strong features are highly correlated with each other,
they will exhibit similar identification ability which are still away from the optimal subset.
The above two problems constitute the motivation of this study and are addressed in the following
approaches: firstly, a large number of features is extracted to cover a wide range of descriptions of
the flight state. Then, a modified distance evaluation algorithm is conducted to obtain a subset of
individually powerful features followed by the combination of a variance inflation factor algorithm
to reduce high dependency among features in the subset. Machine learning models are employed to
evaluate the above method for multiple flight states identification as well as a specific case of stall
detection and alerting.
The main novel aspects of this study include:

(1) A large feature pool is created covering up to 47 different features from the time, frequency and
information domains.
(2) A novel filter feature selection method is developed by combining a modified distance evaluation
algorithm and a variance inflation factor.
(3) The flight state identification is treated as a classification problem by establishing the mapping
relationship from the feature space to the physical space characterized by varying angle of attack
and airspeed of the self-sensing wing structure in wind tunnel experiments.
(4) The application on stall detection and alerting with high identification accuracy provides new
perspectives for autonomous flight control with real-time flight state monitoring.

3. Methodology Development
In this section, a novel filter feature selection method is proposed via the combination of
a modified distance evaluation algorithm and a variance inflation factor. In order to obtain sufficient
feature candidates, a large feature pool is firstly created by extracting features covering a wide range.
The output of this method is a feature subset consisting of most salient features with low correlation,
which is able to represent a lengthy time-series signal of the wing structural response under certain
flight state.

3.1. Feature Extraction

Feature extraction relies heavily on experts’ knowledge, it is encouraged to extract different kinds
of features, as many as possible in case of missing useful ones. In this study, we intend to create a large
feature pool from three main sources, namely the time, frequency and information domains.
In time domain, 25 statistical features are calculated including 12 commonly used features such
as mean, standard deviation, variance, peak, mean absolute deviation, etc. and 13 un-dimensional
features such as crest factor, shape factor and a series of normalized central moments. The expressions
of all time domain features are listed in Table 1. In terms of their physical insights, t1 –t12 may reflect
the vibration amplitude and energy while t13 –t25 may represent the series distribution of the signal in
time domain.
Previous studies employed Fast Fourier Transform (FFT) to convert the time series into frequency
spectrum [19,20]. However, the signal instances from the wind tunnel experiments are samples of
a stochastic process with considerable noise. Welch’s method improves FFT by shortening the signals
and averaging, and thus the peaks are smoothed for noise reduction [28]. Herein, a sample-long
Sensors 2018, 18, 1379 6 of 21

Hamming data window with 90% overlap is used for the Welch-based spectral estimation. A series of
power spectrum y(k) without log transformation is then used for frequency domain feature extraction.
Thirteen statistical features such as mean spectrum, spectrum center, root mean square spectrum, etc.
and their mathematical expressions are shown in Table 2. f 1 may indicate the vibration energy in the
frequency domain. f 2–4 , f 6 , f 10–13 may describe the convergence of the spectrum power. f 5 , f 7–9 may
show the position change of the main frequency.

Table 1. Features in time domain.

Time Domain Feature Parameters Un-Dimensional

 N √ 2
3
N ∑n=1 | x (n)| t9 N
∑n=1 ( x (n)−t1 )
t1 = ∑n=1N
x (n) t13 =
t7 = N t6 t19 = N · t2 3
4
q N N
N
∑n=1 ( x (n)−t1 )
2
∑n=1 | x (n)| t14 = t6 ∑n=1 ( x (n)−t1 )
t2 = t8 = N t8 t20 = N · t2 4
N
N 3 t9
∑n=1 ( x (n)−t1 ) t9 = max( x (n)) t15 =
t3 = N t8
N 4 t9
∑n=1 ( x (n)−t1 ) t10 = min( x (n)) t16 = ...
t4 = N t7
N 2 t3
∑n=1 ( x (n)−t1 ) t11 = t9 − t10 t17 =
t5 = N t6 3
9
q N N
N
∑n=1 ( x (n))
2
∑n=1 | x (n)−t1 | t18 = t4 ∑n=1 ( x (n)−t1 )
t6 = t12 = N t6 4 t25 = N · t2 9
N
Note: x(n) is a signal series for n = 1, 2, . . . , N, N is the number of data points.

Table 2. Features in the frequency domain.

Frequency Domain Feature Parameters

q
K
( f r − f )2 y ( k ) f6
f 1 = ∑k=N
K
1 y(k)
f 6 = ∑k=1 kK 5 f 10 = f5
r
K K 3
f 7 = ∑k=K1 k ∑ k =1 ( f r k − f 5 ) y ( k )
K 2 f r 2 y(k)
∑k=1 (y(k)− f 1 ) f 11 =
f2 = K ∑ k =1 y ( k ) K· f6 3

(y(k)− f )3
K
r
f 3 = ∑ k =1 √ 3 1
K K 4
f 8 = ∑kK=1 k 2 ∑ k =1 ( f r k − f 5 ) y ( k )
f r 4 y(k)
f 12 = K· f6 4
K( f2 ) ∑ k =1 f r k y ( k )
K K
√
K
∑k=1 (y(k)− f 1 )
4
∑ k =1 f r k 2 y ( k ) ∑ k =1 | f rk − f 5 |y(k )
f4 = f9 = √ K f 13 = √
K· f2 2 ∑ k =1y(k)∑kK=1 f rk 4 y(k ) K f6
K
∑k=1 ( f rk ·y(k ))
f5 = K
∑ k =1 y ( k )
Note: y(k) is a spectrum for k = 1, 2, . . . , K, K is the number of spectrum components; f rk is the frequency value of
the kth spectrum line.

In electroencephalograph (EEG) analysis for neural diseases diagnosis and vibration analysis for
mechanical defects, fractal dimensions from computational geometry and entropies from information
theory have demonstrated effectiveness in early diseases/fault diagnosis [29,30]. Inspired by that,
a group of complex features are employed and their terminologies are Multi-Scale Entropy, Partial Mean
of Multi-Scale Entropy, Petrosian Fractal Dimension, Higuchi Fractal Dimension, Fisher Information,
Approximate Entropy, and Hurst Exponent, respectively.
Multi-Scale Entropy (MSE) introduces the scale factor based on the sample entropy to measure
the complexity of signal under different scale factors [31]. It is calculated as:

MSE = {τ SampEn(τ, m, r ) = − ln[Cr,m+1 (r )/Cr,m (r ) ]} (1)

where τ is the scale factor, m is the embedding dimension and r is the threshold. Here m = 2,
r = 0.2 * standard deviation, τ = 12.
Sensors 2018, 18, 1379 7 of 21

The first three values are selected due to the relatively high distinction among different classes.
Also, an integrated non-linear index called Partial Mean of Multi-Scale Entropy (PMMSE) is used to
simultaneously reflect the mean value and variation trend of MSE [32], which is expressed as:

