See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/306260304

Air Data Sensor Fault Detection and Diagnosis with Application to Real Flight
Data

Conference Paper · January 2015

DOI: 10.2514/6.2015-1311

CITATIONS READS

11 860

4 authors, including:

Peng Lu Erik-Jan Van Kampen

The University of Hong Kong Delft University of Technology
65 PUBLICATIONS 1,194 CITATIONS 217 PUBLICATIONS 2,748 CITATIONS

SEE PROFILE SEE PROFILE

Q.P. Chu

265 PUBLICATIONS 5,502 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Sensor Fusion and State Reconstruction for Moving-Base Flight Simulators View project

Reinforcement Learning for Flight Control View project

All content following this page was uploaded by Q.P. Chu on 24 September 2017.

The user has requested enhancement of the downloaded file.

 JOURNAL OF GUIDANCE, CONTROL, AND DYNAMICS
Vol. 39, No. 3, March 2016

Adaptive Three-Step Kalman Filter for Air Data Sensor

Fault Detection and Diagnosis

P. Lu,∗ L. Van Eykeren,† E. van Kampen,‡ C. C. de Visser,§ and Q. P. Chu¶

Delft University of Technology, 2600 GB Delft, The Netherlands
DOI: 10.2514/1.G001313
Air data sensor fault detection and diagnosis is important for the safety of aircraft. In this paper, first an extension
of the robust three-step Kalman filter to nonlinear systems is made by proposing a robust three-step unscented
Kalman filter. The robust three-step unscented Kalman filter is found to be sensitive to the initial condition error when
dealing with air data sensor fault estimation. A theoretical analysis of this sensitivity is presented and a novel adaptive
three-step unscented Kalman filter is proposed which is able to cope with not only the estimation of the air data sensor
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

faults, but also the detection and isolation of faults. The adaptive three-step unscented Kalman filter contains three
steps: time update, fault estimation and measurement update. This approach can reduce the sensitivity to the initial
condition error. Finally, the air data sensor fault detection and diagnosis performance of the adaptive three-step
unscented Kalman filter is validated using simulated aircraft data. Additionally, its performance is further validated
using real flight-test data to demonstrate its performance under realistic uncertainties and disturbances. The results
using both the simulated data and real flight-test data demonstrate the satisfactory fault detection and diagnosis
performance of the adaptive three-step unscented Kalman filter and verify that it can be applied in practice to enhance
the safety of aircraft.

Nomenclature I. Introduction
linear accelerations along body axis, m∕s2
Ax , Ay , Az
Axm , Aym , Azm
=
= measurements of linear accelerations along
body axis, m∕s2
P RESENTLY, fault detection and isolation (FDI) has an important
role in achieving fault tolerance of aircraft [1]. During the past
few decades, many approaches have been proposed for sensor or
f = output faults actuator FDI [2–4]. In aerospace engineering, the FDI of sensors and
f^ = estimation of output faults actuators for fixed-wing aircraft is widely studied, as can be found in
fV , fα , fβ = faults in air data sensors Patton [1], Marzat et al. [5], and Hajiyev and Caliskan [6]. Inves-
L, l, m, p = dimensions of state, input, output, and output tigation of the FDI for unmanned aerial vehicles can also be found
faults, respectively [5,7]. For recent advances, the reader is referred to Goupil [8] and
p, q, r = roll, pitch, and yaw rate along body axis, rad∕s Zolghadri et al. [9,10]. The air data sensors (ADSs) measure the
pm , qm , rm = measurements of roll, pitch, and yaw rate dynamic pressure, airspeed, angle of attack, and angle of sideslip of
along body axis, rad∕s the aircraft, providing essential information on the aircraft states to
T = thresholds for detecting faults the pilot [11]. The ADSs are usually installed outside the aircraft
V = true airspeed, m∕s fuselage and can suffer from icing or water accumulation, which may
V m , αm , β m = air data sensor measurements result in faults such as blockage faults [12]. These faults may nega-
^ P
x, = state estimate and its error covariance matrix tively influence the information provided to the pilot, which can lead
of filter to catastrophic accidents. In the recent past, there have been com-
α, β = angle of attack and sideslip angle, rad mercial aircraft accidents caused by ADS faults. Because of faults in
αvm , βvm = angle of attack and sideslip angle measure- the ADSs, the flight crew of Austral Lineas Aeroeas Flight 2553
ments from vane, rad improperly referenced the airspeed indicator and induced a structure
γ = innovation of filter failure by exceeding safe airspeed limits [13]. More recently, the final
ϕ, θ, ψ = roll, pitch, and yaw angles along body axis, report of the Air France Flight 447 accident stated that erroneous
rad airspeed measurements from the pitot probes were a contributing
ϕm , θm , ψ m = measurements of roll, pitch, and yaw angles factor [14]. Since 2003, commercial aircraft have had more than 35
along body axis, rad recorded incidents of multiple ADS faults [13]. There have also been
accidents of military aircraft caused by ADS faults. The crash of a B-2
bomber is due to a large bias to the ADSs, which is caused by
Presented as Paper 2015-1311 at the AIAA Guidance, Navigation, and moisture in the port transducer units [14]. These facts indicate the
Control Conference, Kissimmee, FL, 5–9 January 2015; received 10 February importance of the fault detection of the ADSs.
2015; revision received 9 May 2015; accepted for publication 8 June 2015; The fault detection of ADSs has been investigated in a number of
published online 13 August 2015. Copyright © 2015 by Peng Lu. Published studies [12,15]. Some researchers propose to use alternative air data
by the American Institute of Aeronautics and Astronautics, Inc., with
sensing systems such as a flush air data sensing system [15,16].
permission. Copies of this paper may be made for personal or internal use, on
condition that the copier pay the $10.00 per-copy fee to the Copyright Nebula et al. proposed a virtual air data system against ADS failures
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923; include [17,18]. Looye and Joos [19] proposed to use the data from a navi-
the code 1533-3884/15 and $10.00 in correspondence with the CCC. gation system to determine the air data information. On the other
*Ph.D. Student, Control and Simulation Division, P.O. Box 5058; P.Lu-1@ hand, the faults of the ADSs can be detected. Houck and Atlas [11]
tudelft.nl. were one of the first to analyze ADS faults. The limitation of their
†
Ph.D. Student, Control and Simulation Division, P.O. Box 5058; approach is that independent static pressure measurements are not
L.VanEykeren@tudelft.nl. always available in unmanned aerial vehicle applications [13].
‡
Assistant Professor, Control and Simulation Division, P.O. Box 5058;
Eubank et al. [13] and Cervia et al. [20] detected the faults using a
e.vankampen@tudelft.nl.
§
Assistant Professor, Control and Simulation Division, P.O. Box 5058; multiple-redundancy air data system. The air data system studied by
c.c.devisser@tudelft.nl. Cervia et al. [13] is based on pseudoquadruplex redundancy, which
¶
Associate Professor, Control and Simulation Division, P.O. Box 5058; employs four self-aligning air data probes. Freeman et al. [12]
q.p.chu@tudelft.nl. investigated analytical redundancy instead of hardware redundancy
590
 LU ET AL. 591

for the ADS fault detection. They used a longitudinal dynamics mod- A. Aircraft Kinematic Model with ADS Faults
el of the aircraft and two linear H ∞ filters are designed to detect the The kinematic model of the aircraft including ADS faults is
faults and provide robustness to model errors. described as
Alternatively, the kinematic model can be used to detect the faults
in the ADSs, thereby reducing the influence of model uncertainties _  f
x
t;
x
t  um
t; t  G
x
tw
t (1)
caused by the calculation of the aerodynamic forces and moments
[21,22]. Van Eykeren and Chu [23] use an adaptive extended Kalman
filter to detect the faults in the ADSs. However, the estimation of
the faults is not addressed in their work. In Lu et al. [24], a selective- y
t  h
x
t; um
t; t  v
t  F
tf
t; t  ti ;
reinitialization multiple-model adaptive estimation approach is (2)
proposed for the ADS fault detection and diagnosis (FDD). The ap- i  1; 2; : : :
proach improved the FDD performance of the multiple-model-based
approaches. However, the computational load of the approach is where x ∈ RL represents the system states, um ∈ Rl is the measured
intensive when dealing with simultaneous faults. input, and y ∈ Rm is the measurement. The functions f and h are
In this paper, a newly developed robust three-step Kalman filter nonlinear functions. G and F are the noise distribution matrix and
(RTS-KF) [25] is combined with the kinematic model to estimate the output fault distribution matrix. The function f ∈ Rp represents
ADS faults. First, the RTS-KF is extended to cope with nonlinear output faults.
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

