You are on page 1of 6

11th

11th IFAC
IFAC Symposium
Symposium on on Nonlinear
Nonlinear Control
Control Systems
Systems
Vienna,
11th
Vienna, Austria,
IFAC Sept.
Sept. 4-6,
Symposium
Austria, 2019
on Nonlinear
4-6, 2019 Control Systems
Available online at www.sciencedirect.com
11th IFAC
Vienna, Symposium
Austria, on Nonlinear
Sept. 4-6, 2019 Control Systems
Vienna, Austria, Sept. 4-6, 2019
ScienceDirect
IFAC PapersOnLine 52-16 (2019) 831–836
Differential
Differential Flatness-Based
Flatness-Based ApproachApproach for for
Differential
Sensors and Flatness-Based
Actuators Fault Approach
Diagnosis for
of
Differential
Sensors Flatness-Based
and Actuators Fault Approach
Diagnosis forof aa
Sensors and Actuators
Multirotor Fault
UAV Diagnosis of a
Sensors and Actuators
Multirotor
Multirotor UAVFault
UAV Diagnosis of a
Multirotor
M. Saied ∗∗∗ T. Mahairy ∗ UAV

∗ C. Francis ∗


∗ H. Shraim ∗

∗ H. Mazeh ∗

M. Saied ∗∗ T. Mahairy ∗∗ C. Francis ∗∗∗ H. Shraim ∗∗ H. Mazeh ∗∗
M. Saied T. Mahairy M. M.C. El Rafei
Francis
El Rafei∗∗∗∗∗ H. Shraim ∗ H. Mazeh ∗
M. Saied ∗ T. Mahairy ∗M. C. ElFrancis
Rafei H. Shraim H. Mazeh


∗ Research Center
∗ Research Center in Engineering,Rafei
M.
in Engineering, El Faculty of
Faculty of Engineering,
Engineering, Lebanese Lebanese


Research University,
Center inHadath
University, Hadath (e-mail:
Engineering,
(e-mail: cfrancis,
Faculty
cfrancis, of hassan.shraim,
Engineering, Lebanese
hassan.shraim,