PMMSE = (1+|Ske|/3) · MSEa (2)

where Ske = 3( MSEa − MSEb )/MSEc , MSEa , MSEb , MSEc represent mean, median and standard
deviation of MSE(τ ) = [ MSE(1), MSE(2), . . . , MSE(12)].
Fractal dimension characterizes the space filling capacity of a pattern that changes with the scale
at which it is measured [33]. Herein, two approaches are used as Petrosian Fractal Dimension (PFD)
and Higuchi Fractal Dimension (HFD). PFD is calculated as:

log10 N
PFD = (3)
log10 N + log10 ( N/( N + 0.4Nδ ))

where N is the length of the signal and Nδ is the number of sign changes in the signal derivative [30].
In terms of HFD, firstly k new series are constructed from the original signal [ x1 , x2 , . . . , x N ] by
[ xm , xm+k , xm+2k , . . . , xm+b( N −m)/kck ], where m = 1, 2, . . . , k. Secondly the length L(m, k) for each new
series is calculated as:
b( N −m)/kc
∑ i =2 xm+ik − xm+(i−1)k ( N − 1)
L(m, k ) = (4)
b( N − m)/kck
j k
and the average length L(k) = ∑ik=1 L(i, k) /k. After kmax repetitions, a least-squares method is used
to obtain the best slope that fits the curve of ln( L(k)) versus ln(1/k), which is defined as the Higuchi
Fractal Dimension. For details, please refer to [34].
Fisher Information (FI) measures the expected value of the observed information [35]. Its
mathematical expression using normalized singular spectrum is:

M −1 ( σ i +1 − σ i )2
FI = ∑ i =1 σi
(5)

where σi is the normalized value through σi = σi /∑ jM=1 σj , and M is the number of singular value.
Approximate Entropy (ApEn) quantifies the amount of regularity and the unpredictability of
fluctuations of a signal [36], which is computed in the following procedures:

(1) Set the input as [ x1 , x2 , . . . , x N ].

(2) Construct the subsequence x (i, m) = [ xi , xi+1 , . . . , xi+m−1 ] for 1 ≤ i ≤ N − m, where m is the
subsequence length.
(3) Construct a set of subsequences { x ( j, m)} = { x ( j, m)| j ∈ [1, . . . N − m]}, where x ( j, m) is defined
in Step (2).
(
−m
∑N
j =1 k j 1 if | x (i, m) − x ( j, m)|< r
(4) For each x (i, m) ∈ { x ( j, m)}, C (i, m) = N −m , where k j = .
0 otherwise
(5) ApEn is calculated as:
" #
N −m
1 C (i, m)
ApEn(m, r, N ) =
N−M ∑ ln
C (i, m + 1)
(6)
i =1

Hurst Exponent (HST) measures the long-term memory of a signal. It is used to quantify the relative
tendency of the signal either to regress to the mean or to cluster in a direction [37]. For time series
Sensors 2018, 18, 1379 8 of 21

X = [ x1 , x2 , . . . , x N ], its accumulated deviation within range T is calculated as X (t, T ) = ∑it=1 ( xi − x),

where x = T1 ∑iT=1 xi , t ∈ [1, 2, . . . , N ]. Then:

R( T ) max( X (t, T )) − min( X (t, T ))

= q (7)
S( T )
(1/T )∑tT=1 [ x (t) − x ]2

The slope of ln( R(n)/S(n)) versus ln(n) for n ∈ [2, 3, . . . , N ] is defined as the Hurst Exponent.
In summary, abbreviations of the complex features extracted from information domain are listed
in Table 3.

Table 3. Features in information domain.

Information Domain Feature Parameters

I1 = MSE [1] I4 = PMMSE I7 = FI
I2 = MSE [2] I5 = PFD I8 = ApEn
I3 = MSE [3] I6 = HFD I9 = HST

3.2. Feature Selection

Feature extraction guarantees a wide coverage of the object descriptions from various aspects
while feature selection ensures that a set of most salient descriptions can be utilized. For large-scale
models, feature selection is of utter importance in computation reduction and efficiency improvement.
The distance evaluation technique ranks the feature importance independent of the model used
for classification, which belongs to the filter category as mentioned in the Introduction. Salient
features result in minimum inner-class distances of the same class while have maximum margins for
different classes. It has been widely used in fault diagnosis of rotating machinery [20,21,38]. Suppose

a feature set has K conditions, qi,k,j , i = 1, 2, . . . , Ik ; k = 1, 2, . . . , K; j = 1, 2, . . . , J , where qi,k,j is the
jth eigenvalue of the ith sample under the kth condition, Ik is the sample number of the kth condition,
and J is the feature number of each sample. Totally Ik × K × J features are obtained in the feature set

qi,k,j . Herein, a modified distance evaluation algorithm is presented as follows:

(1) Calculate the average distance of the same condition samples:

1 Ik
Ik × ( Ik − 1) ∑l,i=1 i,k,j
dk,j = q − ql,k,j , l, i = 1, 2, . . . , Ik , l 6= i (8)

then obtain the average distance of K conditions:

1 K
K ∑k=1 k,j
(w)
dj = d (9)

(2) Calculate the average eigenvalue of all samples under the same condition:

1 Ik
Ik ∑i=1 i,k,j
uk,j = q (10)

then obtain the average distance between condition samples:

1 K
∑
(b)
dj = k,e
u − uk,j , k, e = 1, 2, . . . , K, k 6= e
=1 e,j
(11)
K ( K − 1)
Sensors 2018, 18, 1379 9 of 21

(b)
(3) Calculate the variance factor of d j as:

(b)
sum( ue,j − uk,j )
vj = (12)
min( ue,j − uk,j )

(4) Calculate the compensation factor as:

sum(vbj )
δj = (b)
(13)
vj

(b) (w)
(5) Calculate the ratio d j and d j considering the compensation factor:

(b)
dj
α j = δj (w)
(14)
dj

then normalize α j and obtaining the feature importance criteria:

αj
αj = (15)
sum(α j )

A higher α j indicates that the corresponding feature j has greater importance. Features can be
ranked in terms of the α j values in Equation (15) in descending order. This algorithm is referred
to as Modified Distance Evaluation algorithm (MDE). Although the top ranked features have
superior discriminative capability, they may suffer from high multi-collinearity, which refers to the
non-independence among features [39]. Herein, the variance inflation factor (VIF) is used to avoid
high collinearity. Assuming a training sample set X with J features X1 , X2 , . . . , X J and class Y, the VIF
of feature j is calculated as:
1
VIF j = (16)
1 − R2j

where R2j is the R-squared value of the regression equation X j = β 0 + βX 0 , in which X 0 contains all
features except X j . An improved algorithm combining MDE and VIF is presented in Algorithm 1 and
is abbreviated as MDV (Modified Distance evaluation and variance inflation Factor).