systems by proposing a novel robust three-step unscented Kalman The system equation variables are defined as follows:
filter (RTS-UKF). The RTS-UKF is able to reduce linearization
errors. However, it is found that the RTS-UKF is sensitive to the x  V α β ϕ θ ψ T (3)
initial condition error. Second, the sensitivity of this three-step
Kalman filter to the initial condition error is analyzed theoretically. It
is proved that the RTS-UKF does not use some of the measurements
to update the state estimation, which causes the sensitivity to the um   Axm Aym Azm pm qm r m T
initial condition error.
Finally, a novel adaptive three-step unscented Kalman filter (ATS-   Ax Ay Az p q r T  w (4)
UKF) is proposed, which not only estimates the ADS faults, but also
detects and isolates the faults. The ATS-UKF contains three steps:
time update, fault estimation, and measurement update. The fault
detection is performed before the fault estimation. This approach also y   Vm αm βm ϕm θm ψ m T (5)
reduces its sensitivity to the initial condition. The fault detection is
performed by checking the innovation variances. In the presence of
faults, the innovation variance increases. If the innovation variance
exceeds a predefined threshold, then the fault alarm is triggered. The w   wx wy wz wp wq wr  T (6)
FDD performance of the ATS-UKF is tested using simulated aircraft
data with the objective of detecting, isolating, and estimating ADS
faults. Two different fault scenarios (multiple faults and simultaneous
faults) are implemented to test the performance, and the results v   vV vα vβ vϕ vθ vψ T (7)
demonstrate the satisfactory performance of the ATS-UKF. The fault
types contain not only bias and drift fault, but also oscillatory faults.
Furthermore, the FDD performance of the ATS-UKF is validated
using real flight-test data of a Cessna Citation II aircraft. The sensor
f   fV fα f β T (8)
measurements from the real flight test contain biases and uncer-
tainties and are suitable for testing the performance of the ATS-UKF.
Different fault scenarios are generated and the faults are injected into where the input um is the inertial measurement unit (IMU) measure-
the real flight data. The ADS FDD results of the ATS-UKF demon- ment, which measures the linear accelerations Ax , Ay , and Az and
strate its performance and verify that it can be applied in practice to angular rates (roll rate p, pitch rate q, and yaw rate r) of the aircraft; y
enhance the safety of the aircraft. is the output measurement, which measures the air data information
The structure of the paper is as follows: in Sec. II, the ADS FDD (true airspeed V, angle of attack α, and angle of sideslip β) and Euler
problem is formulated. The kinematic model including ADS faults is angles (roll angle ϕ, pitch angle θ, and yaw angle ψ);  fV fα fβ T
introduced. Section III extends the RTS-KF to cope with nonlinear are the faults of the ADSs (i.e., fV , fα , and fβ are the faults in the
systems by proposing the RTS-UKF. The RTS-UKF is applied to velocity sensor, angle of attack sensor, and angle of sideslip sensor,
estimate the ADS faults, which turns out to be sensitive to the initial respectively). It is assumed that there are no faults in the attitude and
condition. The sensitivity problem is analyzed theoretically and a heading reference system, which measures the Euler angles, and the
novel ATS-UKF is proposed to deal with not only the estimation of influence of changing wind such as turbulence is limited. Therefore,
the ADS faults, but also the detection and isolation of the faults. The the input noise vector w
t can be assumed to be a continuous-time
performance is tested using a simulated aircraft model. In Sec. IV, the white noise process, whereas the output noise vector v
t can be
performance of the ATS-UKF is further validated using the real flight assumed to be a discrete time noise sequence:
data of the Cessna Citation II aircraft. The performance is shown and
some remarks are given. Finally, the conclusions are made in Sec. V. Ew
t  0
Ew
twT
t τ   Qδ
t − τ;
Q diag
σ 2wx ; σ 2wy ; σ 2wz ; σ 2wp ; σ 2wq ; σ 2wr ;
II. Air Data Sensor FDD Using the Kinematic Model
The objective of this paper is the FDD of the aircraft ADSs. Ev
t  0 (9)
However, model-based approaches are sensitive to model uncer-
tainties. To make the proposed approach more robust, the kinematic
model of aircraft, which does not involve the computation of Ev
ti vT
tj   Rδ
ti − tj ; R  diag
σ 2vV ;σ 2vα ; σ 2vβ ; σ 2vϕ ;σ 2vθ ;σ 2vψ 
aerodynamic forces and moments, is used instead of the aerody-
(10)
namic model.
 592 LU ET AL.

Ew
tvT
ti   0; t  ti ; i  1; 2; : : : (11) ψ m  ψ  vψ (24)

The kinematic model is given as follows [23,24]:

The measurement model can be rewritten into
V_ 
Axm − g sin θ cos α cos β 
Aym  g sin ϕ cos θ sin β
y
t  x
t  F
tf
t  v
t; t  ti ; i  1; 2; : : :

Azm  g cos ϕ cos θ sin α cos β (12)
(25)

where
1
α_ 
−Axm sin α  Azm cos α  g cos ϕ cos θ cos α
V cos β
F   I3 03×3 T (26)
 g sin θ sin α  qm −
pm cos α  rm sin α tan β (13)

The objective of the ADS FDD problem is to detect, isolate, and

estimate f   fV fα fβ T . This paper assumes that there are no
1 faults in the IMU sensors. If there are faults in the IMU sensors, they
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

β_  −
Axm − g sin θ cos α sin β 
Aym  g sin ϕ cos θ cos β
V
−
Azm  g cos ϕ cos θ sin α sin β  pm sin α − rm cos α (14)
ϕ_  pm  qm sin ϕ tan θ  rm cos ϕ tan θ
θ_  qm cos ϕ − rm sin ϕ
can be detected and estimated by other methods using another set of
kinematic models [26].
B. Fault Scenarios for the ADS FDD
In this paper, two different fault scenarios are used to test the
(15) performance of the approaches. The fault scenario for multiple ADS
faults is given in Table 1, and that for simultaneous ADS faults is
given in Table 2. The fault type, magnitude, and unit are given in the
table. The units of the drift faults are given by the units of the drift
rates. It can be seen that the fault types not only contains bias faults,
(16)
but also drift faults and oscillatory faults.

C. State Observability and Fault Reconstructability

sin ϕ cos ϕ This section checks the observability of the system described by
ψ_  qm  rm (17)
cos θ cos θ Eqs. (1) and (2). The observability analysis of the system can be
performed by checking the rank of the following observability
and G
x
t is defined as matrix:

2 − cos α cos β − sin α cos β − sin α cos β 0 0 0 3

6 sin α∕
V cos β 0 − cos α∕
V cos β cos α tan β −1 sin α tan β 7
6 7
6 cos α sin β∕V − cos β∕V sin α sin β∕V − sin α 0 cos α 7
G
x
t  6 7 (18)
6 0 0 0 −1 − sin ϕ tan θ − cos ϕ tan θ 7
4 5
0 0 0 0 − cos ϕ sin ϕ
0 0 0 0 − sin ϕ∕ cos θ − cos ϕ∕ cos θ

Therefore, the measurement model including the ADS faults is

Table 1 Fault scenario of multiple faults
V m  V  fV  vV (19)
Time interval Sensor Fault type Fault magnitude Fault unit
10 < t < 20 s V Bias 2 m∕s
30 < t < 40 s α Drift 0.01 t rad∕s
50 < t < 60 s β Oscillatory −2π sin
πt∕180 Rad
αm  α  fα  vα (20)

