9th IFAC Symposium on Fault Detection, Supervision and

9th IFAC Symposium on Fault Detection, Supervision and

9th IFAC
Safety of Symposium on Fault Detection, Supervision and
Technical Processes
9th IFAC
Safety of Symposium on Fault Detection, Supervision and
Technical Processes
Safety of Technical
September 2-4, Processes
2015.
Safety of Technical Arts Available online
et Mtiers ParisTech,
Processes
at www.sciencedirect.com
Paris, France
September 2-4, 2015. Arts et Mtiers ParisTech, Paris, France
September 2-4, 2015. Arts et Mtiers ParisTech, Paris, France
September 2-4, 2015. Arts et Mtiers ParisTech, Paris, France
ScienceDirect
IFAC-PapersOnLine 48-21 (2015) 541546
Diagnosis
Diagnosis of
of Wing
Wing Icing
Icing Through
Through Lift
Lift and
and
Diagnosis
Diagnosis of Wing Icing
of Wing ChangeThrough
Icing Through Lift and
Liftfor
and
Drag
Drag Coefficient
Coefficient Change Detection
Detection for
Drag Coefficient
Drag Small
Coefficient Change Detection
ChangeAircraft for
 for
Detection
Small Unmanned
Unmanned Aircraft 
Small Unmanned Aircraft 
Small Unmanned Aircraft
,
Kim
Kim Lynge
Lynge Srensen
Srensen Mogens Blanke ,
Mogens Blanke ,
Kim Lynge
LyngeTor Srensen
Arne Mogens Blanke ,
Kim Tor
Tor Arne Johansen
Srensen Mogens Blanke
Johansen
Tor Arne
Arne Johansen
Johansen

Autonomous Marine Operations and Systems, Department of
Autonomous Marine Operations and Systems, Department of
Autonomous
Engineering
Autonomous Marine
Marine Operations
Cybernetics, the Norwegian
Operations and
and Systems,
University
Systems, Department
of Science
Department of
and
of
Engineering
Engineering Cybernetics,
Cybernetics, the
the Norwegian
Norwegian University
University of
of Science
Science and
and
Technology,
Engineering
Technology, O.
O. S.
S. Bragstads
Cybernetics,
Bragstads thePlass
Plass 2D,
Norwegian
2D, NO-7491
University
NO-7491 Trondheim,
of
Trondheim, Science Norway
and
Norway
Technology,
(e-mail:
Technology, O. S. Bragstads
kim.sorensen@itk.ntnu.no, Plass 2D, NO-7491 Trondheim,
mogens.blanke@itk.ntnu.no,
O. S. Bragstads Plass 2D, mogens.blanke@itk.ntnu.no,
NO-7491 Trondheim, Norway Norway
(e-mail:
(e-mail: kim.sorensen@itk.ntnu.no,
(e-mail: kim.sorensen@itk.ntnu.no,
kim.sorensen@itk.ntnu.no, mogens.blanke@itk.ntnu.no,
tor.arne.johansen@itk.ntnu.no).
mogens.blanke@itk.ntnu.no,
tor.arne.johansen@itk.ntnu.no).