Algorithm 1: MDV Algorithm.

(1) Set the selected future subset Fsub = ∅, j = 1;

(2) Rank the J features in terms of the α j defined in Equation (15) in descending order. Set Fr to represent the
index list of the ranked features. Add the first feature in Fr to Fsub , j = j + 1;
(3) while j < J :
calculate the VIFj of the jth feature in Fr with the features in Fsub ;
if VIFj < 10:
add the jth feature in Fr to Fsub ;
end
j = j + 1;
end

The MDV algorithm describes the feature-subset selection for multi-class classification based on
the filter method with the MDE and VIF. The threshold of 10 in MDV is an empirical value. A larger
Sensors 2018, 18, 1379 10 of 21

threshold will result in a higher correlation of the selected feature in Fr with the existing features in
Fsub [23].

4. Case Study

4.1. Data Prepraration

A series of wind tunnel experiments of the self-sensing composite wing were conducted under
various angles of attack (AoAs) and freestream velocities at Stanford University. The open-loop wind
tunnel with a square test section of 0.76 m by 76 m was used and a basis was designed to supported
the composite wing allowing adjustments in the angle of attack (AoA). The composite wing dimension
is outlined in Table 4.

Table 4. Wing Dimension.

Wing Geometry
Chord 0.235 m
Span 0.86 m
Area 0.2 m2
Aspect ratio 3.66

Compared with the size of the wind tunnel test section, the additional 0.1 m extension of the
wing span was attached to the wing fixture. The AoAs range from 0 degree up to 18 degrees with an
incremental step of 1 degree. At each degree, data were collected for all velocities ranging from 9 to
22 m/s (incremental step of 1 m/s). For experimental details, please refer to [2].
PZT signals reflect the coupled airflow-structural dynamics through the wing structural vibration
and each time series contains coupled behavior with repeated patterns of a certain flight state. This
study focuses on the usage of PZT sensor signals for flight state identification. In each experiment,
the structural vibration responses (60,000 data points) were recorded from the PZT located near the
wing root at 1000 Hz sampling frequency. For each flight state, data are prepared in two steps: (1) the
entire signal of 60,000 data points is divided into 60 segments (1000 data points for each segment) to
ensure enough samples for training while each segment has sufficient data points for feature extraction;
(2) the first order difference and zero-mean are conducted for each sample sequence in order to
eliminate the influence of zero drift. To evaluate the effectiveness of the proposed method and apply it
for dangerous state pre-warning, two sets of data are collected for general flight state identification
and stall detection and alerting.

4.2. General Flight State Identification

The first data set includes PZT signals with a coarse resolution covering the range of 16 flight
states corresponding to combinations of four AoAs (1, 5, 9, 13 degrees) and four airspeeds (10, 13, 16,
19 m/s). Four signal segments are shown in Figure 3 under a series of AoAs and a fixed airspeed of
10 m/s as an example.
It is noticed that the flight state with AoA of 13 degrees and velocity of 10 m/s can be obviously
identified since the amplitude of the voltage distinguishes it from other signals (it is because this
flight state is close to the stall condition which will be discussed later). The second largest amplitude
comes with 9 degrees which can be separated to a certain extent but already has overlaps with the
rest two. In the study, the identification of the different flight states relies on the features selected by
the developed method in Section 3. To compare the feature selection effectiveness, four other feature
selection methods are employed including Univariate Feature Selection based on mutual information
(UFS_m), Support Vector Machine with L1 regularization (SVM_L1), Gradient Boosted Decision Tree
(GBDT) and Stability selection (STAB). These methods cover three main feature selection categories.
A brief introduction is presented as follows:
Sensors 2018, 18, 1379 11 of 21

SensorsUFS_m
(1) 2018, 18, is
x a commonly used filter method. It performs test on each feature by evaluating the
10 of 21
relationship between the feature and the response variable based on mutual information [40],
which is defined as
p( x, y)
Compared with the size of I ( X, = ∑tunnel
theY )wind
y ∈Y ∑ x ∈ X
test psection,
( x, y) logthe
( additional ) 0.1 m extension of (17) the
wing span was attached to the wing fixture. The AoAs range from (0xdegree p ) p(y) up to 18 degrees with an
incremental step of 1 degree. At each degree, data were collected for all velocities ranging from 9 to
It measures the mutual dependence between variable X and Y. Features with low rankings
22 m/s (incremental step of 1 m/s). For experimental details, please refer to [2].
are removed.
PZT signals reflect the coupled airflow-structural dynamics through the wing structural
(2) SVM_L1 is one of the embedded methods, which selects salient features as part of the learning
vibration and each time series contains coupled behavior with repeated patterns of a certain flight
system [18]. Support Vector Machine (SVM) is a popular machine learning method based on
state. This study focuses on the usage of PZT sensor signals for flight state identification. In each
structural risk minimization principle. It constructs a hyperplane that has the largest distance to
experiment, the structural vibration responses (60,000 data points) were recorded from the PZT
the nearest training data points, which are so called support vectors. An appropriate separation
located near the wing root at 1000 Hz sampling frequency. For each flight state, data are prepared in
can reduce the generalization error of the classifier [41]. L1 is a regularization item added to the
two steps: (1) the entire signal of 60,000 data points is divided into 60 segments (1000 data points for
loss function as |W|, where W standards for the parameter matrix of the learning model [42].
each segment) to ensure enough samples for training while each segment has sufficient data points
This is a penalty item to make the model sparse with fewer useful input dimensions.
for feature extraction; (2) the first order difference and zero-mean are conducted for each sample
(3) GBDT
sequence is a tree-based
in order model
to eliminate belonging
the influence ofto thedrift.
zero embedded category.
To evaluate It combinesofweak
the effectiveness decision
the proposed
method trees in apply
and an iterative manner based
it for dangerous onpre-warning,
state gradient descent
two through additive
sets of data training.for
are collected Trees are added
general flight
at each iteration with modified parameters
state identification and stall detection and alerting. learned in the direction of residual loss reduction [43].
(4) Stability selection is a kind of wrapper method, in which features are selected based on the
established
4.2. General Flightmodels using different subsets, model could be of various types and structures such
State Identification
as logistic regression, SVM, etc. By calculating the frequency of a feature ended up being selected
The first data set includes PZT signals with a coarse resolution covering the range of 16 flight
as important from a feature subset being tested, powerful features are expected to have high
states corresponding to combinations of four AoAs (1, 5, 9, 13 degrees) and four airspeeds (10, 13, 16,
scores close to 100%, weaker features will have lower score and the least useful ones will close to
19 m/s). Four signal segments are shown in Figure 3 under a series of AoAs and a fixed airspeed of
zero [44]. Herein, a randomized logistic regression is used as the selection model.
10 m/s as an example.