βm  β  fβ  vβ (21) Table 2 Fault scenario of simultaneous faults

Sensor Fault type Fault magnitude Fault unit

Time interval 10 < t < 20 s
V Oscillatory 2 sin
πt m∕s
ϕm  ϕ  vϕ (22) α Drift 0.01 t rad∕s
β Drift −0.01 t rad∕s
Time interval 30 < t < 40 s
V Drift −0.2 t m∕s2
α Bias −2π∕180 Rad
θm  θ  vθ (23) β Oscillatory −2π sin
πt∕180 Rad
 LU ET AL. 593

2δx h 3
w
m
i  w
c
i  1∕f2
L  γ 0 g; i  1; 2; : : : ; 2L (31c)
δ
6 x f
L h 7
6 7
O6 .. 7 (27) with X i;k−1 as the sigma points of the states (dimension L) at
4 . 5
δx
LL−1 h step k − 1; w
mi and w
c
i are the weights associated with the ith
f
point with respect to x^ k−1jk−1 and Pk−1jk−1 , respectively; γ 0 
α20
L  κ − L is a scaling factor, α0 determines the spread of
where the Lie derivative is defined as follows:
the sigma points around x^ k−1jk−1 , κ is a secondary scaling factor,
and β0 is used to incorporate the prior knowledge of the
Lf h  δx h · f
distribution of x. In this paper, κ  0, α0  0.8, and β0  2.
.. (28) After the creation of the sigma points through the nonlinear
.
LL−1 h  δ
LfL−2  transformation, the predicted mean and covariance are com-
 x  h · f
f puted as follows:
It can be readily checked that O is of full rank. Therefore, the Zt
k
system state is observable. To reconstruct the faults, additional X i;kjk−1  X i;k−1   i;k−1 ; u
τ; τ dτ
f
X (32)
tk−1
conditions are required, which are given in condition (29).
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

III. Extension of the Robust Three-Step Kalman Filter

X
2L
This section extends the RTS-KF to estimate output faults. First, in x^ kjk−1  w
m
i X i;k;k−1 (33)
Sec. III.A, the RTS-KF is extended to nonlinear systems by propos- i0
ing an RTS-UKF. This RTS-UKF is applied to the ADS fault estima-
tion problem and is found to be sensitive to the initial condition errors.
This sensitivity problem is analyzed theoretically in Sec. III.B. Then,
in Sec. III.C, an ATS-UKF is proposed that can detect, isolate, and X
2L
estimate the faults. Finally, the ATS-UKF is applied to the ADS FDD Pkjk−1  w
c
i X i;kjk−1 − x
^ kjk−1 X i;kjk−1 − x^ kjk−1 T Qd
problem in Sec. III.D to demonstrate its FDD performance. i0
(34)
A. Robust Three-Step Unscented Kalman Filter
The RTS-KF [25] can be used for output FDD. Consider the
aircraft kinematic model described by Eqs. (1) and (2). For this p p
system, because the system state is observable, the existence condi- X i;kjk−1  X 0: 2L;kjk−1 X 0;kjk−1 − ν Qd X 0;kjk−1  ν Qd i
tion of an RTS-KF is [25] (35)
m ≥ p; rank Fk  p (29)

In this study, m  6, p  3, and rank Fk  3. Therefore, an RTS- Y i;kjk−1  h
X i;kjk−1  (36)
KF can be designed to estimate the ADS faults.
However, the RTS-KF is designed for linear systems, whereas the
kinematic model is nonlinear. Therefore, the RTS-KF needs to be
extended to cope with nonlinear systems. The UKF is a nonlinear
filter that can achieve a better level of accuracy than the extended X
2La
y^ k  w
m
i Y i;kjk−1 (37)
Kalman filter (EKF) [27,28]. This section extends the RTS-KF to i0
nonlinear systems by proposing an RTS-UKF.
According to the technique in Lu et al. [22], the RTS-UKF can be
derived as follows:
Step 1) Sigma points calculation and time update X
2La
Pxy;k  w
c
i X i;kjk−1 − x^ kjk−1 Y i;kjk−1 − y^ k T (38)
X 0;k−1  x^ k−1jk−1 (30a) i0

q
X i;k−1  x^k−1jk−1 −

L  γ 0 Pk−1jk−1  ; i  1; 2; : : : ; L X
2La
i Pyy;k  w
c
i Y i;kjk−1 − y^k Y i;kjk−1 − y^ k T  R (39)
(30b) i0

q
X i;k−1  x^k−1jk−1 

L  γ 0 Pk−1jk−1 
i−L
i  L  1; L  2; : : : ; 2L
p
where La  2L, ν  L  γ 0 , w
m i and w
c
i are calculated
similar to Eq. (31) with the replacement of L by La. Qd is
approximated by G
x^ kjk−1 QGT
x^kjk−1 Δt where Δt 
; tk − tk−1 .
(30c) Step 2) Estimation of the faults

γ k 
yk − y^ k  (40)

w
m
0  γ 0 ∕
L  γ 0  (31a)
N k 
FTk P−1 −1 T −1
yy;k Fk  Fk Pyy;k (41)

w
c
0  γ 0 ∕
L  γ 0  
1 − α0  β 0 
2 (31b) f^k  N k γ k (42)
 594 LU ET AL.

Pfk 
FTk P−1

yy;k Fk 
−1 (43) It can be seen that the measurement update of the normal UKF, as
given by Eqs. (47–49), does not take the fault estimation and error
where γ k is the innovation, f^k is the estimation of fk , and Pfk is its covariance into account. Also note that the normal UKF does not
error covariance matrix. N k is the gain matrix, which can estimate the faults, which means that it does not contain Eqs. (40–43).
achieve an unbiased estimation of fk . The ADS fault estimation using the RTS-UKF is shown in the
Step 3) Measurement update following.
The performance of the RTS-UKF will be demonstrated under
Kk  Pxy;k P−1 (44) different initial conditions. The simulation data are taken from the
yy;k
simulation model of a Cessna Citation II aircraft. During
10 < t < 17 s, there is a 3-2-1-1 command on the aileron. The fault
scenario is given in Table 1. The true initial state x0 is as follows:

x^ kjk  x^ kjk−1  Kk
yk − y^k − Fk f^k  (45) x0  90; 0.056; 0; 0; 0.0037; 0T (50)

First, the true initial condition (50) is used as the initial guess x^0 in
the filter. P0  10−3 · I 6 . The standard deviations of the
Pkjk  Pkjk−1 − K k
Pyy;k − Fk Pfk FTk K Tk (46) measurement noises are
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

σ wx  σ wy  σ wz  0.001 m∕s2
This measurement update is different from that of the normal UKF
[22,27], which is given next for comparison and quick reference: σ wp  σ wq  σ wr  0.000018 rad∕s
σ vV  0.1 m∕s; σ vα  σ vβ  0.0018 rad
Kk  Pxy;k P−1 (47)
yy;k
σ vϕ  σ vθ  σ vψ  0.0018 rad

x^ kjk  x^ kjk−1  K k
yk − y^ k 
Pkjk  Pkjk−1 − Kk Pyy;k K Tk
Therefore, Q and R can be inferred from Eqs. (9) and (10). The
(48) results are shown in Fig. 1.
The estimation errors of V, α, and β, as shown in Fig. 1a, are close
to zero mean. The estimation errors of ϕ, θ, and ψ using the RTS-UKF
are given in Fig. 1b. It can be seen that the estimation errors are
zero mean except during the period when there is a maneuver
(49) (10 < t < 17 s). However, during this period, the estimation errors

−3
0.1 x 10
2
ΔV (m/s)

Δφ (rad)

0 0

−0.1 −2
0 10 20 30 40 50 10 20 30 40 50
−3
0.01 x 10
1
Δα (rad)

Δθ (rad)

0 0

−0.01 −1
10 20 30 40 50 10 20 30 40 50
−3
0.01 x 10
1
Δβ (rad)

Δψ (rad)

0 0

−0.01 −1
10 20 30 40 50 10 20 30 40 50
time (s) time (s)
a) Error of estimation of V, , and b) Error of estimation of , , and

estimation
5 true 0.5
Δ fV (m/s)
fV (m/s)

0 0

−5 −0.5
0 10 20 30 40 50 60 0 10 20 30 40 50
0.2 0.01
Δ fα (rad)
f (rad)

0.1
0
0
α

−0.1 −0.01
0 10 20 30 40 50 60 0 10 20 30 40 50
0.05 0.01
Δ fβ (rad)
fβ (rad)