Research Center inHadath
University, Engineering,
(e-mail:Faculty
maher.elrafei@ul.edu.lb).
cfrancis,
maher.elrafei@ul.edu.lb). of hassan.shraim,
Engineering, Lebanese
University, Hadath (e-mail: cfrancis, hassan.shraim,
maher.elrafei@ul.edu.lb).
maher.elrafei@ul.edu.lb).
Abstract:
Abstract: This This paper
paper deals
deals withwith sensors
sensors and and actuators
actuators fault
fault detection
detection in in multirotor
multirotor unmannedunmanned
Abstract:
aerial vehiclesThis paper
whose deals
dynamics witharesensors
modeled and actuators
as nonlinear
aerial vehicles whose dynamics are modeled as nonlinear systems. The considered fault detection
systems. The in multirotorapproach
considered unmanned
approach is
is
Abstract:
based
aerial on
vehiclesThis
the paper
differential
whose deals
dynamics with
flatness aresensors
property
modeled andof actuators
flat
as systems
nonlinear fault detection
that
systems.
based on the differential flatness property of flat systems that provide analytical redundancy provideThe in multirotor
analytical
considered unmanned
redundancy
approach is
aerial
where
based vehicles
every
on the whose
control
differential dynamics
input and
flatness are
system modeled
propertystate of as
can nonlinear
be
flat expressed
systems systems.
where every control input and system state can be expressed as a function of the flat outputs.that as a The
function
provide considered
of
analyticalthe approach
flat outputs.
redundancy is
based
By
where on
comparing
every the differential
the
control real
input flatness
measurements
and property
system with
state of
canflat
their be systems
corresponding
expressed
By comparing the real measurements with their corresponding estimations deduced from the that as provide
estimations
a function analytical
ofdeduced
the redundancy
flat from
outputs.the
where every control
By comparing
differentially
differentially flat
flat input
theequations,
real
equations, andresidual
measurementssystem signals
residual state can
with their
signals arebecorresponding
are expressedand
computed
computed as estimations
and ausedfunction
used for ofdeduced
for fault
fault thediagnosis.
flat from
diagnosis.outputs.the
For
For
By comparing
multirotor
differentially UAV the
flat real
systems, measurements
equations, only one set
residual ofwith
flat
signals their
outputs
are corresponding
can
computed be found.
and
multirotor UAV systems, only one set of flat outputs can be found. In this paper, it is shown that estimations
In
used thisforpaper,deduced
fault it is from
shown
diagnosis. the
that
For
differentially
this unique
multirotor UAV flat
set can equations,
be
systems, used only toresidual
detect
one set signals
and
of isolate
flat are
outputs computed
bothcan sensors
be and
and
found.
this unique set can be used to detect and isolate both sensors and actuators faults and failures. used for
actuators
In this fault
faults
paper, itdiagnosis.
and
is shown For
failures.
that
multirotor
The
this unique
The UAV
effectiveness
set can
effectiveness systems,
of
ofbethis
thisused only to one
approach
approachdetectset
is of flat
is illustrated
and outputs
isolate
illustrated by
both
by can be found.
numerical
sensors
numerical and In
simulations this paper,
actuators
simulations on it isandshown
Matlab/Simulink,
onfaults
Matlab/Simulink, that
failures.
this
The unique
and also
and also a set
effectiveness
a real can of be
this
real experimental used to
approach
experimental application detect is and isolate
illustrated
application on both
by
on aa hexarotor sensors
numerical
hexarotor UAV. UAV. and actuators
simulations on faults and
Matlab/Simulink, failures.
The also
and effectiveness of this approach
a real experimental is illustrated
application by numerical
on a hexarotor UAV.simulations on Matlab/Simulink,
© 2019,
and also IFAC
a real(International
experimental Federation of Automatic
application on a Control) Hosting
hexarotor UAV. by Elsevier Ltd. All rights reserved.
Keywords:
Keywords: Fault Fault Detection,
Detection, Flat Flat Systems,
Systems, Unmanned
Unmanned Aerial Aerial Vehicles,
Vehicles, Hexarotor.
Hexarotor.
Keywords: Fault Detection, Flat Systems, Unmanned Aerial Vehicles, Hexarotor.
Keywords: Fault Detection, Flat Systems, Unmanned
1. INTRODUCTION
INTRODUCTION Aerial Vehicles, Hexarotor.
Differential
1. Differential flatness-based
flatness-based fault fault diagnosis
diagnosis for for multirotor
multirotor
1. INTRODUCTION systems
Differential was first presented
flatness-based in
fault (Nan et
diagnosis al.for(2008)). Dif-
multirotor
1. INTRODUCTION systems
Differential
systems was first presented
flatness-based in
fault (Nan et
diagnosis al.
al.for(2008)). Dif-
multirotor
In recent years, methods for sensors and actuators fault ferential was
ferential flatness
flatnessfirstis ispresented
a property
a property inof
of(Nan
someetnonlinear
some nonlinear(2008)). Dif-
dynamic
dynamic
In
In recent
recent of years,
years, methods
methods for sensors
for sensors and actuators
andvehicles
actuators fault systems
ferential was
that
flatnessfirst
aims is presented
toa study
property if in
a (Nan
system
of some et
of al. (2008)).
differential
nonlinear Dif-
equa-
dynamic
diagnosis
diagnosis of multirotor
multirotor unmanned
unmanned aerial
aerial vehicles havefault
have ex- systems that aims to study if a system of differential equa-
ex-
In recent
diagnosis years,
of
perienced increasing methods
multirotor
increasing attention. for
unmanned sensors
attention. This aerial
This is and
is due actuators
vehicles
due toto the have
the perspec- fault
ex- ferential
tions
systems flatness
could
that beaims istoa property
parameterized
study if
perspec- tions could be parameterized by a set of variables called a of
by some
a
system set nonlinear
of
of variables
differential dynamic
called
equa-
perienced
diagnosis
perienced
tives
tives of of multirotor
increasing