Department tor.arne.johansen@itk.ntnu.no).
of Electrical Engineering,
tor.arne.johansen@itk.ntnu.no). Technical University of
Department of Electrical Engineering, Technical University of
Department of Electrical Engineering, Technical University
Denmark,
Department
Denmark, DK-2800
ofDK-2800 Kgs.
ElectricalKgs. Lyngby,
Engineering,
Lyngby, Denmark
Technical(e-mail:
Denmark University of
(e-mail: of
Denmark,
Denmark, DK-2800DK-2800 Kgs. Lyngby,
mb@elektro.dtu.dk) Denmark
Kgs. Lyngby, Denmark (e-mail: (e-mail:
mb@elektro.dtu.dk)
mb@elektro.dtu.dk)
mb@elektro.dtu.dk)
Abstract:
Abstract: This
This paper
paper address
address the
the issue
issue of
of structural
structural change,
change, caused
caused by
by ice
ice accretion,
accretion, on
on
Abstract:
UAVs
Abstract: by This
utilising
This paper
a
paper address
Neyman
address the
Pearson
the issue
(NP)
issue of
of structural
based change,
statistical
structural change, caused
change
caused by
detection
by ice
ice accretion,
approach,
accretion, on
for
on
UAVs
UAVs by
by utilising
utilising a
a Neyman
Neyman Pearson
Pearson (NP)
(NP) based
based statistical
statistical change
change detection
detection approach,
approach, for
for
the
UAVs
the identification
by utilising
identification of
a
of structural
Neyman
structural changes
Pearson
changes of
(NP)
of fixed
based
fixed wing
wing UAV
statistical
UAV airfoils.
change
airfoils. A
A structural
detection
structural analysis
approach,
analysis is
for
is
the identification
performed
the identificationon the of
of structural
nonlinear
structural changes
aircraft
changes of
system fixed
and
of fixed wing
residuals
wing UAV UAV are airfoils.
generated,
airfoils. A structural
where
A structural a analysis
analysis is
generalised is
performed
performed on
on the
the nonlinear
nonlinear aircraft
aircraft system
system and
and residuals
residuals are
are generated,
generated, where
where aa generalised
generalised
likelihood
performed
likelihood ratio
on the
ratio test
test is
is applied
nonlinear
applied to
aircraft
to detect
system
detect faults.
and Numerical
faults. residuals are
Numerical simulations
generated,demonstrate
simulations where a generalised
demonstrate aa robust
robust
likelihood
detection
likelihood ratio
with test
test is
ratioadequate is applied
balance
applied to detect
tobetween faults.
detect false
faults. Numerical
alarm rate
Numerical and simulations
sensitivity.
simulations demonstrate
demonstrate aa robust robust
detection
detection with
with adequate
adequate balance
balance between
between false
false alarm
alarm rate
rate and
and sensitivity.
sensitivity.
detection with adequate balance between false alarm rate and sensitivity.
2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords:
Keywords: Structural
Structural ChangeChange Detection;
Detection; Aircraft
Aircraft IceIce accretion;
accretion; Aircraft
Aircraft IcingIcing Detection;
Detection; UAV; UAV;
Keywords:
Fixed-Wing
Keywords: Structural
UAV;
Structural Change
Statistical
Change Detection;
Change
Detection; Aircraft
Detection;
Aircraft Ice
Ice accretion;
accretion; Aircraft
Aircraft Icing
Icing Detection;
Detection; UAV; UAV;
Fixed-Wing
Fixed-Wing UAV; Statistical Change Detection;
Fixed-Wing UAV; UAV; Statistical
Statistical Change
Change Detection;
Detection;
1. INTRODUCTION
1. INTRODUCTION
INTRODUCTION temperature at the point of the droplets impact with the
1. temperature
temperature at the point of the droplets impact with the
1. INTRODUCTION aircraft,
temperature is 0 at
C
at the
or
the point
colder.
point of
of the
the droplets
droplets impact
impact with
with the
the
aircraft,
aircraft, is
is 0
0 C or colder.
C or colder.
Structural changes due to ice accretion are common causes aircraft, is 0 C or colder.
Structural changes
Structural changes due due to to ice
ice accretion
accretion are are common
common causes causes The The risks
risks of
of aircraft
aircraft icing
icing have
have been
been known
known since
since the
the early
early
for
for unmanned
Structural
unmanned aerial
changes
aerialdue vehicle
to ice incidents
vehicle accretion are
incidents in
in cold
common
cold and
and humid
causes The
humid risks
1900s. Theof aircraft
effects icing
of have
icing been
depends known
upon since
the the early
location,
for unmanned aerial vehicle incidents in cold and humid The risks
1900s. Theof aircraft
effects icing
of have
icing been
depends known
upon since
the the early
location,
regions.
for
regions. For
unmanned
For fixed
aerial
fixed wing
wing UAVs
vehicle
UAVs the
incidents
the leading
in
leading edge
cold
edge and of
of airfoil
humid
airfoil 1900s.
and the Thetype,effects
of of
the icing
formed depends
ice. upon
Icing the
can location,
occur on
regions. For fixed wing UAVs the leading edge of airfoil 1900s.
and the Thetype,effects
of of
the icing
formed depends
ice. upon
Icing the
can location,
occur on
surfaces is the primary surface exposed to these changes, and the type, of the formed ice. Icing can occur
regions.
surfaces is
surfaces
For
is the fixed
the primary wing
primary surface UAVs the
surface exposed leading
exposed to edge
to these of
these changes,airfoil
changes, wings, wings,
and the control surfaces,
type,surfaces,
control horizontal
of the horizontal
formed ice. and
and vertical
Icing
vertical occur on
stabilizers,
can stabilizers, on
causing a significant
surfaces aissignificant
the primary reduction
surface inin aerodynamic
exposed to these ability, i.e.
changes, wings,
fuselage control
nose, surfaces,
landing horizontal and vertical stabilizers,
causing
causing a significant reduction
reduction in aerodynamic
aerodynamic ability,
ability, i.e.
i.e. wings,
fuselage control
nose, surfaces,gear
landing gear doors,
horizontal
doors, engine
and vertical
engine intakes,
intakes, fuselage
stabilizers,
fuselage
decreasing lift and manoeuvrability, and increasing drag,
i.e. fuselage nose, landing gear
causing
decreasing
decreasing
a significant
lift and
lift
reduction in aerodynamic
and manoeuvrability,
manoeuvrability, and
and increasing
ability,
increasing drag,
drag,
air data
fuselage
air data ports
nose,
ports and
landing
and gear doors,
sensors,
sensors, and
doors,
and
engine
drain
engine
drain
intakes,
system
intakes,
system
fuselage
outputs.
fuselage
outputs. In
In
weight,
decreasing
weight, and
and consequently
lift and
consequently power
manoeuvrability,
power consumption.
and
consumption. Timely
increasing
Timely de-
drag,
de- air
this data
paper ports
focus andis sensors,
on detectingand drain
ice system
forming on outputs.
the leading In