0 0

−0.05 −0.01
0 10 20 30 40 50 60 0 10 20 30 40 50
time (s) time (s)

c) Estimation of fV, f , and f d) Estimation error of fV, f , and f

Fig. 1 Result of state and ADS fault estimation using RTS-UKF approach and initial condition (50) in presence of multiple faults.
 LU ET AL. 595

are small (e.g., the maximum estimation error of ϕ is less than
2 × 10−3 rad). The estimation of fV , fα , and fβ is given in Fig. 1c. As
can be seen, all the faults are estimated in an unbiased sense. The
estimation errors can be found in Fig. 1d.
Next, the performance with two different initial conditions for x^ 0 is
tested. The two initial conditions are as follows:
x^ 0  90; 0; 0; 0; 0; 0T
x^ 0  1; 0; 0; 0; 0; 0T
The initial condition (52) significantly deviates from the true initial
condition (50) whereas condition (51) slightly deviates from
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313
condition (50). P0 is the same with the previous simulation and
is 10−3 · I 6 .
The state estimation errors of the RTS-UKF using the initial
condition (51) are shown in Figs. 2a and 2b. As can be seen from
Fig. 2a, the estimation errors of V, α, and β are larger than those
shown in Fig. 1a. The estimation errors of ϕ, θ, and ψ, shown in
Fig. 2b, are the same as those shown in Fig. 1a.
(51) The state estimation errors of the RTS-UKF using the initial
condition (52) are shown in Figs. 2c and 2d. The estimation errors of
V, α, and β, shown in Fig. 2c, are significantly worse than those
shown in Figs. 1a and 2a. However, the estimation errors of ϕ, θ, and
(52) ψ, shown in Fig. 2d, are still zero mean.
The estimates of fV , fα , and fβ using the initial condition (51) and
(52) are demonstrated in Figs. 2e and 2f, respectively. As can be seen
from Fig. 2e, when the initial x0 deviates from the true state, the

−3
0.5 x 10
2
ΔV (m/s)

Δφ (rad)
0 0

−0.5 −2
0 10 20 30 40 50 60 10 20 30 40 50
−3
0.06 x 10
1
Δα (rad)

Δθ (rad)
0.05 0

0.04 −1
0 10 20 30 40 50 60 10 20 30 40 50
−3
0.05 x 10
1
Δβ (rad)

Δψ (rad)

0 0

−0.05 −1
0 10 20 30 40 50 60 10 20 30 40 50
time (s) time (s)
a) Error of estimation of V, , and using (51) b) Error of estimation of , , and using (51)

−3
100 x 10
2
ΔV (m/s)

Δφ (rad)

0
0

−100 −2
0 10 20 30 40 50 60 10 20 30 40 50
1 −3
x 10
1
Δα (rad)

Δθ (rad)

0
0

−1 −1
0 10 20 30 40 50 60 10 20 30 40 50
−3
2 x 10
1
Δβ (rad)

Δψ (rad)

0
0

−2 −1
0 10 20 30 40 50 60 10 20 30 40 50
time (s) time (s)
c) Error of estimation of V, , and using (52) d) Error of estimation of , , and using (52)

estimation estimation
5 true 100 true
fV (m/s)
fV (m/s)

0 0

−5 −100
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.2 1
fα (rad)
fα (rad)

0.1
0
0
−0.1 −1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.05 2
fβ (rad)
f (rad)

0 0
β

−0.05 −2
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

e) Estimation of f V, f , and f using (51) f) Estimation of f V, f , and f using (52)

Fig. 2 Result of state and fault estimation using RTS-UKF approach and two different initial conditions in presence of multiple faults.
 596 LU ET AL.

estimates of the faults also deviate from their true magnitudes, where Lk is defined as
especially that of fα . When the initial condition deviates significantly
from the true initial condition, the performance becomes significantly Lk ≔ Kk
I − Fk N k  (54)
worse, as can be seen in Fig. 2f.
Based on the preceding simulation results, it is seen that the RTS-
UKF is sensitive to the initial condition errors. This sensitivity The covariance matrix Pxy;k , fault distribution matrix Fk , and the
problem will be analyzed theoretically in the following section. innovation γ k and x can be partitioned as follows:
" #  
B. Problem Analysis of the Robust Three-Step Filter P11 P12 Ip
In the preceding sections, it was shown that the performance of Pxy;k  ; Fk  ;
P21 P22 0
the RTS-UKF is influenced by the given initial condition. This " #
γp  x 
section analyzes the problem of the sensitivity to the initial condition. p
Rewrite Eq. (45) into γk  ; x (55)
γ m−p xL−p

x^ kjk  x^ kjk−1  Lk γ k (53) Because Pyy;k is invertible, its inverse can be partitioned as follows:
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

−3
x 10
0.05 2
ΔV (m/s)

Δφ (rad)
0 0

−0.05 −2
10 20 30 40 50 0 10 20 30 40 50 60
−3 −3
x 10 x 10
1 2
Δα (rad)

Δθ (rad)
0 0

−1 −2
10 20 30 40 50 0 10 20 30 40 50 60
−3 −3
x 10 x 10
1 5
Δψ (rad)
Δβ (rad)

0 0

−1 −5
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

a) Error of estimation of V, , and b) Error of estimation of , , and

estimation
1
2 true
f (m/s)
FAV

0.5
0
V

0
−2
0 10 20 30 40 50 60
0 10 20 30 40 50
1 0.1
f (rad)
FAα

0.5 0
α

0
0 10 20 30 40 50 60 −0.1
0 10 20 30 40 50
1 0.05
fβ (rad)
FAβ

0.5 0

0 −0.05
0 10 20 30 40 50 60 0 10 20 30 40 50
time (s) time (s)

c) Fault detection and isolation d) Estimation of fV, f , and f

estimation
2 true 0.5
fV (m/s)

Δ fV (m/s)

0 0
−2
0 10 20 30 40 50 −0.5
0 10 20 30 40 50
0.1 0.1
Δ f (rad)
fα (rad)

0 0
α

−0.1 −0.1
0 10 20 30 40 50 0 10 20 30 40 50
0.05 0.02
Δ f (rad)
f (rad)

0 0
β
β

−0.05 −0.02
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)

e) Weighted estimation of fV, f , and f f) Error of estimation of fV, f , and f

Fig. 3 Result of state estimation and ADS FDD using proposed ATS-UKF approach in presence of multiple faults.
 LU ET AL. 597

 
R~ 11 R~ 12 x^ kjk  x^kjk−1 (63)
P−1
yy;k  (56)
R~ 21 R~ 22
This means that all the states are not updated by their measure-
where R~ 11 ∈ Rp×p , R~ 12 ∈ Rp×
m−p , R~ 21 ∈ R
m−p×p , and ments. In this situation, all the state estimations will be sensitive to the
R~ 22 ∈ R
m−p×
m−p . Therefore, Eq. (41) can be computed by initial condition.
  Through the analysis in this section and the performance demon-
R~ R~ 12 stration of the RTS-UKF in Sec. III.A, the need for a modification of
N k   R~ −1
11 0  11   I p R~ −1
11 R12 
~ (57)
0 0 the RTS-UKF is emphasized. In real life, the exact initial condition is
difficult to obtain due to uncertainties in the system (which can also
Substituting Eq. (57) into Eq. (54), it follows that be found in Sec. V). The RTS-UKF will interpret the initialization
  error as a fault, which results in wrong fault estimation. Therefore, the
0 −R~ −1 ~
11 R12 RTS-UKF cannot be applied to the FDD of the ADSs.
Lk  Pxy;k P−1
yy;k (58)
0 I m−p
C. Novel Adaptive Three-Step Unscented Kalman Filter for ADS
FDD
  Having found the cause for performance degradation of the RTS-
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

0 L12 UKF, this section proposes a novel ATS-UKF to solve the ADS FDD.
 (59)
0 L22 The sensitivity to the initial condition of the RTS-UKF can be solved
by performing the measurement update of normal UKF.
where L12 and L22 are defined as It should be noted that the RTS-UKF only considers the estimation
of the faults. It does not detect and isolate the faults. The proposed
L12 ≔ P12
R~ 22 − R~ 21 R~ −1
11 R12 
~ L22 ≔ P22
R~ 22 − R~ 21 R~ −1
11 R12 
~ ATS-UKF deals with not only the estimation of the faults, but also the
detection and isolation.
Therefore, the measurement update of the robust three-step filter, In the following, the initial measurement update and FDI scheme
denoted in Eq. (53), can be further written as follows: are introduced. Then the complete FDD system is introduced.
 