Figure 3. Indicative signals under a set of AoAs and a constant velocity of 10 m/s.
Figure 3. Indicative signals under a set of AoAs and a constant velocity of 10 m/s.

It is noticed that the flight state with AoA of 13 degrees and velocity of 10 m/s can be obviously
4.3. Application to Stall Detection and Identification
identified since the amplitude of the voltage distinguishes it from other signals (it is because this
flightThe second
state data
is close to set
the covers a higher which
stall condition resolution
will of
beflight stateslater).
discussed (AoAs: 11,second
The 12, 13 degrees, airspeeds:
largest amplitude
10,
comes13, 16,
with19 m/s) for critical
9 degrees whichstates
can be alerting. In aerodynamics,
separated stall phenomenon
to a certain extent but already ishas
oneoverlaps
of the dangerous
with the
conditions wherein a sudden reduction of the lift coefficient occurs as the angle of attack
rest two. In the study, the identification of the different flight states relies on the features selected by increases
beyond a critical
the developed point. According
method in Section 3.to To
previous
compare analysis [2], theselection
the feature signal energy can be used
effectiveness, fourasother
an indicator
feature
of the liftmethods
selection loss of the
areself-sensing wing. From
employed including the wind
Univariate tunnelSelection
Feature experiments,
based the mean values
on mutual of the
information
signal
(UFS_m), energy for aVector
Support series of AoAs (from
Machine 0 toregularization
with L1 17 degrees) under four airspeeds
(SVM_L1), Gradient (10, 13, 16,
Boosted 19 m/s)Tree
Decision are
obtained
(GBDT) and andStability
shown in Figure 4.
selection (STAB). These methods cover three main feature selection categories.
A briefTheintroduction
signal energy variation as
is presented with respect to the angle of attack is similar under four different
follows:
airspeeds. It is noticed that for relatively low velocities (10 m/s, 13 m/s & 16 m/s), the significant
(1) UFS_m is a commonly used filter method. It performs test on each feature by evaluating the
increase occurs approximately after 14 degrees while for the relatively high speed (19 m/s), stall
relationship between the feature and the response variable based on mutual information [40],
happens much early at 13 degrees. It should be noted that the data were stopped recording after
which is defined as
p( x, y)
I ( X , Y )   yY  xX p( x, y) log( ) (17)
p ( x) p ( y )
 as logistic regression, SVM, etc. By calculating the frequency of a feature ended up being selected
as important from a feature subset being tested, powerful features are expected to have high
scores close to 100%, weaker features will have lower score and the least useful ones will close
to zero [44]. Herein, a randomized logistic regression is used as the selection model.
Sensors 2018, 18, 1379 12 of 21
4.3. Application to Stall Detection and Identification
The second
13 degrees with thedata
highset covers
speed of 19a m/s,
higher resolution
which of flight
is reflected in thestates (AoAs:
red line 11, 12,
with zero 13 degrees,
energy starting
airspeeds:
from 10, 13, Therefore,
14 degrees. 16, 19 m/s)wefordefine
criticalthe
states
orangealerting.
shaded In aerodynamics,
area starting from stall 13
phenomenon is one
degrees as the of
stall
the dangerous
region conditions
which should whereinMoreover,
be avoided. a suddenitreduction
is observed of the
thatlift coefficient
at 12 degrees,occurs as the
the signal anglefor
energy of
attackflight
some increases
statesbeyond a critical
has certain point.
increase According
compared with to the
previous analysis
rest small [2],This
angles. the signal
degreeenergy can as
is defined be
used
the as an
alert indicator
region as theoftransition
the lift loss of the the
between self-sensing
safe region wing. Frominthe
marked wind
light tunnel
green andexperiments, the
the critical stall
mean values
region. Whenof theself-sensing
the signal energywingforcomes
a seriestoofthis
AoAs (from
region, 0 to 17 degrees)
warnings should beunder four airspeeds
provided (10,
to the flight
13, 16, 19
control form/s)
angleare obtained and shown in Figure 4.
reduction.

Figure4.4. Signal
Figure Signalenergy
energyunder
undervarious
variousflight
flightstates.
states.

The signal
5. Results energy variation with respect to the angle of attack is similar under four different
and Discussion
airspeeds. It is noticed that for relatively low velocities (10 m/s, 13 m/s & 16 m/s), the significant
5.1. General
increase Flightapproximately
occurs State Identification
after 14 degrees while for the relatively high speed (19 m/s), stall
happens muchdata
The first early
setatwith
13 degrees. It should
a relatively low be noted that
resolution of the
16 data
flightwere stopped
states is usedrecording afterthe
to evaluate 13
degrees with the high speed of 19 m/s, which is reflected in the red line with zero
performance of six feature selection methods, which include Univariate Feature Selection based onenergy starting
mutual information (UFS_m), Support Vector Machine with L1 regularization (SVM_L1), Gradient
Boosted Decision Tree (GBDT) and Stability selection (STAB), Modified Distance Evaluation (MDE),
and our proposed filter method Modified Distance Evaluation with Variance Inflation Factor (MDV).
Feature rankings are obtained and the top 10 features for different methods are listed in Table 5 and
their detailed expressions are listed in Appendix A.

Table 5. Top 10 ranking matrix.

Ranking UFS_m SVM_L1 GBDT STAB MDE MDV

1 F25 F41 F47 F47 F35 F35
2 F34 F43 F40 F12 F26 F30
3 F6 F39 F46 F21 F2 F5
4 F2 F25 F14 F20 F6 F28
5 F5 F46 F39 F19 F31 F42
6 F4 F19 F44 F18 F30 F45
7 F40 F33 F41 F17 F12 F41
8 F23 F13 F1 F16 F8 F46
9 F42 F44 F21 F15 F36 F14
10 F17 F10 F45 F14 F10 F23
 Sensors 2018, 18, 1379 13 of 21

It is observed from the table that the ranking results vary with the different methods. An intuitive
evaluation is to simply visualize the features distribution under various flight states. For example, four
features are plotted in Figure 5 including: F1 (mean value in time domain), F29 (spectrum kurtosis in
frequency domain), F35 (spectrum power convergence in frequency domain), and F47 (Hurst Exponent
in information domain). The x axis denotes the 16 flight states while the y axis is the feature value before
normalization. The shaded area along each vertical line segment represents the feature distribution
in a single flight state and each subplot of Figure 5 describes a feature distribution on 16 flight
states. As mentioned in Section 3, F1 (mean value) has no effects in classification. Correspondingly,
F1 has the highest overlap among flight states. Similarly, F47 has large overlaps which exhibits pool
classification capability. Theoretically, the ranking of F1 and F47 should be low but they are ranked
high in GBDT and STAB. In comparison, F30 and F35 show smaller overlap and thus have better
classification performance. This may provide some physical insights of the effectiveness of different
Sensors 2018, 18, x 13 of 21
feature selection methods.