weight, and consequently power consumption. Timely de- air
this data
paper ports
focus andis sensors,
on detectingand drain
ice system
forming on outputs.
the leading In
tection
weight,
tection of
and
of such changes
consequently
such changes could
power
could potentially
consumption.
potentially prevent
Timely
prevent icing
de-
icing this
edge paper
of a focus
UAV is
wings.on detecting ice forming on the leading
tection of this
edge paper
of a focus
UAV is
wings.on detecting ice forming on the leading
related
tection UAV
related of such
UAV such changes
changes could
incidents.
incidents. could potentially
potentially preventprevent icingicing edge edge of aa UAV
UAV wings.
ofliterature wings.
related
related UAV
UAV incidents.
incidents. In
In the
the literature there
there exists
exists numerous
numerous approaches
approaches to
to fault
fault
The
The use
use of
of unmanned
unmanned aerial
aerial vehicles
vehicles (UAV)s
(UAV)s has
has increased
increased In the
detectionliterature
and there
isolation exists
(FDI) numerous
techniques approaches
applied to
forfault
the
The In the
detectionliterature
and there
isolation exists
(FDI) numerous
techniques approaches
applied to
forfault
the
The use
use of
of unmanned
significantly
significantly within
unmanned
within the
the
aerial
last
aerial
last
vehicles
decade,
vehicles
decade,
(UAV)s
operating
(UAV)s
operating
has inincreased
hasin surveil- detection and isolation (FDI)
increased
surveil- detection and isolation of techniques
actuator
(FDI) techniquesand applied
sensor
applied for
faults
for the
on
the
significantly
lance and within
reconnaissancethe last decade,
primarily. operating
UAVs are in surveil-
very well detection
detection and
and isolation
isolation of
of actuator
actuator and
and sensor
sensor faults
faults on
on
significantly
lance and within
reconnaissancethe last decade,
primarily. operating
UAVs are in surveil-
very well UAVs,
detection see Bateman
and isolation et ofal. (2011),
actuator Ducard
and sensor(2009),
faults and
on
lance
suited and
for reconnaissance
operating in primarily.
conditions UAVs
that are are very
deemed well
un- UAVs,
UAVs, see
see Bateman
Bateman et
et al.
al. (2011),
(2011), Ducard
Ducard (2009),
(2009), and
and
lance
suited and
for reconnaissance
operating in primarily.
conditions UAVs
that are are very
deemed well
un- Ducard
UAVs, and
see Geering
Bateman (2008)
et al. for
(2011), the former
Ducard and
(2009),for the
and
suited forhumans,
operating in Ducard and Geering (2008) for the former and for the
safe
suitedfor
safe forforhumans,
operating
humans, in conditions
Arctic
Arctic operations
conditions
operationsthat
that
being
being
are deemed
are relevant
deemed and
relevant
un-
un- Ducard
and later
Ducard see and Geering
Fravolini
and Geering et al.(2008)
(2009).
(2008) for the
the former
Detection
for of
former and
and for
control the
surface
for the
safe for
significant mentioning. Arctic operations
Consequently, being
reliable relevant
and and
efficient later
later see
see Fravolini
Fravolini et
et al.
al. (2009).
(2009). Detection
Detection of
of control
control surface
surface
safe for humans,
significant mentioning. Arctic operations reliable
Consequently, being relevant
and and later
efficient defectsseewas studied
Fravolini in
et al. Blanke
(2009). and
Detection Hansen (2013),
of control where
surface
significant
UAV mentioning.
operation in harsh Consequently,
environments, reliable
as and
the efficient defects was studied in Blanke and Hansen (2013), where
significant
UAV operation mentioning.
operation in harsh
harsh Consequently,
environments, reliable
as the andArctic,
the Arctic, is
efficient
is faultdefects
defects was
was studied
parameter studied in
in Blanke
adaptationBlankeand and
andfault Hansen
Hansen (2013),
diagnosis
(2013),was where
at-
where
UAV
desirable. in environments, as Arctic, is fault
fault parameter
parameter adaptation
adaptation and
and fault
fault diagnosis
diagnosis was
was at-
UAV operation
desirable.
desirable.
in harsh environments, as the Arctic, is tempted simultaneously.
fault parameter
tempted adaptation
simultaneously. For
For the
the specific
andspecific detection
fault detection was at-
problem
diagnosis problem at-
desirable.
In aviation, icing conditions are atmospheric conditions tempted
addressed simultaneously.
in this
tempted simultaneously. paper, For the
literature specific
For the specificis more detection
sparse.
detection problem
In Tousi
problem
In aviation,
aviation, icingicing conditions
conditions are are atmospheric
atmospheric conditions
conditions addressedaddressed in
in this paper, literature is more sparse. In Tousi
In
that
In can
aviation, lead to
icing the formation
conditions of ice
are atmospheric on aircraft.
conditionsIn- and
and Khorasani
addressed
Khorasani in this
this paper,
(2009)
paper,
(2009) and
and
literature
Tousi
literature
Tousi
is
is more
and
and Khorasani
more
Khorasani
sparse. In
In Tousi
sparse.(2011)
(2011)Tousiic-
ic-
that
that can
can lead
lead to
to the
the formation
formation of
of ice
ice on
on aircraft.
aircraft. In-
In- and
ing Khorasani
is diagnosed (2009)
through and an Tousi and
observer-based Khorasani fault (2011)
diagnosis ic-
flight
that canicing can
lead occur
to when
thewhen an
formation aircraft passes
of icepasses through
on aircraft. air
In- and
ing Khorasani
is diagnosed (2009)
through and an Tousi and
observer-based Khorasani fault (2011)
diagnosis ic-
flight
flight icing
icing can
can occur
occur when an
an aircraft
aircraft passes through
through air
air ing is diagnosed
technique that through
detects andan observer-based
estimates the fault
percentage diagnosis
of ice
that
flight contain
icing droplets
can occur of
whenwateran (humid
aircraft air),
passes and where
through the
air ing is diagnosed
technique that through
detects andanestimates
observer-based the fault diagnosis
percentage of ice
that contain
contain droplets of of water (humid
(humid air), and and where the the technique that detects and
that
that contain droplets
droplets of water water (humid air), air), and wherewhere the technique
present
present on
on the
that
the aircraft
detects wing,
aircraft and estimates
wing, relying
estimates
relying
the
theaa percentage
on
on linearised
percentage
linearised
of
of ice
lateral
ice
lateral
present
model
present of on
onthethe
the aircraft
aircraft.
aircraft Inwing, relying
Cristafaro
wing, relying eton
onal.a
a linearised
(2015)
lineariseda lateral
multiple
lateral
model
model of
of the
the aircraft. In Cristafaro et al. (2015) a multiple