L12 γ m−p 1. Initial Measurement Update
x^ kjk  x^kjk−1  (60)
L22 γ m−p The solution to reduce the sensitivity of the RTS-UKF to the
initial condition is proposed in this subsection, which is to use the
It can be seen that γ p is not used in the measurement update. measurement update of normal UKF [Eqs. (47–49)] when the state
Because γ p is not used, the estimation of xp is not updated by estimation is influenced by the initialization error. However, when
measurements of xp . Therefore, the estimation of xp (V, α, and β), is the correction is sufficient (i.e., when the measurement update of the
sensitive to the initial condition. If the initial x0 significantly deviates UKF is sufficient) needs to be determined. This paper proposes a
from the true value, it will not be corrected to the true value. However, criteria that can determine whether the measurement update of the
the estimation of ϕ, θ, and ψ is not influenced because they are UKF is sufficient. The details are given as follows:
updated by the measurement. This is consistent with the result shown Let Cii;k , i  1, 2, 3 denote the ith diagonal elements of the
in Figs. 2b and 2d, in which the estimation of ϕ, θ, and ψ is still good innovation covariance matrix Ck associated with the measurements
even when that of V, α, and β is not. that are not used in the update of the RTS-UKF at time step k (i.e., the
In case p  m and rank Fk  m, it can be found that measurement of V, α, and β, respectively, in this paper).
Define the change of the innovation variance ΔCii;k as
N k  F−1
k (61)
ΔCii;k ≔ Cii;k − Cii;k−1 ; i  1; 2; 3 (64)

When the following inequality holds, the measurement update can

Lk  0 (62)
be regarded as sufficient. The inequality is
Consequently, the measurement update of the three-step Kalman
filter is ΔCii;k < ηi ; i  1; 2; 3 (65)

OFC
1 1
V

V
OFC

0.5 0.5
FA

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50
1 1
OFCα

FA α

0.5 0.5

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50
1 1
β

FA β
OFC

0.5 0.5

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50
time (s) time (s)
a) Detection of oscillatory faults, OFC indicates the b) Fault detection and isolation including oscillatory
detection of oscillatory faults fault indication
Fig. 4 Result of fault detection and isolation using proposed ATS-UKF approach in presence of multiple faults.
 598 LU ET AL.

where ηi , i  1, 2, 3 are predefined constants that can be tuned to 2. Fault Detection and Isolation
stop the measurement update. The principle is that, if there are The fault detection is performed by monitoring the innovation
initialization errors, Cii;k is not constant. When the filter achieves variance of the filter. In the presence of the ith fault, Cii;k increases.
steady state, Cii;k is approximately constant. Therefore, ΔCii;k should The fault detection and isolation logic at time step k is
be small. If ΔCii;k is smaller than ηi , then it indicates that the filter has
reached steady state and the measurement update of the UKF is
sufficient. If ηi is chosen to be small, then the number of the initial
measurement update will be bigger, whereas the influence of the if Cii;k > T i ; FAi  1. otherwise FAi  0; i  1; 2; 3
initial condition error will be less. Ck can be estimated using the
following [29,30]:
where FA   FAV FAα FAβ T are the alarm indicators, T i are the
1 X k
C^ k  γ γT (66) thresholds that are designed to detect the faults in the V, α, and β
N jk−N1 j j sensors, respectively. These thresholds are designed based on the
fault-free case. It can be seen that the fault detection and isolation are
where γ j denotes the innovation at time step j. simultaneously realized.
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

−3
x 10
0.05 2
ΔV (m/s)

Δφ (rad)
0 0

−0.05 −2
10 20 30 40 50 0 10 20 30 40 50
−3 −3
x 10 x 10
1 2
Δα (rad)

Δθ (rad)
0 0

−1 −2
10 20 30 40 50 0 10 20 30 40 50
−3 −3
x 10 x 10
1 5
Δψ (rad)
Δβ (rad)

0 0

−1 −5
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)

a) Error of estimation of V, , and b) Error of estimation of , , and

OFC estimation
1 5
true
f (m/s)
V

0.5 0
FA

0 −5
0 10 20 30 40 50 0 10 20 30 40 50
1 0.2
f (rad)
FAα

0.5 0
α

0 −0.2
0 10 20 30 40 50 0 10 20 30 40 50
1 0.2
fβ (rad)
FA β

0.5 0

0 −0.2
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)
c) Fault detection and isolation d) Estimation of fV, f , and f

estimation
2 true 1
fV (m/s)

Δ f (m/s)

0
0
V

−2
0 10 20 30 40 50 −1
0 10 20 30 40 50
0.1
0.1
f (rad)

Δ f (rad)

0 0
α

−0.1 −0.1
0 10 20 30 40 50 0 10 20 30 40 50
0.1 0.01
Δ f (rad)
fβ (rad)

0 0
β

−0.1 −0.01
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)