of integration
integration unmanned
ofattention.
of these
these Thisaerial
vehicles
vehicles is due
into
into vehicles
the
the to
airthe
air have ex- systems
perspec-
transporta-
transporta-
flat
flat outputs
tions thatbe
could
outputs aims
and
and to study
aa finite
parameterized
finite if aby
number
number system
of setofof
ofatheir
their differential
derivatives
variables (Fliess
derivatives equa-
called
(Fliess
perienced
tives
tion increasing
of integration
system, whereofthey attention.
theseshould
they This
vehicles is due
into
imperativelythe to
airthe perspec-
transporta-
meet flight et tions
flat
et al. could
outputs
al. (1995)). be
and
(1995)). The parameterized
a finite
The analytical number by
analytical redundancy of a set
their of
redundancy obtained variables
derivatives
obtained thanks called
(Fliess
thanks
tion
tives system,
tion where
of integration
system, where ofthey
these should
vehicles
should imperatively
into the
imperatively meet
airlike flight
flight flat
transporta-
meet to
et al.outputs
the main
(1995)). and
property
The a finite of number
the
analytical of their derivatives
differentially
redundancy flat
obtainedsystems (Fliess
can
thanks
conditions
conditions and
and comply
comply with
with air
air traffic
traffic rules
rules like manned
manned to the main property of the differentially flat systems can
tion system,
conditions
aircrafts. andwhere
comply they should
with air imperatively
traffic rules meet
like flight
manned et
be
to al.
the (1995)).
exploited
main to The analytical
generate
property of residual
the redundancy
signals
differentially
be exploited to generate residual signals for fault diagnosis obtained
for fault
flat systems thanks
diagnosis
can
aircrafts.
conditions and comply with air traffic rules like manned to
aircrafts. be the
purpose. main
exploited property
to
Residues generateare of the differentially
residual
computed
purpose. Residues are computed by comparing the mea- signals
by flat
for
comparing faultsystems can
diagnosis
the mea-
A multirotor unmanned aerial vehicle is a typical nonlinear be exploited
purpose.
sured to
Residues generateare residual
computed signals
by for
comparing fault diagnosis
the mea-
aircrafts.
A
A multirotor
multirotor
system. Its
unmanned
unmanned aerial
aerial vehicle
vehicle is a
a typical
is using
typical nonlinear
nonlinear sured variables with the estimated ones deduced from the
variables with the estimated ones deduced from the
system. Its dynamics
dynamics can
can be
be modeled
modeled using Lagrange
Lagrange or
or purpose.
differentially
sured Residues
variables
differentially flat
flatwith are
thecomputed
equations.
equations. The
estimated
The
by comparing
approaches
ones
approachesdeduced the
frommea-
proposed
proposed in
the
in
A multirotor
system.
Newton-Euler unmanned
Its dynamics
formalism aerial
can vehicle is using
be modeled
(Bouabdallah a(2007)).
typical nonlinear
Lagrange
Recently, or suredthe variables
literature
differentially aim
flatwith to the
find
equations. estimated
two or
The more ones
sets
approachesdeduced
of flat from
outputs
proposed the
to
in
Newton-Euler
system. formalism
Its dynamics (Bouabdallah
canfor bestate
modeled (2007)). Recently, the literature aim to find two or more sets of flat outputs to
Newton-Euler
the
the techniques
techniques
formalism
proposed
proposed
(Bouabdallah
for state and using
and
(2007)).
parameter
parameter
Lagrange
Recently,
estima-
estima-
or differentially
the literature
increase the flat
aim
number equations.
to find
of two
decoupledThe
or more approaches
sets
residues of flat
which proposed
outputs
increasesin
to
Newton-Euler
the
tion techniques
of nonlinear formalism
proposed
systems (Bouabdallah
for
have state
contributed (2007)).
and parameterto the estima- increase
Recently,
develop- increase
the
the number of decoupled residues which increases
literature
the
possibility aim
number
to to find
isolateof two
decoupled
every or more
single sets
residues
fault. of flat
which
For a outputs
increases
multirotor to
tion of nonlinearproposed
the techniques
tion ofof systems have
for
havestatecontributed
and parameterto
to the
the develop-
estima- the possibility to isolate every single fault. For a multirotor
ment
ment ofnonlinear
purely systems fault
purely nonlinear
nonlinear fault contributed
diagnosis
diagnosis approaches.
approaches. develop-
With
With increasesystem,
the the
it
it was
possibility number
toshown
isolateof decoupled
that
everyonly single residues
one set
fault. which
of
of flat
set For increases
outputs
a multirotor
tion ofofnonlinear
ment
applicationpurely systems fault
nonlinear
to multirotor
multirotor have
UAVs, contributed
diagnosis to the
sliding approaches.
mode With system,
develop-
techniques
techniques the
can possibility
be
system, found
it
was
was to
in
shown
isolate
(Torres
shown
that
every
only
(2013)),
that single
only
one
and
onefault.
only
set For
of
flat
a
additive
outputs
multirotor
flat faults
outputs
application
ment used
of purely
application to nonlinearetfaultUAVs, sliding
diagnosis mode
approaches. With system, can be found in (Torres (2013)), and only additive faults
were
were used into(Chandra
in
multirotor
(Chandra et
UAVs,
al. (2015),
al. (2015),slidingSaied
Saied
mode
et al.
et
techniques
al. (2015a),
(2015a), can be
affecting it
found was
sensors in shown
(Torres
and that only
(2013)),
control inputs one
and set
only
were of flat
additive
considered. outputs
faults
The
application
were
Saied used
et al.into multirotor
(Chandra
(2015b), et
Avram UAVs,
al. (2015),
et al. slidingSaied
(2017)) mode
et
for techniques
al. (2015a),
sensors and affecting sensors and control inputs were considered. The
Saied et al. (2015b), Avram et al. (2017)) for sensors and can
work
work be
affectingin
infound
our
sensors
our in
paper
paper (Torres
and is
is (2013)),
based
control
based on
inputs
on and
the
the only