models
model
models the aircraft.
ofadaptive
adaptive aircraft. In
In Cristafaro
estimation
estimation framework
Cristafaro
framework
et
et al.
al.is
is
(2015)
proposed
(2015)
proposed
a
a multiple
and
multiple
and in
in

This
 work has been carried out at the Centre for Autonomous Marine models
Cristafaro adaptive
models adaptive and estimation
Johansen
estimation framework
(2015) an
framework input is proposed
observer
is proposed and
and in
frame- in
This work has been carried out at the Centre for Autonomous Marine Cristafaro
Cristafaro and
and Johansen
Johansen (2015)
(2015) an
an input
input observer
observer frame-
frame-
This work has
Operations
This work has and been
been
carried
Systems
carried
out at the
(AMOS),
out at the
Centre for
supported
Centre for byAutonomous
the ResearchMarineCoun- work for
Cristafaro
work for designing
and Johansen
designing icing
icing diagnosis
(2015) anfilters
diagnosis inputis
filters is developed
observer
developed and
frame-
and
Operations and Systems (AMOS), supported byAutonomous
the ResearchMarineCoun- work for
Operations
cil of Norway
Operations
cil of Norway
andthrough
Systemsthe
andthrough
Systemsthe
(AMOS),
Centres
(AMOS),
Centres
supported by thefunding
of Excellence
supported
Researchscheme,
by thefunding
of Excellence
Coun-
Researchscheme,
Coun- work for designing
validated
validated through
designing
through aaicing
case
icing
case
diagnosis
study.
diagnosis
study.
filters
filters is is developed
developed and and
cil of Norway
Project number
cil of Norway
through
223254 the
- Centres
AMOS. The of Norwegian
Excellence Research
funding scheme,
Council validated through a
validated through a case study. case study.
Project numberthrough
223254 the Centres
- AMOS. The of Norwegian
Excellence Research
funding scheme,
Council
Project
is number 223254
acknowledged - AMOS.
as the main TheofNorwegian
sponsor AMOS. Research Council
Project number 223254
is acknowledged - AMOS.
as the main TheofNorwegian
sponsor AMOS. Research Council
is acknowledged as the main sponsor of AMOS.
is acknowledged as the main sponsor of AMOS.
Copyright
2405-8963 2015 IFAC 541 Hosting by Elsevier Ltd. All rights reserved.
Copyright 2015, IFAC (International Federation of Automatic Control)
2015 IFAC 541
Copyright
Peer review 2015 responsibility
IFAC 541Control.
Copyright under
2015 IFAC of International Federation of Automatic
541
10.1016/j.ifacol.2015.09.582
SAFEPROCESS 2015
542
September 2-4, 2015. Paris, France Kim Lynge Srensen et al. / IFAC-PapersOnLine 48-21 (2015) 541546