Figure
Figure5.5.Pool
Pooland
and superior featuresagainst
superior features against1616 flight
flight states.
states.

TheThe
lastlast
column
columnMDVMDVin in
Table 4 is
Table anan
4 is improvement
improvementof ofMDE
MDE forfor preventing highcollinearity.
preventing high collinearity. To
examine the effects of the proposed algorithm, Correlation analysis is conducted for MDV and MDE
To examine the effects of the proposed algorithm, Correlation analysis is conducted for MDV and MDE
as shown in Figure
as shown 6. 6.
in Figure
It is obvious that the top 10 features selected by MDE are highly correlated with each other.
In comparison, the overall collinearity of the features in MDV is much lower except for the small
region of the top three.
To visualize the feature selection performance by MDV, t-Distributed Stochastic Neighbor
Embedding (t-SNE) is employed which is a relatively new method of dimension reduction particularly
suitable for non-linear and high-dimensional datasets. It is a kind of manifold learning technique by
mapping to probability distributions through affine transformation. For detailed algorithm, please
refer to [45]. The 3D visualization by t-SNE is shown in Figure 7. The left figure is the visualization
using the entire feature pool while the right figure uses only top six features obtained by MDV. It can
 Sensors 2018, 18, 1379 Figure 5. Pool and superior features against 16 flight states. 14 of 21

The last column MDV in Table 4 is an improvement of MDE for preventing high collinearity. To
examine the the
be seen that effects of the
feature proposed
subset algorithm,
through Correlation
MDV selection analysis
exhibits better is conducted for
classification MDV
effects and MDE
compared to
as
theshown
entire in Figure
feature 6.
pool.

Sensors 2018, 18, x 14 of 21

(a) (b)
obtained by MDV. It can be seen
Figure that the feature subset through MDV(b).selection exhibits better
Figure 6.
6. Correlation
Correlation between
between features
features by
by MDV
MDV (a)
(a) and MDE
and MDE (b).
classification effects compared to the entire feature pool.
It is obvious that the top 10 features selected by MDE are highly correlated with each other. In
comparison, the overall collinearity of the features in MDV is much lower except for the small region
of the top three.
To visualize the feature selection performance by MDV, t-Distributed Stochastic Neighbor
Embedding (t-SNE) is employed which is a relatively new method of dimension reduction
particularly suitable for non-linear and high-dimensional datasets. It is a kind of manifold learning
technique by mapping to probability distributions through affine transformation. For detailed
algorithm, please refer to [45]. The 3D visualization by t-SNE is shown in Figure 7. The left figure is
the visualization using the entire feature pool while the right figure uses only top six features

(a) (b)
Figure 7. 3D
Figure visualization
7. 3D visualization by t-SNE:(a)(a)
by t-SNE: t-SNE
t-SNE usingusing original
original features;features;
(b) t-SNE(b) t-SNE
using using
selected selected
features.
features.
Further, machine learning techniques are used to quantify the flight state identification process.
Further,
For machine
each feature learning
selection techniques
method, the mostaresalient
used to quantify
6 features arethe flight state
obtained identification
as model inputs andprocess.
the
For each
16 flight states are set as model outputs. Five supervised learning models are employed including the
feature selection method, the most salient 6 features are obtained as model inputs and
16 flight states
Logistic are set (LR),
Regression as model
Support outputs. Five supervised
Vector Machine (SVM), Naïve learning
Bayes models are employed
(NB), Random Forest (RF),including
and
Neural Network (NN). Cross-validation is used in each model and
Logistic Regression (LR), Support Vector Machine (SVM), Naïve Bayes (NB), Random Forestthe average accuracy value of five(RF),
tests is computed
and Neural Networkto(NN). reduce the unbalance influence
Cross-validation is usedbetween
in eachtraining
model and and testing samples.
the average It should
accuracy value
be noted that since the objective of the case study is to compare the effects of different
of five tests is computed to reduce the unbalance influence between training and testing samples. It feature selection
methods instead of obtaining the optimized parameter setting for each machine learning model to
should be noted that since the objective of the case study is to compare the effects of different feature
achieve the highest accuracy level, default parameter settings in Python scikit-learn package for LR,
selection methods instead of obtaining the optimized parameter setting for each machine learning
SVM, NB and RF are used and remain the same for all feature selection methods while for NN, the
model to achieve the highest accuracy level, default parameter settings in Python scikit-learn package
parameter setting is as follows: {hidden layer size = 20, solver = ‘lbfgs’, activation function = ’relu’,
for LR, SVM,rate
learning NB= and
0.001,RF are usediteration
maximum and remain
= 100}.theThesame for all feature
identification selection
results are shown methods
in Figure 8.while for
NN, the parameter setting is as follows: {hidden layer size = 20, solver = ‘lbfgs’, activation function =
’relu’, learning rate = 0.001, maximum iteration = 100}. The identification results are shown in
Figure 8.
 selection methods instead of obtaining the optimized parameter setting for each machine learning
model to achieve the highest accuracy level, default parameter settings in Python scikit-learn package
for LR, SVM, NB and RF are used and remain the same for all feature selection methods while for
NN, the parameter setting is as follows: {hidden layer size = 20, solver = ‘lbfgs’, activation function =
Sensors’relu’, learning
2018, 18, 1379 rate = 0.001, maximum iteration = 100}. The identification results are shown in 15 of 21
Figure 8.

Figure
Figure 8. Identification
8. Identification accuracyagainst
accuracy against different
different feature
featureselection methods.
selection methods.

It can be observed that our proposed method MDV achieves the highest identification accuracy
Itincan be observed
all five that ourmodels
machine learning proposed
and method MDV
particularly, achieves
there the highest
is a significant identification
improvement accuracy
in Logistic
in all five machine learning models and particularly, there is a significant improvement in Logistic
Regression. This demonstrates the superior effectiveness of MDV. The comparison between MDV
and MDE shows that a group of individually powerful features with low collinearity can lead to
better results.

5.2. Stall Detection and Alerting

So far, the developed MDV algorithm has achieved the best performance in feature selection and
the final flight state identification accuracy is up to 100%. Herein, the second dataset with higher
resolution is used for the application of stall detection and alerting. Similarly, totally 47 features as
discussed in Section 3 are extracted and the most salient 6 features are selected by MDV as model
inputs. A neural network is employed with the same parameter settings as the first case. The split rule
is 80% samples for training and 20% samples for testing.
The classification report is shown in Table 6 including three criteria: Precision, Recall and
F1-score. Precision is the ratio of correctly predicted positive observations to the total predicted
positive observations while Recall is the ratio of correctly predicted positive observations to the all
observations in the actual class. F1-Score is the weighted average of Precision and Recall: F1-Score =
2 * (Recall * Precision)/(Recall + Precision) [46]. Safe, Alert, and Stall regions are divided with
corresponding flight states. The overall identification accuracy is 98%.