results
were additive
presented
considered.
results presented faults
in
The
in
were
Saied used
et
actuators al.in
fault(Chandra
(2015b),
diagnosis et
Avramin al. (2015),
et al.
quadrotors Saied
(2017))and et
for
actuators fault diagnosis in quadrotors and octorotors. The affecting al. (2015a),
sensors
octorotors. and
The (Torres
work in sensors
(2013))
our to
paper and
show
iscontrol
that
based itinputs
is
on were
possible
the considered.
to
results isolate
presented The
sensors
in
Saied et al.
actuators
performance (2015b),
faultofdiagnosis
slidingAvramin
mode et observer-based
al. (2017))
quadrotors for sensors
and octorotors.
techniques and work
The (Torres (2013))
in(2013))
our paper to show that
is based it is
on possible
the to
results isolate
presentedsensors
in
performance of sliding mode observer-based techniques and
(Torresactuators failures
to show using
that only
it is one
possible settoof flat
isolate outputs.
sensors
actuators
was shownfault
performance
was shown to of
to
diagnosis
be
be
sliding
robust to
robust
in
mode
to quadrotors
parameter
parameter
and octorotors.
observer-based
uncertainties
uncertainties in the and
inThe
techniques
the (Torres
and actuators
(2013))
The actuators
main
failures
to showusing
failures
contribution that
using
of
only
theitonly
one
is possible
paper one set of flat
flat outputs.
settoproposition
is the
the ofisolate
proposition sensors
outputs.
of a
a
performance
was shown
system model to ofbe sliding
and robust mode
to
disturbances. observer-based
parameter
Thau uncertainties
observer-basedtechniques
inFDI
the and The main contribution of the paper is of
system model and disturbances. Thau observer-based FDI fault
The
fault actuators
diagnosis
main failures
approach
contribution
diagnosis approach using
offor
for only
paperone
theactuators
actuators set proposition
and
is the
and of flat outputs.
sensors
sensors failures
of a
failures
was shown
system
approachesmodel to
werebe proposed
and robust to in
disturbances. parameter
(CenThau et uncertainties
observer-based
et al. (2013),
(2013), Cen inFDI
the The
and main contribution
of aa multirotor UAV. offor
the paper is the proposition of a
approaches
system (2013))
approaches
Noura model
were
wereand
to
proposed
disturbances.
proposed
estimate
in (Cen
inthe
(CenThau
al.
observer-based
et al.
quadrotor (2013),
system
Cen
Cen
and
and of
FDI
states
fault diagnosis
multirotor UAV. Most
approach Most of
of the
the papers
actuators papers andin
in the
the literature
sensors failures
literature
Noura (2013)) to estimate the quadrotor system states fault
consider
of a diagnosis
only
multirotor approach
oneUAV. type of
Most for actuators
failures:
of the sensors
papers and insensors
or the failures
actuators.
literature
approaches
Noura
and build awere
(2013)) settoofproposed
ofestimate inthe
(Cen
offset residuals
residuals ettoal.
quadrotor (2013),
system
indicate Cen
actuators and consider
states of
only one type of failures: sensors or actuators.
a multirotor
consider only one UAV. typeMost of the papers
of failures: sensors in orthe literature
actuators.
and build a set offset to indicate actuators The paper is
Noura
and
faults. An
faults. (2013))
build
An LPVa set to ofestimate
offset
LPV model-based the
residuals
model-based sensor quadrotor
to
sensor fault system
indicate
fault diagnosis states
actuators
diagnosis for for aa consider
The paper onlyis organized
organized
one type of as follows:
asfailures: Section
follows: sensors
Section or II presents
presents the
II actuators. the
and build
faults. An aUAV
LPVset model-based
of offset residuals
sensor tofault
indicate
diagnosisactuators
foral. dynamics
The paperof ofisthe hexarotor
organized UAV.
UAV. Section
as follows: SectionIIIIIIIIis devoted
ispresents
devotedthe to
quadrotor
quadrotor UAV was
was presented
presented in
in (Lopez-Estrada
(Lopez-Estrada et
et al.a dynamics the hexarotor Section to
faults. An
quadrotor
(2015)). In LPV
UAV
order model-based
was
to presented
detect and sensor fault
in (Lopez-Estrada
isolate diagnosis
sensor faults, for
et
a al.a The
set the paperofisthe
dynamics
the flatness-based
flatness-based organized fault as
hexarotor
fault follows:
UAV. Section
detection
detection Section
and
and IIIIIispresents
isolation
isolation devoted
of sensors
of the
to
sensors
(2015)).
quadrotor
(2015)). In order
UAV
In order to
was detect
presented
to detectwith and isolate
in
and aisolate sensor
(Lopez-Estrada
sensor faults,
faults, a
et set
set and actuators failures. The results are then validated to
al. dynamics
and
the of
actuators
flatness-basedthe hexarotor
failures.
fault The UAV.results
detection Section
and are III
then
isolation is devoted
validated
of in
sensors
of residuals
of residuals was generated
was generated with a bank of
bank of observers
observers soathat
so that in
(2015)).
of residualsIn
each residual order
was
residual is to detect
generated
is sensitive
sensitive to and
with
to only isolate
a
only onebank
one fault. sensor
of
fault. faults,
observers so a set
that the
and flatness-based
simulations
actuators and in fault
failures. real detection
experiments
The results
simulations and in real experiments in section IV. The and are isolation
in section
then of sensors
IV.
validated The
in
each
of residuals
each residual wasis generated
sensitive towith onlya one
bankfault.of observers so that and paper actuators
simulations
ends and
with failures.
in
a real The
conclusion results
experiments
and
paper ends with a conclusion and perspectives on future arein then
section
perspectives validated
onIV. in
The
future
each simulations
paper
works. ends and
with in
a real experiments
conclusion and in section
perspectives onIV. The
future
 Thisresidual
 This work has is
work has sensitive
been
been funded with
funded
to only
with support
support
onefromfault.
from thethe research
research support
support works.