This paper addresses the issue of detecting ice when it
occurs on the leading edge of small scale UAV wings, in
flight, by means of a structural change detection and iden-
tification approach using a nonlinear longitudinal model
of the aircraft. Under the assumption that model uncer-
tainties, such as unknown aerodynamic coefficients, are
a priori estimated, the scheme is based on a structural
analysis performed on the system, with residual generation
as an outcome. A generalised likelihood ratio test (GLRT)
is then utilised to detect any changes in the generated
residuals, where any constant change would constitute an
occurance of a fault.
with acceleration components ax , ay , and az decomposed
in the BODY frame. Let the wind velocity vector relative
to Earth be denoted vw = (uw , vw , ww ). The aircraft
velocity relative to the wind velocity is then vr = vg vw .
Consequently the relative velocity in the BODY frame,
denoted vbr = (ur , vr , wr ), can be written as
     
ur u uw
vr = v Rn vw , b
(4)
wr w ww
where Rnb is the rotation matrix from NED to BODY frame
defined by the Euler angles roll (), pitch (), and yaw ()

The remainder of the paper is organized as follows. In The relative velocity components ur , vr , and wr are related
Section 2, background theories and modelling used in the to the airspeed Va through
paper are reviewed. Section 3 provides the main contri- ur = Va cos() cos(), (5)
bution of the paper, where the proposed fault diagnosis vr = Va sin(), (6)
system for detecting and isolating faults caused by icing on wr = Va sin() cos(), (7)
UAVs is presented. A performance assessment is conducted 
through simulations in Section 4. Finally, the paper is Va = u2r + vr2 + wr2 , (8)
concluded in Section 5. with and denotes the angle-of-attack (AOA) and
sideslip angle (SSA). For small scale UAVs the AOA and
2. PRELIMINARIES SSA are generally not measured, nor is the wind speed,
instead they require estimation. In Long and Song (2009)
This section provides background information needed to a sensor fusion approach is applied for just this purpose,
develop the proposed fault diagnosis solution. First the where the estimation is based on estimates of velocities
consequences of icing is described followed by a presenta- and wind speeds, and on the general relations (see Beard
tion of the aircraft model used throughout the paper. and McLain (2012), Long and Song (2009), and Stevens
and Lewis (2003))
2.1 Icing Consequences    
1 wr 1 vr
= tan , = sin .
In flight icing forming on the leading edge of an aircraft ur Va
causes a disruption to the airflow around the wings, i.e.
a change in the aerodynamic properties of the wing. A similar objective, but different approach, is utilised
The consequences are identified as a reduction in lift in Johansen et al. (2015), where wind velocities, AOA,
and controllability, as well as an increase in drag and and SSA are estimated through kinematic relationships
weight. Most UAVs are equipped with an autopilot. The and a Kalman filter, thereby avoiding the need to know
autopilot will control the aircraft to maintain a certain aerodynamic models or other aircraft parameters. Based
height above ground, a certain forward airspeed, and a on either of the aforementioned estimation methods , ,
course, as specified by any user. When ice forms on the vw , and consequently Va (denoted Va ) is assumed known.
leading edge of the wings of the aircraft, the autopilot will,
in an attempt to maintain the height and speed, increase 2.4 Aircraft Model - Forces and Aerodynamics
the angle-of-attack of the aircraft and the thrust of the
engine. Let the specific force vector, decomposed in the BODY
frame, be (fx , fy , fz ), defined as
2.2 Sensors fx = ax + g sin () , (9)
Utilisation of the following sensor suite is assumed fy = ay g sin () cos () , (10)
fz = az g cos () cos () , (11)
An IMU procures measurements of angular velocities
where g is the gravitational constant. (fx , fy , fz ) is related
and specific force
to the aerodynamic forces and thrust force of the aircraft
A pitot-static tube providing measurements of the
by
relative velocity in the longitudinal axis of the aircraft
An engine speed sensor 1
fx = (Fax + Ft ) , (12)
m
2.3 Aircraft Model - Kinematics 1
fy = F a y , (13)
m
Let vbg = (u, v, w) denote the decomposition of the ground 1
velocity vector vg , defined in the Earth-fixed North-East- f z = Fa z , (14)
m
Down (NED) frame, into the BODY frame, and let p, q, where m is the aircraft mass and Ft denotes the thrust
and r be the angular rates. This allows for the aircraft force, which is assumed
kinematics to be written as  to be aligned with the longitudinal
axis of the aircraft. Fax , Fay , Faz are aerodynamic forces
u rv + qw = ax , (1) represented as a vector decomposed in the BODY frame.
v pw + ru = ay , (2)
The thrust developed by the engine, as presented in Beard
w qu + pv = az , (3) and McLain (2012), is described by

542
 SAFEPROCESS 2015
September 2-4, 2015. Paris, France Kim Lynge Srensen et al. / IFAC-PapersOnLine 48-21 (2015) 541546 543