e) Weighted estimation of fV, f , and f f) Error of estimation of fV, f , and f

Fig. 5 Result of state estimation and ADS FDD using proposed ATS-UKF approach in presence of simultaneous faults.
 LU ET AL.
The weighted fault estimation can be calculated as follows:
fi;k  FAi f^i;k ; i  1; 2; 3 (67)
3. Adaptive Three-Step Unscented Kalman Filter
When the initial measurement update is sufficient, there are two
options to achieve FDD, which are as follows:
1) After the initial measurement update, the FDI scheme is used to
detect and isolate the faults. The RTS-UKF is used to estimate the
faults.
2) After the initial measurement update, the FDI scheme is used to
detect and isolate the faults. If there are no faults detected, the UKF is
used and the fault estimation is considered to be zero. If there are
faults detected, then the RTS-UKF is used for the fault estimation and
measurement update.
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313
The ATS-UKF proposed in this paper is based on the latter one
because it can reduce the computational load. The measurement
update of the ATS-UKF switches adaptively between that of the
normal UKF and that of the RTS-UKF through the FDI scheme. The
specific three steps of the ATS-UKF are given as follows:
1) Time update: This is the same as in the UKF, which also includes
the sigma point calculation. The steps are described by Eqs. (30–39).
2) Fault estimation: Before estimating the faults, the FDI,
introduced earlier, is performed. If FA  0, then f^k  0. If FA ≠ 0,
then the faults are estimated using the RTS-UKF, which is described
by Eqs. (40–43). This step is the FDD.
3) Measurement update: As mentioned, during the initial measure-
ment update, the measurement update of the normal UKF is applied.
When the initial measurement update is done, the measurement
update of the ATS-UKF is as follows:
a) If FA  0, use the measurement update of the normal UKF. In
this situation, there are no faults detected in the system. To reduce the
computational load, the measurement update of the UKF is used. This
means that the faults are considered to be zero, so that the fault
estimation and measurement update of the RTS-UKF are not needed.
The steps are described by Eqs. (47–49).
b) If FA ≠ 0, use the measurement update of the RTS-UKF. In
this situation, faults are detected. Therefore, the measurement update
of the RTS-UKF is needed to obtain an unbiased state estimation and
fault estimation, which cannot be achieved using the normal UKF.
The steps are described by Eqs. (44–46).
D. ADS FDD Using the ATS-UKF
In this section, the FDD, as well as the state estimation perfor-
mance of the proposed ATS-UKF, is demonstrated using two dif-
ferent fault scenarios. The initial condition is the same as in Eq. (52).
The threshold to stop the initial measurement update is η 
5 × 10−3 ; 2 × 10−5 ; 2 × 10−5 T and the threshold to detect the fault
is T  0.2; 1 × 10−4 ; 5 × 10−5 T .
1. Multiple FDD
In this scenario, consecutive ADS faults are generated, which are
shown in Table 1. The results using the ATS-UKF are shown in Fig. 3.
It is found that using the preceding thresholds, the initial measure-
ment update is only performed for two time steps. The estimation
error of V, α, and β is shown in Fig. 3a. Despite the fact that the initial
x0 significantly deviates from the true state, the estimation is still
Table 3 Update frequencies of different
measurements
Measurements Unit Update frequency, Hz
V m∕s 100
un , vn , wn m∕s 1 faults are oscillatory faults.
α, β rad 100
ϕ, θ rad 100
ψ rad 10
599
Fig. 6 Vanes on boom for measuring angle of attack and angle of
sideslip. (Photo credits by Daan Pool.)
satisfactory. This means that the sensitivity to the initial condition of
the RTS-UKF is tackled by the ATS-UKF. The estimation error of ϕ,
θ, and ψ, as shown in Fig. 3b, is also satisfactory. This demonstrates
the state estimation performance of the ATS-UKF.
The fault detection and isolation is achieved by checking FAV , FAα ,
and FAβ , which is shown in Fig. 3c. From the figure, it can be seen that
fV is detected instantaneously. The detection of fα takes longer than
that of fV . This is because fα is a drift fault, which is a slow time-
varying fault; fβ is also detected instantaneously. However, FAβ
switches from one to zero nine times. This is as expected because the
oscillatory fault crosses zero nine times. When the magnitude of the
fault is zero, it can be regarded as no fault. From the figure, it is
obvious that fault isolation is also achieved. For instance, when
FAV  1, both FAα and FAβ are equal to zero, which means only fV
occurs.
The estimation of fV , fα , and fβ is shown in Fig. 3d. As can be
seen, all the faults are estimated in an unbiased sense. The weighted
fault estimation, calculated using Eq. (67), is shown in Fig. 3e. The
error of the estimation of fV , fα , and fβ is shown in Fig. 3f. It is seen
that all the estimation errors are zero mean. This demonstrates the
fault estimation performance of the ATS-UKF. It is also noticed that,
when there are no faults or the faults are not detected, the estimates of
the fault are zero and so are the estimation errors. This is because,
when there are no faults detected, the measurement update of the
UKF is used and the faults are considered to be zero.
The fault detection of the oscillatory faults, shown in Fig. 3c,
shows a chattering behavior. To detect the presence of oscillatory
failures, the detection logic of oscillatory faults in Goupil [31] is used.
The basic idea is to count the crossings of the fault estimate (shown in
Fig. 3e) through a positive and negative threshold within a sliding
time window. In this paper, the oscillatory faults are detected if
one full oscillation is detected. The result of detecting the oscillatory
fault is shown in Fig. 4a. In the figure, OFC denotes oscillatory
failure case (OFC). As can be seen, an oscillatory fault is only
detected in the β sensor. If we take the bigger value of FAi and OFCi (i
is associated with the sensor of V, α, and β), the fault detection,
including the detection of the oscillatory fault, can be obtained, which
is demonstrated in Fig. 4b. In the figure, the dashed line indicates that
the detected fault is an oscillatory fault.
2. Simultaneous FDD
In this scenario, simultaneous faults are generated, which are
shown in Table 2. The ATS-UKF is used to detect, isolate, and
estimate these faults. The results are given in Fig. 5.
The estimation errors of V, α, and β and ϕ, θ, and ψ using the ATS-
UKF are shown in Figs. 5a and 5b, respectively. As can be seen, even
in the presence of simultaneous faults, the state estimation
performance of the ATS-UKF is still satisfactory.
The fault detection and isolation performance is shown in Fig. 5c.
As can be seen, there are no false alarms, which demonstrates its good
performance. For the detection of fα and fβ during 10 < t < 20 s
and fV during 30 < t < 40 s, there are detection delays because there
are drift faults. The dashed lines in the figure indicate that the detected
Figures 5d and 5e show the estimation and weighted estimation of
fV , fα , and fβ , respectively. It can be seen that the fault estimation
performance is satisfactory. All faults are estimated in an unbiased
 600 LU ET AL.

sense including the oscillatory faults. The estimation error of fV , fα ,
and fβ is shown in Fig. 5f. It can be seen that the error is zero mean,
which confirms the good estimation performance of the ATS-UKF.
IV. Performance Validation Using Real Flight Data
In the preceding section, the FDD performance of the ATS-UKF
is tested using simulated aircraft data. In this section, the FDD
performance of the ATS-UKF is validated using real flight-test data
of the Cessna Citation II aircraft. Air data information such as α and β
are measured for postflight analysis. The real flight data contain
uncertainties such as biases and spikes. Additionally, in real flight,
external disturbances, such as changing wind, can also influence the
air data measurements. Therefore, the real flight data pose challenges
to the ADS FDD problem and provide a realistic validation of the
performance of FDD approaches such as the ATS-UKF.
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313
The primary objective of the flight test is aerodynamic model
identification where a number of maneuvers were performed by the
aircraft to obtain sufficient excitation. Because there were no faults
during the flight, sensor faults are injected into the real flight data to
validate the FDD performance of the ATS-UKF. Besides the fault
scenarios presented in Tables 1 and 2, a fault-free case is also studied.
The real flight data used in this paper are the same as that in Lu and
van Kampen [26]. In [26], the estimated wind turns out to be time
varying. This can test the ADS FDD performance of the ATS-UKF
under the condition of winds.
The update rates of the onboard sensors are given in Table 3.
A. Real-Life Measurement Model
For simulated aircraft data, the measurement model is given in
Eqs. (19–24). If f  0, the measurements are only corrupted by

true
108 0.5 estimation
V (m/s)

f (m/s)
106 0

V
104 −0.5
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.1 0.1
α (rad)

f (rad)
0.08 α
0

0.06 −0.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.04 0.01
0
β (rad)

f (rad)

0.02 −0.01
β

−0.02
0 −0.03
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

a) Estimation of V, , and using the RTS-UKF b) Estimation of fV, f , and f using the RTS-UKF

108 0.1
φ (rad)
V (m/s)

106 0

104 −0.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.1 0.1
θ (rad)
α (rad)

0.05 0.05

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.05 4.7
ψ (rad)
β (rad)

−0.05 4.65
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

c) Estimation of V, , and using the ATS-UKF d) Estimation of , , and using the ATS-UKF

true
1
2 estimation
f (m/s)
V

0.5
FA

0
V

0
−2
0 10 20 30 40 50 60
0 10 20 30 40 50 60
1 0.1
f (rad)
α

0.5
FA

0
α

0
−0.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
1 0.05
f (rad)
β

0.5
FA

0
β

0 −0.05
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

e) Fault detection and isolation using the ATS-UKF f) Weighted estimation of fV, f , and f using the ATS-UKF
Fig. 7 State estimation and ADS FDD of real-life measurement model of aircraft using RTS-UKF and ATS-UKF approaches in absence of faults.
 LU ET AL. 601

white Gaussian noises, as can be seen from the equations. However, xβ r zβ p

this is never the case in real life. In this flight test, the air data βm  Cβ0 
1  Csi β −   vβ (70)
V V
information, such as α and β, are measured by multiple vanes on a
boom (shown in Fig. 6), which is mounted on the nose of the aircraft.
The angle of attack and angle of sideslip measured by the vanes are
denoted by αvm and βvm , respectively. The measurements αvm and βvm ϕm  ϕ  v ϕ (71)
are different from Eqs. (20) and (21), respectively. The measurement
model for the real-life measurements is given as follows [32–34]:

V m  V  vV (68)
θ m  θ  vθ (72)

xα q
αm  Cα0 
1  Cup α   vα (69) ψ m  ψ  vψ (73)
V
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

108 true
5 estimation
V (m/s)

fV (m/s)
106
0

104
0 10 20 30 40 50 60 −5
0 10 20 30 40 50 60
0.1 0.2
α (rad)

fα (rad)
0.08 0

0.06 −0.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.04 0.05
β (rad)

fβ (rad)

0.02 0

0 −0.05
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

a) Estimation of V, , and using the RTS-UKF b) Estimation of fV, f , and f using the RTS-UKF

108 0.1
V (m/s)

φ (rad)

106 0

104 −0.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.1 0.1
α (rad)

θ (rad)

0.05 0.05

0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
0.05 4.7
ψ (rad)
β (rad)