program of the Lebanese University.
paper
works. ends with a conclusion and perspectives on future
This work has been funded
program of the Lebanese University. with support from the research support

program of the
This work hasLebanese
been funded University.
with support from the research support
works.
program of the Lebanese University.
program
2405-8963of©
©the Lebanese
2019, University. Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
IFAC (International
Copyright
Copyright 2019
2019 IFAC
© under IFAC 1481
1481Control.
Peer review
Copyright © 2019 IFAC responsibility of International Federation of Automatic
1481
10.1016/j.ifacol.2019.12.066
Copyright © 2019 IFAC 1481
2019 IFAC NOLCOS
832
Vienna, Austria, Sept. 4-6, 2019 M Saied et al. / IFAC PapersOnLine 52-16 (2019) 831–836

3. FLATNESS-BASED FAULT DIAGNOSIS

A dynamical system,
ẋ = f (x, u); y = h(x) (5)

with x ∈ Rn and u ∈ Rm , is flat if and only if there exists


a flat output vector z ∈ Rm such that:
• The flat output vector is expressed as function of the
state x and the control input u and a finite number
of its time derivatives:
z = Φz (x, u, u̇, ..., u(p) ) (6)
Fig. 1. The PPNNPN hexarotor UAV configuration and • The state x and the control input u are expressed as
reference frames functions of the vector z and a finite number of its
time derivatives:
2. MODELING OF THE HEXAROTOR UAV
x = Φx (z, ż, ..., z (q) )
(7)
The equations governing the motion of the hexarotor UAV u = Φu (z, ż, ..., z (q+1) )
shown in Fig. 1 are derived in this section, where the
commonly employed multirotor UAV model (Chamseddine where z (q) denotes the q th time derivative of z.
et al. (2012)) is used:
uf k1 3.1 Flat Model
ẍ = (cosφsinθcosψ + sinφsinψ) ∗ − ẋ
m m
uf k2 The nonlinear model of the hexarotor presented in the
ÿ = (cosφsinθsinψ − sinφcosψ) ∗ − ẏ
m m previous section consists of 12 states:
uf k3
z̈ = −g + (cosφcosθ) ∗ − ż X = [ x y z φ θ ψ ẋ ẏ ż φ̇ θ̇ ψ̇ ]T = [x1 x2 ... x12 ]T
m m and the control inputs are u = [ uf τ˜φ τ˜θ τ˜ψ ] =
Iyy − Izz Jr τφ k4 (1)
φ̈ = τ˜φ = θ̇ψ̇ − θ̇Ω + −l φ̇ [ u1 u2 u3 u4 ]. With negligible drag coefficients at low
Ixx Ixx Ixx Ixx
Izz − Ixx Jr τθ k5 speeds (Lai et al. (2006)), the system is rewritten as:
θ̈ = τ˜θ = φ̇ψ̇ + φ̇Ω + −l θ̇ 
Iyy Ixx Iyy Iyy  ẍ = (cosφsinθcosψ + sinφsinψ) ∗ u1 /m
Ixx − Iyy τψ k6 
 ÿ = (cosφsinθsinψ − sinφcosψ) ∗ u1 /m
ψ̈ = τ˜ψ = φ̇θ̇ + − ψ̇ 

Izz Izz Izz  z̈ = −g + (cosφcosθ) ∗ u1 /m
φ̈ = u2 (8)


where x, y and z are the coordinates of the hexarotor 

center of gravity in the earth-fixed frame. φ, θ and ψ are  θ̈ = u3

respectively the roll, pitch and yaw Euler angles. m is the ψ̈ = u4
mass of the vehicle, Ixx , Iyy and Izz are the moments of
inertia along the x, y and z directions. ki (i = 1...6) are the Since there are four control inputs then there are four flat
drag coefficients and l is the arm length. Ω is a disturbance outputs in each set of flat outputs. Observing the above
depending on the motors speeds: Ω = ω2 + ω3 + ω5 − ω1 − equations one can notice that we can find only one set
ω4 − ω6 . of flat outputs za = [x y z ψ] (Torres (2013)). The goal
is to express all the states and all the control inputs as
uf is the total thrust applied to the hexarotor UAV in functions of these flat outputs and a finite number of their
the z-direction of the body-fixed frame. τφ , τθ and τψ are time derivatives. Let us take:
respectively the roll, pitch and yaw torques in the φ, θ and
ψ directions. They are given as follows: z1 = x; z2 = y; z3 = z; z4 = ψ; (9)
 
  T1 + T2 + T3 + T4 + T5 + T6 The parametrization of φ and θ in function of the flat
uf  l 
 2T6 ) ∗  outputs is:
 τφ   (−T1 + T2 + 2T3 + T4 − T5 − √ 2
τ = 3l
 (2) cos(z4 )z¨1 + sin(z4 )z¨2
θ   θe = atan( )
τψ  (T 1 + T 2 − T 4 − T 5 ) ∗ 
2 z¨3 + g (10)
−Q1 + Q2 + Q3 − Q4 + Q5 − Q6 cos(θe )(sin(z4 )z¨1 − cos(z4 )z¨2 )
φe = atan( )
z¨3 + g
The generated thrust from a rotor i having an angular
velocity ωi is: For simplification, we assume that the system is in a
Ti = Kf ωi2 (3) hovering state with a small yaw angle ψ, and φ and θ
are re-written as:
The aerodynamic moment Qi of a propeller i is: z¨1
θe = atan( )
Qi = Kt ωi2 (4) z¨3 + g
−cos(θe )z¨2 (11)
Kf and Kt are the thrust and drag coefficients. φe = atan( )
z¨3 + g

1482
2019 IFAC NOLCOS

Vienna, Austria, Sept. 4-6, 2019 M Saied et al. / IFAC PapersOnLine 52-16 (2019) 831–836 833