1   SATool presented in Blanke and Lorentzen (2006), as

Ft = Sp Cp kp2 p2 Va2 , (15)
2 {c1 , c2 }
where is the air density, Sp and Cp are propeller coef-  2  
r1 = y Sy5 + Sp Cp kp2 y32 y42 y1 , (20)
ficients. kp is a constant that specify the motor efficiency 2m 4
and p denotes the angular velocity of the propeller. 2
r2 = y Sy6 y2 , (21)
2m 4
Focus of the fault diagnosis approach presented in this pa-
per is confined to regard longitudinal and vertical change
detection. For that purpose the relevant aerodynamic 3.2 Simulator
forces can be described by
1 To asses the performance of the proposed icing detection
Fax = Va2 SCX (), (16) solution, numerical simulations have been conducted util-
2
1 ising Matlab and Simulink, with a sample time of 0.01s.
Faz = Va2 SCZ (), (17) The simulations have been based on the complete, 6-degree
2
of freedom Zagi model, of the small unmanned aircraft
where S is the wing surface area and 12 Va2 represents the system presented in Beard and McLain (2012), including
dynamic pressure. CX and CZ are composed of lift CL the autopilot module. Measurement noise is modelled as
and drag CD coefficients and depend on , as described in zero-mean white Gaussian noise, N (0, m, ), with stan-
Beard and McLain (2012) and Stevens and Lewis (2003), dard deviations aligned with the ones found in Langelaan
i.e. et al. (2011), i.e. m,ur = 0.1 [m/s] and m,fx , m,fz = 0.1
CX = CL () sin() CD () cos(), (18) [m/s2 ]. Wind is modelled as a constant wind field with
CZ = CL () cos() CD () sin(). (19) added turbulence, which is generated as white noise fil-
tered through a Dryden gust model, an approach presented
The aerodynamic coefficients, CX and CZ , are unknown by Langelaan et al. (2011) and utilised Beard and McLain
quantities. In Hansen and Blanke (2013) estimation is (2012). The Dryden transfer functions for the wind turbu-
achieved using a standard adaptive observer updating ap- lence are defined by
proach. However, within the scope of this paper nominal 
2Va 1
values of these estimates, i.e. no icing, are assumed known. Hu (s) = D,u , (22)
The estimated parameters are denoted CX and CZ , respec- L s + Va /Lu
 u
tively. 3Va s + Va /( 3Lv )
Hv (s) = D,v , (23)
L (s + Va /Lv )2
 v
3. FAULT DIAGNOSIS 3Va s + Va /( 3Lw )
Hw (s) = Dw , (24)
Lw (s + Va /Lw )2
The fault diagnosis objective is to detect and isolate the where D and L are the turbulence intensities and spa-
structural fault that occurs when icing forms on the lead- tial wavelength along the vehicle frames, respectively. For
ing edge of aircraft wings. The diagnosis is complicated by the simulations the Dryden model has been implemented
model uncertainties, which can be attributed to the non with a constant nominal airspeed Va = Va0 . The gust
measurable coefficients CX and CZ found in the expres- model used is for low altitude, moderate turbulence, with
sions for the longitudinal and lateral aerodynamic forces. numerical values for the turbulence parameters presented
The objective is achieved through residual generation and in Table 1
a NP change detection solution. The former is obtained
as presented in Blanke et al. (2006) and the latter as Table 1. Measurement noise levels and Dryden
presented in Kay (1998). gust parameters
altitude, 50 [m]
3.1 Structural Analysis
Lu , L v , 200 [m]
Lw , 50 [m]
Given the nonlinear model of the aircraft, described in the
previous section, the following constraints (or subsystems) D,u , D,v , 2.12 [m/s]
for the structural analysis can be formulated D,w , 1.4 [m/s]
1   V a0 , 14 [m/s]
c 1 : fx = Fax (Va , CX ) + FT (Va , p ) ,
m
1 The structural fault, icing on the leading edge of the
c2 : fz = Faz (Va , CZ ), wing entails, has been imposed upon the aircraft as a
m
10% increase in the drag coefficient and a 10% decrease
m1 : y 1 = fx , m2 : y2 = f z , m3 : y 3 = p , in the lift coefficient, which are well within the penalties
e1 : y4 = Va , e2 : y5 = CX , e3 : y6 = CZ , described in Lynch and Khodadoust (2001). For clarifying
purposes Figures 1 and 2 display the responses of the
aircraft system to icing. The former show responses in
Note that the constraints denoted e1 , e2 , and e3 are
altitude, airspeed, and pitch angle. The latter display
considered measured parameters.
autopilot responses, i.e. thrust and elevator displacements.
Residuals are identified using the methods described in The fault is imposed at t = 500s with ice forming over a
Blanke et al. (2006) and confirmed using the Matlab tool period lasting 25 seconds.

543
SAFEPROCESS 2015
544
September 2-4, 2015. Paris, France Kim Lynge Srensen et al. / IFAC-PapersOnLine 48-21 (2015) 541546

Fig. 1. Response to structural fault- 1. altitude, 2. airspeed,

3. pitch angle. Fig. 3. Residuals r1 and r2 , with distributions for the
faultless and faulty operations.

as the following detection problem.

H0 : x[n] = w[n] n = 0, 1, . . . N 1, (27)
H1 : x[n] = A + w[n] n = 0, 1, . . . N 1, (28)
where A is unknown and w[n] is white Gaussian noise
with unknown variance 2 . N is the window size. The
H0 hypothesis describes the case where the signal contain
the expected noise only, whereas the alternative hypoth-
esis H1 , other than containing the expected noise, also
contain an offset A from zero. If an offset is identified it
implies a significant difference between the model and the
measurement, hence a fault is concluded to be present. In
Figures 4 and 5 the probability plot and autocorrelation
of the two relevant residuals are displayed. As seen in
the figures the residuals contain uncorrelated samples.