−0.05 4.65
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)
c) Estimation of V, , and using the ATS-UKF d) Estimation of , , and using the ATS-UKF

OFC
true
1
2 estimation
fV (m/s)
FA V

0.5 0
0 −2
0 10 20 30 40 50 60 0 10 20 30 40 50 60
1 0.1
f (rad)
FA α

0.5 0
α

0
−0.1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
1 0.05
f (rad)
FA β

0.5 0
β

0 −0.05
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (s) time (s)

e) Fault detection and isolation using the ATS-UKF f) Weighted estimation of fV, f , and f using the ATS-UKF
Fig. 8 State estimation and ADS FDD of real-life measurement model of aircraft using RTS-UKF and ATS-UKF approaches in presence of multiple
faults.
 602 LU ET AL.

where xα , xβ , and zβ are the position of the vanes in the body frame,
and Cα0 , Cβ0 , Cup , and Csi are the boom correction parameters. In this
paper, zβ is assumed to be zero. The parameter estimation can be
found in Lu et al. [34]. For the ADS FDD using real flight data, this
measurement model is used. However, it should be noted that, in this
real-life measurement model, boom bending is considered to be
negligible for the maneuvers flown.
B. ADS FDD Using Real Flight Data in the Absence of Faults
Using the real-life measurements, uncertainties and disturbances
such as varying winds can also have a negative influence on the FDD
performance, which can result in false alarms. Under this condition,
the FDD approaches should not give false alarms. In this section, the
ATS-UKF is tested in a fault-free case to verify whether it gives false
alarms.
To show the effectiveness of the ATS-UKF, the RTS-UKF is also
applied to estimate the ADS faults. The initial condition x0 given for
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313
C. ADS FDD Using Real Flight Data in the Presence of Multiple
Faults
In this subsection, the ADS FDD performance of the ATS-UKF
will be verified using the real-life measurement model in the presence
of multiple faults (given in Table 1). The initial condition for RTS-
UKF is the same as in Eq. (74) and that of the ATS-UKF is the same as
in Eq. (75). The results using these two approaches are shown
in Fig. 8.
The estimation of V, α, and β using the RTS-UKF is shown in
Fig. 8a. The fault estimation is shown in Fig. 8b. Although the initial
condition of the RTS-UKF is chosen to be the measurements, the
estimation of the faults is still biased. This shows the drawback of the
RTS-UKF when it is used in practice because the initial condition
error will be estimated as a bias fault.
The estimation of V, α, and β and ϕ, θ, and ψ using the ATS-UKF is
presented in Figs. 8c and 8d, respectively. It can be seen that the
estimates of α and β are again different from those of the RTS-UKF

the RTS-UKF is the first measurement which is shown in Fig. 8a.

104.8733; 0.0796; 0.0073; −0.0019; 0.0733; 4.6692T
The initial condition x0 given for the ATS-UKF is
1; 0; 0; 0; 0; 0T
In this manner, the initial condition given for the RTS-UKF is close
to the true state, whereas that of the ATS-UKF significantly deviates
from the true state. The threshold to stop the initial measurement
update is η  5 × 10−3 ; 2 × 10−5 ; 2 × 10−5 T and the threshold to
detect the fault is T  0.2; 1 × 10−4 ; 5 × 10−5 T , which are the same
as those used in the preceding section.
The estimation of V, α, and β using the RTS-UKF is given in
Fig. 7a. The fault estimation using the RTS-UKF is given in Fig. 7b.
As can be seen, the estimated faults deviate from their true
magnitudes. This result shows that the RTS-UKF is not able to be
applied for real applications unless modifications are made.
The results of the ATS-UKF are shown in Figs. 7c–7f. The
estimation of V, α, and β and ϕ, θ, and ψ are shown in Figs. 7c and 7d,
respectively. The estimates of α and β using the ATS-UKF are
different from those using the RTS-UKF. The fault detection result,
shown in Fig. 7e, indicates that there are no faults. This demonstrates
that the ATS-UKF does not give false alarms in the presence of no
faults even when the real flight data are used. The weighted estimates
of fV , fα , and fβ , shown in Fig. 7f, are zero mean. This confirms the
fault estimation performance of the ATS-UKF.
The fault detection and isolation using the ATS-UKF is given in
Fig. 8e. No false alarms are generated and the isolation is also correct.
It can be seen that the performance is as good as that in Fig. 4b where
(74)
the simulation data are used. The oscillatory fault is also detected,
which is shown by the dashed line.
The weighted estimation of fV , fα , and fβ using the ATS-UKF is
presented in Fig. 8f. Even though the initial condition of the ATS-
UKF deviates from the true state significantly, its performance is not
sensitive to the initial condition. Because the faults are estimated in an
(75) unbiased sense, the estimates of α and β using the ATS-UKF are more
reliable than those using the RTS-UKF.
D. ADS FDD Using Real Flight Data in the Presence of Simultaneous
Faults
In this subsection, simultaneous faults (given in Table 2) are
injected into the real flight data to validate the performance of the
ATS-UKF. The result of the RTS-UKF is also presented, which is
given in Fig. 9. From this figure, it is seen that the fault estimation of
the RTS-UKF is again biased, although the initial x0 is chosen to be
the measurements. This highlights the limitation of the RTS-UKF
when used in reality.
The results using the ATS-UKF are shown in Fig. 10. The state
estimation is presented in Figs. 10a and 10b, respectively.
The fault detection and isolation using the ATS-UKF is shown in
Fig. 10c. No false alarms are observed from the figure and the
oscillatory faults are also detected. The performance is as good as that
in Fig. 5c, in which the simulation data are used. This confirms the
FDI performance of the ATS-UKF.
Estimation and weighted estimation of fV , fα , and fβ using the
ATS-UKF are shown in Figs. 10d and 10e, respectively. All fault
estimates achieve an unbiased estimation. The fault estimation errors

true
estimation 0.5
Δ f (m/s)

2
f (m/s)

0 0
V
V

−2
−0.5
0 10 20 30 40 50 0 10 20 30 40 50
0.2 0.06
Δ f (rad)
fα (rad)

0.1
0.05
α

0
−0.1 0.04
0 10 20 30 40 50 0 10 20 30 40 50
0.1 0.04
Δ f (rad)
f (rad)

0
0.02
β

−0.1
β

−0.2 0
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)
a) Estimation of fV, f , and f b) Error of estimation of fV, f , and f
Fig. 9 ADS FDD of real-life measurement model of aircraft using RTS-UKF approach in presence of simultaneous faults.
 LU ET AL. 603

108 0.02

V (m/s)

φ (rad)
106 0

104 −0.02
0 10 20 30 40 50 0 10 20 30 40 50
α (rad) 0.1 0.1

θ (rad)
0.05

0 0.05
0 10 20 30 40 50 0 10 20 30 40 50
0.05 4.68

ψ (rad)
β (rad)

0 4.67

−0.05 4.66
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)
a) Estimation of V, , and b) Estimation of , , and
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313

OFC true
1 5 estimation

f (m/s)
V
FA

0.5 0

V
0 −5
0 10 20 30 40 50 0 10 20 30 40 50
1 0.1

f (rad)
FA α

0.5 0
α
0 −0.1
0 10 20 30 40 50 0 10 20 30 40 50
1 0.1
f (rad)
FA β

0.5 0
β

0 −0.1
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)

c) Fault detection and isolation d) Estimation of fV, f , and f

true
2 estimation 1
f (m/s)

Δ f (m/s)

0
0
V

−2
0 10 20 30 40 50 −1
0 10 20 30 40 50
0.1
0.05
fα (rad)

Δ f (rad)

0 0
α

−0.1 −0.05
0 10 20 30 40 50 0 10 20 30 40 50
0.1 0.02
Δ f (rad)
f (rad)

0 0
β
β

−0.1 −0.02
0 10 20 30 40 50 0 10 20 30 40 50
time (s) time (s)

e) Weighted estimation of fV, f , and f f) Error of estimation of fV, f , and f

Fig. 10 State estimation and ADS FDD of real-life measurement model of aircraft using ATS-UKF in presence of simultaneous faults.