The first order derivatives θ̇ and φ̇ are estimated as: outputs will not trigger all the residues but only the ones
(3) (3)
depending on the measured outputs. For example a fault
−z2 (z¨3 + g) + z¨2 z3 on the sensor measuring the displacement x will affect the
φ̇e =
(z¨3 + g)2 (1 + z¨3z¨+g
2
)2 residues R2x and R3u . A fault on the sensor measuring
(12) the displacement y will affect the residues R1x and R2u .
(3) (3) A fault on the sensor measuring the displacement z will
z (z¨3 + g) − z¨1 z3
θ̇e = 1 trigger the residue R1u . R1x , R2x , R2u and R3u will be
(z¨3 + g)2 (1 + z¨3z¨+g
1
)2 also slightly affected since they depend on the derivatives
of z, however their amplitude will not be large enough to
The parametrization of the control inputs in function of exceed the thresholds. This is indicated by 0∗ in Table
the flat outputs is: 1. The last flat output is ψ, and a fault on the sensor
m(z¨3 + g) measuring the yaw angle will affect only the residue R4u .
u1 = ; u2 = φ¨e Regarding the faults on the sensors measuring φ and θ,
cos(θe )cos(φe ) (13)
the affected residues will be respectively R1x , R1u , R2u
u3 = θ¨e ; u4 = ψ¨e and R2x , R1u , R3u .
where φ¨e and θ¨e are functions of the flat outputs and can The table below presents the residue matrix for isolation
be derived from the expressions of φ˙e and θ˙e presented in of sensors faults and shows that any single sensor fault can
Eq. (12): be easily isolated.
(3) (3)
α(1 + ( z¨3z¨+g
2
)2 ) − 2( z¨−3 +g
z¨2
)(
−z2 (z¨3 +g)+z¨2 z3
)2 Fault R1x R2x R1u R2u R3u R4u
(z¨3 +g)2
φ̈e = Fx 0 1 0 0 1 0
(1 + ( z¨3z¨+g
2
)2 )2 Fy 1 0 0 1 0 0
(14)
β(1 + ( z¨3z¨+g
1
)2 ) − 2( z¨3z¨+g
1
)(
(3)
z1 (z¨3 +g)−z¨1 z3
(3)
)2 Fz 0∗ 0∗ 1 0 0 0
(z¨3 +g)2 Fφ 1 0 1 1 0 0
θ̈e =
(1 + ( z¨3z¨+g
1 2
) ) 2
Fθ 0 1 1 0 1 0
Fψ 0 0 0 0 0 1
with (4) (4) (3) (3)
(−z2 (z¨3 +g)+z¨2 z3 )(z¨3 +g)2 −(−z2 (z¨3 +g)+z¨2 z33 )(2(z3¨+g)z3 ) Table 1: The residue matrix for isolation of sensors faults
α= (z¨3 +g)4
(4) (4) (3) (3) (3)
(z1 (z¨3 +g)−z¨1 z3 )(z¨3 +g)2 −(2z3 (z¨3 +g))(z1 (z¨3 +g)−z¨1 z3 )
β= (z¨3 +g)4
3.4 Detection and Isolation on Actuators Faults

Actuators faults are considered in this section. According


3.2 Residues to the motors distribution on the frame, it can be shown
that any motor failure will have an effect on the different
In case of a hexarotor system, the number of state variables measured variables and therefore all the residuals will be
is n = 12 and the number of control inputs and flat outputs triggered except in the cases of failures of motors 3 and
is m = 4. 12 residues can be obtained. For simplicity the 6 that do not contribute to the pitching movement of the
time derivatives of the three position states and the three hexarotor. We will show below that only three residuals:
orientation states are considered unfaulty at any time. R1x , R2x and R4u , are sufficient for motor failure detection
Therefore the residues are reduced to 6: and isolation. The results are summarized in Table 2
below.
R1x = φm − φe R2x = θm − θe
R1u = u1m − u1e R2u = u2m − u2e (15) • Residue R4u :
R3u = u3m − u3e R4u = u4m − u4e R4u = u4m − u4e where u4m = −Kpψ ψ − Kdψ ψ̇ is
τψ
the output of the yaw controller and u4e = ψ̈ = Izz .
where the subscripts m and e denote measured and es- Then the residual R4u can be written as:
timated respectively. These residuals should satisfy the
following properties (Hwang et al. (2009)): τ ψ t2 τψ τψ
R4u = −Kpψ − Kdψ t−
Izz 2 Izz Izz
• Invariance Relation: When no fault occurs, the mean (16)
of the residual is zero. 1 Kp ψ 2
• Fault Detectability: When a fault occurs, the mean of =− ( t + Kdψ t + 1)τψ
Izz 2
the residual deviates from zero.
with Kpψ and Kdψ are positive gains. The sign of R4u
depends on τψ :
3.3 Detection and Isolation on Sensors Faults · If Q1 = 0 or Q4 = 0 or Q6 = 0 then R4u < 0;
· If Q2 = 0 or Q3 = 0 or Q5 = 0 then R4u > 0;
Multirotor UAVs are often equipped with mutiple sensors • Residue R2x :
such as lightweight inertial measurement unit including R2x = θm − θe with θ̈m  Iτyy θ
and θe = atan( z¨3z¨+g
1
).
3-axis gyro, accelerometer and magnetometer and GPS
for position estimation. The measurements provided by With further simplifications , z¨1 = um 1θ
and z¨3 = um1 −
these sensors could suffer from bias and faults. This g, the residue R2x can be written as:
section considers the fault detection and isolation (FDI) R2x  θm − atan(θm )
problem using the residuals computed in Eq. (15). Since τθ t 2 τθ t2 (17)
 − atan( )
the system is naturally decoupled, the faults on the flat Iyy 2 Iyy 2