Fig. 2. Autopilot response to structural fault- 1. thrust

displacement, 2. elevator displacement.
3.3 Change Detection

Fault detection is in this paper achieved by detecting

changes in the residual signals (20) and (21). The proposed
detection solution relies on CX and CZ under nominal
flight conditions. When CX and CZ display unexpected
changes, a bias is introduced into the residuals and a
structural fault (icing on the leading edge of the airfoil) is
said to have occurred. Note that r1 and r2 can be presented
as
2  
r1 = Va S CX CX , (25)
2m  
2
r2 = V S CZ CZ , (26)
2m a
where C C = 0 whenever icing is forming on the leading
edge of the airfoil. Fig. 4. Residual r1 data distribution fit and autocorrela-
tion.
The response of the two residuals and the distributions for
faultless and faulty evolutions, are displayed in Figure 3. The probability plots show that the residuals follows a
A Generalised Likelihood Ratio Test (GLRT) is used Gaussian distribution with a general form
to distinguish between two hypotheses stated about the 1 (x)2
p(x; , ) = e 22 . (29)
residual signal. The problem is mathematically expressed 2

544
 SAFEPROCESS 2015
September 2-4, 2015. Paris, France Kim Lynge Srensen et al. / IFAC-PapersOnLine 48-21 (2015) 541546 545

where P () is the cumulative distribution function of a

given test statistics distribution.

4. PERFORMANCE ASSESSMENT

The GLRT performance is dependant on the trade-off

between the desire for a high PD , a low PF A , GLRT
window size N , and the time it takes to detect the
occurrence of a fault.
The test statistics, from r1 denoted Tr1 (x), approximately
follow a chi-squared (X2 ) distribution under both hypothe-
ses, with serving as the degree-of-freedom parameter. For
window sizes N = 1000, 2000, corresponding to 10 and 20
seconds, a visual representation can be found in Figure 6,
which also includes thresholds for PF A = 106 . Here it
should be mentioned that parameters of the X2 , used to
fit the data, were estimated utilising MLE.

Fig. 5. Residual r2 data distribution fit and autocorrela-

tion.
The GLRT is based on the likelihood ratio between the
probability of the two hypotheses given a window of data.
The GLRT decides H1 if

p(x; A, 12 , H1 )
LG (x) = > , (30)
02 , H0 )
p(x;
where [A 12 ]T is the maximum likelihood estimate (MLE)
of the vector [A 12 ] under H1 and 02 is the MLE of 02
under H0 . The decision function (or threshold) is denoted

. A, 12 , and 02 are determined by maximising (29) with
= A and = A = 0, respectively. For the MLEs under
H1 this results in
N 1
1  2
12 =
(x[n] x ) , (31)
N n=0
with x
being the sample mean of x[n] and where x For
= A. Fig. 6. Probability plot of T (x), for r1 , under H0 and H1
the MLE under H0 maximising (29), assuming = A = 0, for window length N = 1000 and N = 2000.
leads to
N 1 For r2 , the test statistics, denoted Tr2 (x), also approxi-
1 
02 =
x[n]2 . (32) mately follow a X2 distribution under both hypotheses.
N n=0 Choosing window sizes N = 500, 1000 the probability
With this the following test statistic can be derived (see characteristics of Tr2 (x) are displayed in Figure 7. The
Kay (1998)) thresholds seen in the figure are for PF A = 106 . The
 2

0 performance of the GLRT, for both residuals, are found in
T (x) = N ln . (33) Table 2.
12

Table 2. GLRT performance
A theoretical threshold is determined according to the
Neyman Pearson theorem found in Kay (1998). Given Tr1 (x) Tr2 (x)
a signal f (t) that behaves according to the probability N 1000 2000 500 1000
density function p(f (t); H0 ) under H0 , the threshold that
maximises the probability of detection PD is found from PF A 106 106 106 106

PD 58.89% 99.99% 56.97% 96.46%
PF A = p(f ; H0 )df, (34)
{f :LG (f )>}
where PF A is the desired probaility of false alarm. For Note that wind turbulence levels significantly influence the
the modified GLRT, here denoted T (x), an asymptotic evolution of the residuals, i.e. increased turbulence will
result exists for large data records (N ) Kay (1998), result in a decrease in PD , but this is easily accommo-
Galeazzi et al. (2013). dated by increasing the window size N . It is, however,
the accelerometer measurement noise that is the primary
The probability of detecting a fault under H1 , with prob- contaminant. Consequently a white noise model could have
ability PD for a given threshold is given by limitations, as accelerometer noise might include issues
PD = 1 P (; H1 , , ), (35) such as bias, drift, vibrations, etc. that are correlated.