Table 6. Classification report.

States ID AoA (deg) Speed (m/s) Precision Recall F1-Score

1 11 10 0.92 1.00 0.96
2 11 13 0.92 1.00 0.96
Safe
3 11 16 1.00 0.92 0.96
4 11 19 1.00 0.92 0.96
5 12 10 1.00 1.00 1.00
6 12 13 1.00 0.92 0.96
Alert
7 12 16 0.92 1.00 0.96
8 12 19 1.00 1.00 1.00
9 13 10 1.00 1.00 1.00
10 13 13 1.00 1.00 1.00
Stall
11 13 16 1.00 1.00 1.00
12 13 19 1.00 1.00 1.00
Sensors 2018, 18, 1379 16 of 21

To facilitate detailed analysis, a normalized confusion matrix is presented in Figure 9. Each row
of the matrix represents the test samples in a true class label while each column indicates the samples
in a predicted class label [47]. As can be observed from Table 6, for stall states (ID: 9, 10, 11, 12), Recall
values all equal to 100%, meaning that all the critical states can be successfully identified and there is
no safety risk.
In terms of alert states (ID: 5, 6, 7, 8), Recall value of State 6 is 0.92, which means 92% samples in
State 6 are correctly predicted. By examining the 6th row in the confusion matrix, the rest 8% samples
are misclassified as State 1, which is in the safe region. This situation may lead to dangerous results
since the wing is already in the alert states yet there is no warning. From the other perspective, the
precision value of State 7 is 0.92, which indicates that among all samples predicted as State 7, there are
8% samples actually belonging to State 4 as shown in the 7th column of the confusion matrix. This
value can be interpreted as the false-alarm ratio that the wing flying in the safe region yet receives
a false alert.
For safe states (ID: 1, 2, 3, 4), the misclassified samples are for State 3 and State 4, in which 8%
samples of State 3 are predicted as State 2 while 8% samples of State 4 are identified as State 7, which
is the Sensors
false 2018, 18, x
alarm. 16 of 21

Figure
Figure 9. Confusionmatrix
9. Confusion matrix of
of flight
flight state
stateidentification.
identification.

Further, we select the different number of features from the modified distance evaluation (MDE)
Further, we select
method and use thethe different
same neural number of features
network structure forfrom theand
training modified
testing.distance evaluation
The comparison (MDE)
on the
method and use the same neural network structure for training and testing. The comparison
overall identification accuracy between MDV and various MDE is shown in Figure 10. The x axis on the
overall identification
denotes number ofaccuracy
top rankedbetween
features MDV and various MDE is shown in Figure 10. The x axis
selected.
denotes number of top ranked features selected.
It can be seen that if we use the same number of input as MDV, features selected by MDE lead to
a pool result of 0.33. The identification accuracy reaches the same level as MDV until the number of
top ranked features selected from MDE increases to 20. This shows that our proposed method MDV is
able to address the collinearity problem and uses fewer features to achieve superior performance with
a considerable model complexity reduction.
 Figure 9. Confusion matrix of flight state identification.

Further, we select the different number of features from the modified distance evaluation (MDE)
method and use the same neural network structure for training and testing. The comparison on the
Sensorsoverall
2018, 18,identification
1379 accuracy between MDV and various MDE is shown in Figure 10. The x axis 17 of 21
denotes number of top ranked features selected.

Figure 10. Identification accuracy between MDV and various MDE.

Figure 10. Identification accuracy between MDV and various MDE.
It can be seen that if we use the same number of input as MDV, features selected by MDE lead
6. Conclusions
to a pool result of 0.33. The identification accuracy reaches the same level as MDV until the number
of top ranked features selected from MDE increases to 20. This shows that our proposed method
This paper focuses on the feature engineering in structural vibration signals obtained from
MDV is able to address the collinearity problem and uses fewer features to achieve superior
a self-sensing composite
performance wing through
with a considerable wind
model tunnelreduction.
complexity experiments. In addition to common statistical
features from the time domain and frequency domain, complex features from the information domain
inspired by electroencephalograph analysis and mechanical fault diagnosis are also extracted, some of
which exhibit good classification ability. A novel filter feature selection method (MDV) is proposed by
combining the modified distance evaluation (MDE) algorithm and the variance inflation factor (VIF).
MDE is able to select individually powerful features but cannot address high collinearity. VIF is then
applied for each top ranked feature to remove highly correlated elements. Results from both general
flight state identification and stall detection & alerting demonstrate that this method can reduce the
model complexity with fewer features while maintain a high identification accuracy. Knowledge can
be gained by calculating the limited important features obtained by MDV efficiently for flight state
identification using light-weight machine learning models. This would save considerable efforts in
feature extraction and feature selection by manpower and has the potential to provide autonomous
control with real-time flight state monitoring. For multi-sensor utilizations, this method can be
applied to each sensor and ensemble methods can be developed to fuse multi-source results for more
robust identification.

Author Contributions: X.C. analyzed the data and developed the feature selection method; F.K. and F.-K.C.
designed the self-sensing wing and performed the wind tunnel experiments; Q.W. provides feature extraction
algorithms; H.R. and F.-K.C. coordinated the research and revised the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China (Grant No. 51705242),
Shanghai Sailing Program (Grant No. 16YF1404900), and the U.S. Air Force Office of Scientific Research (AFOSR)
program “Avian-Inspired Multi-functional Morphing Vehicles” under grant FA9550-16-1-0087 with Program
Manager Byung-Lip (“Les”) Lee.
Acknowledgments: The authors would like to thank Pengchuan Wang, Ravi Gondaliya, Jun Wu and Shaobo Liu
for their help during the wind tunnel experiments. Also, the authors would like to acknowledge the support of
Lester Su and John Eaton in the wind tunnel facility at Stanford University.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
The expressions of selected features by different feature selection methods are shown in Table A1.
Sensors 2018, 18, 1379 18 of 21

Table A1. Top 10 ranking feature expressions.