1483
2019 IFAC NOLCOS
834
Vienna, Austria, Sept. 4-6, 2019 M Saied et al. / IFAC PapersOnLine 52-16 (2019) 831–836

The sign of R2x depends on τθ :


· If T1 = 0 or T2 = 0 then R2x < 0;
· If T4 = 0 or T5 = 0 then R2x > 0;
· If T3 = 0 or T6 = 0then R2x = 0;
• Residue R1x :
τφ ¨2
R1x = φm − φe with φ̈m  Ixx and φe  atan( z¨−z
3 +g
).
With further simplifications , z¨2 = − um 1φ
and z¨3 =
u1
m − g, the residue R 1x can be written as:
R1x  φm + atan(φm )
τφ t2 τφ t2 (18)
 + atan( )
Ixx 2 Ixx 2
The sign of R1x depends on τφ :
· If T1 = 0 or T5 = 0 or T6 = 0 then R1x > 0;
· If T2 = 0 or T3 = 0 or T4 = 0 then R1x < 0;
Fault R1x R2x R4u
F1 + - - Fig. 3. Residues after a fault occurring on the sensor
F2 - - + measuring θ at time 50s. R2x , R1u and R3u are
F3 - 0 + triggered as indicated in the fifth row of Table 1.
F4 - + -
F5 + + +
F6 + 0 -

Table 2: The residue matrix for isolation of actuators


faults
4. RESULTS

To validate the proposed FDI, we realize simulations and


real experiments using fault injection.
4.1 Matlab Simulations

The nonlinear model of the hexarotor based on (1) is


implemented in Simulink. Actuators and sensors faults are
simulated by considering total motor failures and bias on
sensors giving state variables measurements respectively.
In order to represent the control inputs in terms of the
flat outputs, the high-order derivatives of these outputs
should be estimated using a high gain observer (Vasiljevic Fig. 4. Residues after a fault occurring on the sensor
and Khalil (2008)). Figures 2, 3 and 4 show the residues measuring z at time 50s. R1u is triggered, R1x and
after faults occurring on sensors measuring x, θ and z R2x are slightly affected as indicated in the third row
respectively at time 50s. of Table 1.

Similarly, figures 5 and 6 show for example the residues


R1x , R2x and R4u after failures on motors 1 and 2
respectively at time 50s.

Fig. 5. Residues after a total failure occurring on motor 1


Fig. 2. Residues after a fault occurring on the sensor at time 50s. R1x , R2x and R4u are triggered and have
measuring x at time 50s. Only R2x and R3u are respectively +,-,- signs as indicated in the first row of
triggered as indicated in the first row of Table 1. Table 2.

1484
2019 IFAC NOLCOS

Vienna, Austria, Sept. 4-6, 2019 M Saied et al. / IFAC PapersOnLine 52-16 (2019) 831–836 835

Fig. 6. Residues after a total failure occurring on motor 2


at time 50s. R1x , R2x and R4u are triggered and have
respectively -,-,+ signs as indicated in the second row
Fig. 8. Motors inputs after a total failure on motor 3 at
of Table 2.
time 81s.
4.2 Experimental Results

In this section, the FD scheme is tested in real experiments


on a hexarotor UAV after motors failures. Fault injection
is used to simulate a motor failure.

Experimental Platform An outdoor test environment has


been created to test the proposed FD scheme (Figure 7).
Fault injection is used to simulate a motor failure. The
experimental UAV is a Tarot hexarotor with six rotors
distributed around the center of gravity to form a PPN-
NPN rotor arrangement. The hexarotor frame is mounted
with a pixhawk flight controller. The data sent from the
GPS to the flight controller is merged with an IMU and
a Barometer raw data using an Extended Kalman Filter
(EKF) to form a high rate positioning system and an Atti-
tude Heading Reference System (AHRS). Mission Planner
ground station application permits to monitor and log the
position, attitude and some necessary parameters of the
vehicle during its mission. The injection of the fault is
performed to imitate a real motor failure by commanding
the desired motor to stop completely.