545
SAFEPROCESS 2015
546
September 2-4, 2015. Paris, France Kim Lynge Srensen et al. / IFAC-PapersOnLine 48-21 (2015) 541546
Fig. 7. Probability plot of T (x), for r2 , under H0 and H1
for window length N = 500 and N = 1000.
Correlated noise can be addressed by pre-whitening. The
issue of correlated noise and solutions is a subject deemed
outside the scope of the work presented here.
5. CONCLUSION
The work presented in this paper proposes a Neyman
Pearson based statistical change detection approach, for
the identification of structural changes of fixed wing UAV
airfoils. The solution employs structural analysis to iden-
tify residuals, which are processed by a generalised likeli-
hood ratio test, utilised for hypotheses tests on potential
changes and severity assessments, supplying estimates of
the magnitude of change.
Simulations were conducted that show the expected per-
formance under both nominal and icing influenced con-
ditions. The proposed fault diagnosis solution ensures a
high level of fault detection in spite of turbulent winds,
measurement noise and model uncertainties. Thresholds
were obtained ensuring a low probability of false alarms
and achieving a high probability of detection. The detec-
tion was analysed using simulation generated data where
a fault was imposed on the system.
REFERENCES
Bateman, F., Noura, H., and Ouladsine, M. (2011). Fault
diagnosis and fault-tolerant control strategy for the
aerosonde uav. Aerospace and Electronic Systems, IEEE
Transactions on, 47(3), 21192137.
Beard, R.W. and McLain, T.W. (2012). Small Unmanned
Aircraft - Theory and Practice. Princeton University
Press.
Blanke, M. and Hansen, S. (2013). Towards self-tuning
residual generators for uav control surface fault diag-
nosis. In Control and Fault-Tolerant Systems (SysTol),
2013 Conference on, 3742.
Blanke, M., Kinnaert, M., Lunze, J., and Staroswiecki, M.
(2006). Diagnosis and Fault-Tolerant Control. Springer
Berlin Heidelberg.
Blanke, M. and Lorentzen, T. (2006). Satool - a soft-
ware tool for structural analysis of complex automation
546
systems. In Proceedings of the 6. IFAC Symposium
on Fault Detection, Supervision and Safety of Technical
Processes, 673678. Elsevier Science.
Cristafaro, A. and Johansen, T.A. (2015). An unknown in-
put observer approach to icing detection for unmanned
aerial vehicles. In American Control Conference.
Cristafaro, A., Johansen, T.A., and Aguiar, A.P. (2015).
Icing detection and identification for unmanned aerial
vehicles: Multiple model adaptive estimation. In Euro-
pean Control Conference.
Ducard, G. (2009). Fault-tolerant Flight Control and
Guidance Systems. Springer London.
Ducard, G. and Geering, H.P. (2008). Efficient nonlinear
actuator fault detection and isolation system for un-
manned aerial vehicles. Journal of Guidance, Control,
and Dynamics, (1), 225237.
Fravolini, M.L., Brunori, V., Campa, G., Napolitano,
M.R., and La Cava, M. (2009). Structural analysis
approach for the generation of structured residuals for
aircraft fdi. Aerospace and Electronic Systems, IEEE
Transactions on, 45(4), 14661482.
Galeazzi, R., Blanke, M., and Poulsen, N. (2013). Early
detection of parametric roll resonance on container
ships. Control Systems Technology, IEEE Transactions
on, 21(2), 489503.
Hansen, S. and Blanke, M. (2013). Diagnosis of airspeed
measurement faults for unmanned aerial vehicles. IEEE
Transactions of Aerospace and Electronic Systems.
Johansen, T.A., Cristafaro, A., Srensen, K.L., Hansen,
J.M., and Fossen, T.I. (2015). On estimation of wind
velocity, angle-of-attack and sideslip angle of small uavs
using standard sensors. In International Conference on
Unmanned Aircraft Systems.
Kay, S.M. (1998). Fundamentals of Statistical Signal Pro-
cessing: Detection theory. Prentice Hall Signal Process-
ing Series. Prentice-Hall PTR.
Langelaan, J.W., Alley, N., and Neidhoefer, J. (2011).
Wind field estimation for small unmanned aerial ve-
hicles. Journal of Guidance, Control, and Dynamics,
34(4), 10161030.
Long, H. and Song, S. (2009). Method of estimating
angle-of-attack and sideslip angel based on data fusion.
In Intelligent Computation Technology and Automation,
2009. ICICTA 09. Second International Conference on,
volume 1, 641644. doi:10.1109/ICICTA.2009.160.
Lynch, F.T. and Khodadoust, A. (2001). Effects of
ice accretions on aircraft aerodynamics. Progress in
Aerospace Sciences, 37(8), 669 767.
Stevens, B. and Lewis, F. (2003). Aircraft Control and
Simulation. Wiley.
Tousi, M. and Khorasani, K. (2009). Fault diagnosis and
recovery from structural failures (icing) in unmanned
aerial vehicles. In Systems Conference, 2009 3rd Annual
IEEE, 302307.
Tousi, M. and Khorasani, K. (2011). Robust observer-
based fault diagnosis for an unmanned aerial vehicle.
In Systems Conference (SysCon), 2011 IEEE Interna-
tional, 428434.