R UFS_m SVM_L1 GBDT STAB MDE MDV

1
N
∑ n =1 ( x (n)−t1 )9 I3 = MSE[3] I9 = HST I9 = HST f 10 =
f6
f 10 =
f6
t25 = N · t2 9 f5 f5
K
∑ k =1 f r k 2 y ( k ) N K K
∑k=1 ( f rk ·y(k))
f9 = √ ∑n=1 | x (n)−t1 |
f 1 = ∑k=N
2 I5 = PFD I2 = MSE[2] y(k) f5 =
K K
∑ k =1 y ( k ) ∑ k =1 f r k 4 y ( k )
t12 = N
1
K
∑ k =1 y ( k )
5
q N
q 2
N
( x (n))2 ∑n=1 ( x (n)−t1 ) N
( x (n)−t )2
N
∑n=1 ( x (n)−t1 )
3 t 6 = ∑ n =1 N I1 = MSE[1] I8 = ApEn t21 = N · t2 5 t 2 = ∑ n =1 N 1 t5 = N
K 3
9 4 ∑k=1 (y(k )− f 1 )
q N N
q
N
( x (n)−t )2 ∑n=1 ( x (n)−t1 ) t6 ∑n=1 ( x (n)−t1 ) N
( x (n))2
4 t 2 = ∑ n =1 N 1 t25 = N · t2 9
t14 = t8 t20 = N · t2 4 t 6 = ∑ n =1 N f3 =
K(
√
f2 )
3

3
2 N
q
N
∑n=1 ( x (n)−t1 ) ∑n=1 ( x (n)−t1 ) K
( f r − f )2 y ( k )
5 t5 = N
I8 = ApEn I1 = MSE[1] t19 = N · t2 3 f 6 = ∑k=1 kK 5 I4 = PMMSE
K 3
N
∑n=1 ( x (n)−t1 )
4 N
∑n=1 ( x (n)−t1 )
3
t4 ∑k=1 (y(k)− f 1 )
6 t4 = t19 = I6 = HFD t18 = t6 4
f3 = √ 3 I7 = FI
N N · t2 3 K( f2 )
r
K
∑ k =1 f r k 4 y ( k ) t17 = t3 N
∑n=1 | x (n)−t1 |
7 I2 = MSE[2] f8 = K I3 = MSE[3] t6 3 t12 = I3 = MSE[3]
∑ k =1 f r k 2 y ( k ) N
N 7
8 ∑n=1 ( x (n)−t1 ) t13 = t9 N
∑ n =1 x ( n ) t16 = t9 N
∑n=1 | x (n)| I8 = ApEn
t23 = N · t2 7 t6 t1 = N t7 t8 = N
5 K 3
9 I4 = PMMSE I6 = HFD
N
∑ n =1 ( x (n)−t1 ) t15 = t9 ∑ k =1 ( f rk − f 5 ) y(k )
t14 = t6
t21 = N · t2 5 t8 f 11 = K· f6 3 t8
N 7
10 t17 = t3
t10 = min( x (n)) I7 = FI t14 = t6
t10 = min( x (n)) ∑n=1 ( x (n)−t1 )
t6 3 t8 t23 = N · t2 7
Sensors 2018, 18, 1379 19 of 21

References
1. NASA Fly-By-Feel Systems Represent The Next Revolution In Aircraft Controls. Available online:
https://www.nasa.gov/centers/dryden/news/X-Press/aerovations/2011/fly-by-feel.html (accessed on
16 May 2017).
2. Kopsaftopoulos, F.; Nardari, R.; Li, Y.H.; Chang, F.K. A stochastic global identification framework for
aerospace structures operating under varying flight states. Mech. Syst. Signal Process. 2018, 98, 425–447.
[CrossRef]
3. Brenner, M.J. Controller Performance Evaluation of Fluy-by-feel (FBF) Technology. Available online:
https://nari.arc.nasa.gov/node/448 (accessed on 16 May 2017).
4. Mangalam, A.S.; Brenner, M.J. Fly-by-Feel Sensing and Control: Aeroservoelasticity. AIAA Atmos. Flight
Mech. Conf. 2014. [CrossRef]
5. Suh, P.; Chin, A.; Mavris, D. Virtual Deformation Control of the X-56A Model with Simulated Fiber Optic
Sensors. AIAA Atmos. Flight Mech. Conf. 2013. [CrossRef]
6. Suh, P.M.; Chin, A.; Mavris, D.N. Robust Modal Filtering and Control of the X-56A Model with Simulated
Fiber Optic Sensor Failures. AIAA Atmos. Flight Mech. Conf. 2014. [CrossRef]
7. Lanzara, G.; Feng, J.; Chang, F.-K. Design of Micro-Scaled Highly Expandable Networks of Polymer Based
Substrates for Macro-Scale Applications. Smart Mater. Struct. 2010, 19, 045013. [CrossRef]
8. Salowitz, N.; Guo, Z.; Li, Y.H.; Kim, K.; Lanzara, G.; Chang, F.K. Bio-inspired stretchable network-based
intelligent composites. J. Compos. Mater. 2013, 47, 97–105. [CrossRef]
9. Salowitz, N.; Guo, Z.; Roy, S.; Nardari, R.; Li, Y.H.; Kim, S.J.; Kopsaftopoulos, F.; Chang, F.K.
Recent advancements and vision toward stretchable bio-inspired networks for intelligent structures.
Struct. Heal. Monit. 2014, 13, 609–620. [CrossRef]
10. Guo, Z.; Aboudi, U.; Peumans, P.; Howe, R.T.; Chang, F.K. A Super Stretchable Organic Thin-Film Diodes
Network That Can Be Embedded into Carbon Fiber Composite Materials for Sensor Network Applications.
J. Microelectromech. Syst. 2016, 25, 524–532. [CrossRef]
11. Kopsaftopoulos, F.P.; Nardari, R.; Li, Y.-H.; Wang, P.; Ye, B.; Chang, F.-K. Experimental identification of
structural dynamics and aeroelastic properties of a self-sensing smart composite wing. In Proceedings of the
10th International Workshop on Structural Health Monitoring, Stanford, CA, USA, 1–3 September 2015.
12. Roy, S.; Lonkar, K.; Janapati, V.; Chang, F.-K. A novel physics-based temperature compensation model for
structural health monitoring using ultrasonic guided waves. Struct. Heal. Monit. Int. J. 2014, 13, 321–342.
[CrossRef]
13. Huang, R.; Zhao, Y.; Hu, H. Wind-Tunnel Tests for Active Flutter Control and Closed-Loop Flutter
Identification. AIAA J. 2016, 54, 1–11. [CrossRef]
14. Pang, Z.Y.; Cesnik, C.E.S. Strain state estimation of very flexible unmanned aerial vehicle. In Proceedings of
the 57th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, San Diego,
CA, USA, 4–8 January 2016.
15. Sodja, J.; Werter, N.; Dillinger, J.K.; De Breuker, R. Dynamic Response of Aeroelastically Tailored Composite
Wing: Analysis and Experiment. In Proceedings of the 57th AIAA/ASCE/AHS/ASC Structures, Structural
Dynamics, and Materials Conference, San Diego, CA, USA, 4–8 January 2016.
16. Kopsaftopoulos, F.; Nardari, R.; Li, Y.-H.; Wang, P.; Chang, F.-K. Stochastic global identification of
a bio-inspired self-sensing composite UAV wing via wind tunnel experiments. In Proceedings of the
Health Monitoring of Structural and Biological Systems 2016. International Society for Optics and Photonics,
Las Vegas, NV, USA, 2016.
17. Kopsaftopoulos, F.P.; Fassois, S.D. Vector-dependent functionally pooled ARX models for the identification
of systems under multiple operating conditions. IFAC Proc. Vol. 2012, 16, 310–315. [CrossRef]
18. Guyon, I. Feature Extraction Foundations and Applications; Springer: Berlin, Germany, 2006; Volume 207,
ISBN 9783540354871.
19. Samanta, B. Gear fault detection using artificial neural networks and support vector machines with genetic
algorithms. Mech. Syst. Signal Process. 2004, 18, 625–644. [CrossRef]
20. Shen, Z.; Chen, X.; Zhang, X.; He, Z. A novel intelligent gear fault diagnosis model based on EMD and
multi-class TSVM. Meas. J. Int. Meas. Confed. 2012, 45, 30–40. [CrossRef]
Sensors 2018, 18, 1379 20 of 21