Fig. 9. Residues after a total failure occurring on motor


3 at time 81s. R2x is not triggered, R1x and R4u are
Fig. 7. Overview of the test environment triggered and have respectively -,+ signs as indicated
in the third row of Table 2.
Results The hexarotor was hovering in a stable position threshold since motor 3 does not contribute to the pitching
when failures were injected on motors 3 and 6. Single movement. The motor failure is detected in less than 1 s.
failure is assumed in each scenario. Figures 8 and 9 show The same test is repeated on motor 6 at time 122 s as
respectively the motors inputs in case of a total failure on shown in figures 10 and 11.
motor 3 at time 81 s and the corresponding residues on In this approach, most of the UAV dynamics were ne-
the roll and pitch angles and the yaw control input. The glected. In order to obtain more accurate results, the
hexarotor is recovered after a delay of 1s. The black crosses hexarotor model should be identified and the different
on each figure indicate the fault injection and recovery nonlinear effects should be considered.
times. Residues are examined for the likelihood of faults
by using fixed thresholds. The thresholds are determined 5. CONCLUSION AND PERSPECTIVES
from experiments and are selected in a way to avoid wrong
decisions, such as false alarms or absences of detection. It This paper investigates the problem of fault detection and
is shown that the residue on pitch angle do not exceed the isolation for a hexarotor UAV. Flatness technique has been

1485
2019 IFAC NOLCOS
836
Vienna, Austria, Sept. 4-6, 2019 M Saied et al. / IFAC PapersOnLine 52-16 (2019) 831–836

REFERENCES
Avram, R., Zhang, X., and Muse, J. (2017). Quadrotor
sensor fault diagnosis with experimental results. Journal
of Intelligent and Robotic Systems, 86, 115–137.
Bouabdallah, S. (2007). Design and Control of Quadrotors
with Application to Autonomous Flying. Ph.D. thesis,
Ecole Polytechnique Federale de Lausanne.
Cen, Z. and Noura, H. (2013). An adaptive thau observer
for estimating the time-varying loe fault of quadrotor
actuators. In 2013 Conference on Control and Fault-
Tolerant Systems (SysTol).
Fig. 10. Motors inputs after a total failure on motor 6 at Cen, Z., Noura, H., Susilo, T., and Younes, Y.A. (2013).
time 122.2 s. Robust fault diagnosis for quadrotor uavs using adap-
tive thau observer. Journal of Intelligent and Robotic
Systems, 73, 573–588.
Chamseddine, A., Zhang, Y., Rabbath, C., Join, C., and
Theilliol, D. (2012). Flatness-based trajectory plan-
ning/replanning for a quadrotor unmanned aerial ve-
hicle. IEEE Transactions on Aerospace and Electronic
Systems, 48, 2832–2848.
Chandra, K., Alwi, H., and Edwards, C. (2015). Fault
reconstruction for a quadrotor using an lpv sliding mode
observer. In IFAC-PapersOnLine, volume 48, 374–379.
Fliess, M., Levine, J., Martin, P., and Rouchon, P. (1995).
Flatness and defect of non-linear systems: introductory
theory and examples. nternational journal of control,
61, 1327–1361.
Hwang, I., Kim, S., Kim, Y., and Seah, C. (2009). A survey
of fault detection, isolation, and reconfiguration meth-
ods. IEEE Transactions on Control Systems Technology,
18, 636–653.
Lai, L., Yang, C., and Wu, C. (2006). Time-optimal
control of a hovering quad-rotor helicopter. Journal of
Intelligent and Robotic Systems, 45, 115–135.
Lopez-Estrada, F., Ponsart, J., Theilliol, D., Zhang, Y.,
and Astorga-Zaragoza, C. (2015). Lpv model-based
tracking control and robust sensor fault diagnosis for
a quadrotor uav. Journal of Intelligent and Robotic
Systems, 84, 163–177.
Nan, Z., Antoine, D., Andrei, D., and Felixo, M. (2008).
A differential flatness approach for rotorcraft fault de-
tection. In Proceedings of the 27th Chinese Control
Conference, 237–241.
Saied, M., Lussier, B., Fantoni, I., Francis, C., and Shraim,
H. (2015a). Fault tolerant control for multiple successive
failures in an octorotor: Architecture and experiments.
Fig. 11. Residues after a total failure occurring on motor 6 In IEEE/RSJ International Conference on Intelligent
at time 122.2s. R2x is not triggered, R1x and R4u are Robots and Systems (IROS 2015), 40–45.
triggered and have respectively +,- signs as indicated Saied, M., Lussier, B., Fantoni, I., Francis, C., Shraim,
in the sixth row of Table 2. H., and Sanahuja, G. (2015b). Fault diagnosis and
fault-tolerant control strategy for rotor failure in an
used for residual generation. In general, this approach octorotor. In Robotics and Automation (ICRA), 2015
takes advantage of the non-uniqueness property of the IEEE International Conference on, 5266–5271.
flat output vector to increase the number of generated Torres, C. (2013). Fault Tolerant Control By Flatness
residues. For a hexarotor, only one set of flat outputs can Approach. Ph.D. thesis, Universidad Autonoma De
be found. Due to the internal decoupling in the hexarotor Nuevo Leon.
model, each single sensor fault could be detected and Vasiljevic, L. and Khalil, H. (2008). Error bounds in
isolated. We have analytically and experimentally proved differentiation of noisy signals by high-gain observers.
that actuators faults could be also detected and isolated. Systems & Control Letters, 57, 856–862.
However sensor faults were conducted by simulations only.
In future works we intend to test online a complete archi-
tecture including sensors and actuators failures detection
coupled with system recovery.

1486

You might also like