21. Xi, Y.L.; Xi, Z.H.; Xi, Y.Z. Fault Diagnosis of Rotating Machinery Based on Multiple ANFIS Combination
with Gas Fault diagnosis of rotating machinery based on multiple ANFIS combination with GAs. Mech. Syst.
Signal Process. 2007, 21, 2280–2294. [CrossRef]
22. Guyon, I.; Elisseeff, A. An Introduction to Variable and Feature Selection. J. Mach. Learn. Res. 2003, 3,
1157–1182. [CrossRef]
23. Zhou, L.; Si, Y.W.; Fujita, H. Predicting the listing statuses of Chinese-listed companies using decision trees
combined with an improved filter feature selection method. Knowl.-Based Syst. 2017, 128, 93–101. [CrossRef]
24. Chandrashekar, G.; Sahin, F. A survey on feature selection methods. Comput. Electr. Eng. 2014, 40, 16–28.
[CrossRef]
25. Kohavi, R.; John, G.H. Wrappers for feature subset selection. Artif. Intell. 1997, 97, 273–324. [CrossRef]
26. Saeys, Y.; Inza, I.; Larrañaga, P. A review of feature selection techniques in bioinformatics. Bioinformatics
2007, 23, 2507–2517. [CrossRef] [PubMed]
27. Chen, X.; Kopsaftopoulos, F.; Cao, H.; Chang, F.-K. Intelligent flight state identification of a self-sensing wing
through neural network modelling. In Proceedings of the 11th International Workshop on Structural Health
Monitoring, IWSHM 2017, Stanford, CA, USA, 12–14 September 2017.
28. Welch, P.D. The Use of Fast Fourier Transform for the Estimation of Power Spectral: A Method Based on Time
Averaging Over Short Modified Periodograms. IEEE Trans. Audio Electroacoust. 1967, 15, 70–73. [CrossRef]
29. He, Y.; Huang, J.; Zhang, B. Approximate entropy as a nonlinear feature parameter for fault diagnosis in
rotating machinery. Meas. Sci. Technol. 2012, 23, 045603. [CrossRef]
30. Petrosian, A. Kolmogorov complexity of finite sequences and recognition of different preictal EEG patterns.
In Proceedings of the Eighth IEEE Symposium on Computer-Based Medical Systems, Lubbock, TX, USA,
9–10 June 1995.
31. Costa, M.; Goldberger, A.L.; Peng, C. Multiscale Entropy Analysis of Complex Physiologic Time Series.
Phys. Rev. Lett. 2002, 89, 6–9. [CrossRef] [PubMed]
32. Jiang, W.; Dong, K.; Zhu, Y.; Wang, H. Fault Feature Identification Based on Partial Mean of Multi-scale
Entropy for Hydraulic Pump. Hydraul. Pneum. 2016, 4, 518–522. [CrossRef]
33. Falconer, K. Fractal geometry: Mathematical foundations and applications, 2nd ed.; John Wiley & Sons:
West Sussex, UK, 2003.
34. Higuchi, T. Approach to an irregular time series on the basis of the fractal theory. Phys. D Nonlinear Phenom.
1988, 31, 277–283. [CrossRef]
35. James, C.J.; Lowe, D. Extracting multisource brain activity from a single electromagnetic channel.
Artif. Intell. Med. 2003, 28, 89–104. [CrossRef]
36. Pincus, S.M.; Gladstone, I.M.; Ehrenkranz, R. A regularity statistic for medical data analysis. J. Clin. Monit.
1991, 7, 335–345. [CrossRef] [PubMed]
37. Balli, T.; Palaniappan, R. A combined linear & nonlinear approach for classification of epileptic EEG signals.
In Proceedings of the 2009 4th International IEEE/EMBS Conference on Neural Engineering, Antalya, Turkey,
29 April–2 May 2009; pp. 714–717.
38. Yang, B.; Kim, K. Application of Dempster–Shafer theory in fault diagnosis of induction motors using
vibration and current signals. Mech. Syst. Signal Process. 2006, 20, 403–420. [CrossRef]
39. Dormann, C.F.; Elith, J.; Bacher, S.; Buchmann, C.; Carl, G.; Carré, G.; Marquéz, J.R.G.; Gruber, B.;
Lafourcade, B.; Leitão, P.J.; et al. Collinearity: A review of methods to deal with it and a simulation
study evaluating their performance. Ecography 2013, 36, 27–46. [CrossRef]
40. Peng, H.; Long, F.; Ding, C. Feature selection based on mutual information: Criteria of max-dependency.
IEEE Trans. Pattern Anal. Mach. Intell. 2005, 27, 1226–1238. [CrossRef] [PubMed]
41. Burges, C.J.C. A tutorial on support vector machines for pattern. Data Min. Knowl. Discov. 1998, 2, 955–974.
[CrossRef]
42. Ogutu, J.O.; Schulz-Streeck, T.; Piepho, H.-P. Genomic selection using regularized linear regression models:
Ridge regression, lasso, elastic net and their extensions. BMC Proc. 2012, 6, S10. [CrossRef] [PubMed]
43. Natekin, A.; Knoll, A. Gradient boosting machines, a tutorial. Front. Neurorobot. 2013, 7, 21. [CrossRef]
[PubMed]
44. Meinshausen, N. Stability selection (Slides). J. R. Stat. Soc. Ser. B Stat. Methodol. 2009, 72, 1–30. [CrossRef]
45. Van Der Maaten, L.; Hinton, G. Visualizing Data using t-SNE. J. Mach. Learn. Res. 2008, 620, 267–284.
[CrossRef]
Sensors 2018, 18, 1379 21 of 21

46. Davis, J.; Goadrich, M. The relationship between Precision-Recall and ROC curves. In Proceedings of the
23rd International Conference on Machine Learning—ICML ’06, New York, NY, USA, 25–29 June 2006;
pp. 233–240.
47. Powers, D.M.W. Evaluation: From Precision, Recall and F-Measure To Roc, Informedness, Markedness &
Correlation. J. Mach. Learn. Technol. 2011, 2, 37–63.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).