are demonstrated in Fig. 10f. Although the errors are not zero mean,
they are small compared with the states.
V. Conclusions
This paper deals with the ADS FDD of aircraft. First, the RTS-KF
is extended to the RTS-UKF to cope with nonlinear systems. Second,
the RTS-UKF is found to be sensitive to the initial condition. The
problem is analyzed theoretically, and subsequently, a novel ATS-
UKF is proposed to detect, isolate, and estimate the ADS faults.
The ATS-UKF contains three steps: time update, fault estimation,
and measurement update. The ATS-UKF is validated using simulated
aircraft data, which show good FDD performance.
The performance of the ATS-UKF is further validated using real
flight data of the Cessna Citation II aircraft to test its performance
under real-life uncertainties. It was found that, although the measure-
ment data contain biases that cannot be removed completely and the
initial condition is far from the true state, the ATS-UKF is still able to
maintain its satisfactory FDD performance. This demonstrates that it
can be applied in practice.
This ATS-UKF, which deals with ADS FDD, can be incorporated
into fault-tolerant control systems to further enhance the safety of the
aircraft. It can detect faults without giving false alarms. In addition, it
can provide both unbiased state estimation and fault estimation,
which are important for the recovery from sensor faults.
In the future, the ATS-UKF should be integrated into a fault-
tolerant control system. Finally, it is highly recommended that a
 604 LU ET AL.
real-world flight experiment is designed and executed to detect and
estimate ADS faults in aircraft during flight.
References
[1] Patton, R. J., “Fault-Tolerant Control Systems: The 1997 Situation,”
Proceedings of IFAC Symposium on Fault Detection, Supervision and
Safety for Technical Processes, International Federation of Automatic
Control, Laxenburg, Austria, 1997, pp. 1033–1054.
[2] Chen, J., and Patton, R. J., Robust Model-Based Fault Diagnosis for
Dynamic Systems, Kluwer Academic, Norwell, MA, 1999, pp. 19–21.
[3] Isermann, R., “Model-Based Fault-Detection and Diagnosis–Status
and Applications,” Annual Reviews in Control, Vol. 29, No. 1, 2005,
pp. 71–85.
doi:10.1016/j.arcontrol.2004.12.002.
[4] Venkatasubramanian, V., Rengaswamy, R., and Yin, K., “A Review of
Process Fault Detection and Diagnosis Part I: Quantitative Model-Based
Methods,” Computers and Chemical Engineering, Vol. 27, No. 3,
March 2003, pp. 293–311.
Downloaded by TECHNISCHE UNIVERSITEIT DELFT on August 17, 2016 | http://arc.aiaa.org | DOI: 10.2514/1.G001313
doi:10.1016/S0098-1354(02)00160-6
[5] Marzat, J., Piet-Lahanier, H., Damongeot, F., and Walter, E., “Model-
Based Fault Diagnosis for Aerospace Systems: A Survey,” Journal of
Aerospace Engineering, Vol. 226, No. 10, 2012, pp. 1329–1360.
doi:10.1177/0954410011421717
[6] Hajiyev, C., and Caliskan, F., Fault Diagnosis and Reconfiguration in
Flight Control Systems, Kluwer Academic, Norwell, MA, 2003, pp. 1–
24.
[7] Hajiyev, C., and Soken, H. E., “Robust Adaptive Kalman Filter for
Estimation of UAV Dynamics in the Presence of Sensor/Actuator
Faults,” Aerospace Science and Technology, Vol. 28, No. 1, 2013,
pp. 376–383.
doi:10.1016/j.ast.2012.12.003
[8] Goupil, P., “AIRBUS State of the Art and Practices on FDI and FTC in
Flight Control System,” Control Engineering Practice, Vol. 19, No. 6,
2011, pp. 524–539.
doi:10.1016/j.conengprac.2010.12.009
[9] Zolghadri, A., “Advanced Model-Based FDIR Techniques for Aero-
space Systems: Today Challenges and Opportunities,” Progress in
Aerospace Sciences, Vol. 53, Aug. 2012, pp. 18–29.
doi:10.1016/j.paerosci.2012.02.004
[10] Zolghadri, A., Henry, D., Cieslak, J., Efimov, D., and Goupil, P., Fault
Diagnosis and Fault-Tolerant Control and Guidance for Aerospace
Vehicles, From Theory to Application, Advances in Industrial Control,
Springer–Verlag, Berlin, 2013, pp. 1–22.
[11] Houck, D., and Atlas, L., “Air Data Sensor Failure Detection,” AIAA/
IEEE/SAE 17th Digital Avionics Systems Conference, IEEE Publ.,
Piscataway, NJ, 1998, pp. 1–8.
[12] Freeman, P., Seiler, P., and Balas, G. J., “Air Data System Fault
Modeling and Detection,” Control Engineering Practice, Vol. 21,
No. 10, 2013, pp. 1290–1301.
doi:10.1016/j.conengprac.2013.05.007
[13] Eubank, R. D., Atkins, E. M., and Ogura, S., “Fault Detection and Fail-
Safe Operation with a Multiple-Redundancy Air-Data System,” AIAA
Guidance, Navigation, and Control Conference, AIAA, Reston, VA,
2010, pp. 1–14; also AIAA Paper 2010-7855, 2010.
[14] Lombaerts, T., Fault Tolerant Flight Control–A Physical Model
Approach, Ph.D. Thesis, Delft Univ. of Technology, Delft, The
Netherlands, 2010.
[15] Ellsworth, J. C., and Whitmore, S. A., “Reentry Air Data System for
a Sub-Orbital Spacecraft Based on X-34 Design,” AIAA Aerospace
Sciences Meeting and Exhibit, AIAA, Reston, VA, 2007, pp. 1–31; also
AIAA Paper 2007-1200, 2007.
[16] Rohloff, T. J., Whitmore, S. A., and Catton, I., “Fault-Tolerant Neural
Network Algorithm for Flush Air Data Sensing,” Journal of Aircraft,
Vol. 36, No. 3, 1999, pp. 541–549.
doi:10.2514/2.2489
[17] Nebula, F., Palumbo, R., Morani, G., and Corraro, F., “Virtual Air Data
System Architecture for Space Reentry Applications,” Journal of Space-
craft and Rockets, Vol. 46, No. 4, 2009, pp. 818–828.
doi:10.2514/1.42485
[18] Nebula, F., Palumbo, R., and Morani, G., “Virtual Air Data: A Fault-
Tolerant Approach Against ADS Failures,” AIAA Guidance, Navigation,
View publication stats
and Control and Co-Located Conferences, AIAA, Reston, VA, 2013,
pp. 1–14.
[19] Looye, G., and Joos, H. D., “Design of Robust Dynamic Inversion
Control Laws Using Multi-Objective Optimization,” AIAA Guidance,
Navigation, and Control Conferences and Exhibit, AIAA, Reston, VA,
2001, pp. 1–10; also AIAA Paper 2001-4285, 2001.
[20] Cervia, F., Denti, E., Galatolo, R., and Schettini, F., “Air Data
Computation in Fly-by-Wire Flight Control Systems,” Journal of
Aircraft, Vol. 43, No. 2, 2006, pp. 450–455.
doi:10.2514/1.16270.
[21] Van Eykeren, L., and Chu, Q., “Fault Detection and Isolation for Inertial
Reference Units,” AIAA Guidance, Navigation, and Control Confer-
ence, AIAA, Reston, VA, 2013, pp. 1–10; also AIAA Paper 2013-4697,
2013.
[22] Lu, P., Van Eykeren, L., van Kampen, E., Chu, Q., and Yu, B., “Adaptive
Hybrid Unscented Kalman Filter for Aircraft Sensor Fault Detection,
Isolation and Reconstruction,” AIAA Guidance, Navigation, and Con-
trol Conference, AIAA, Reston, VA, 2014, pp. 1–18; also AIAA Paper
2014-1145, 2014.
[23] Van Eykeren, L., Chu, Q., and Mulder, J. A., “Sensor Fault Detection
and Isolation Using Adaptive Extended Kalman Filter,” Eighth IFAC
Symposium on Fault Detection, Supervision and Safety of Technical
Processes, International Federation of Automatic Control, 2012,
pp. 1155–1160.
doi:10.3182/20120829-3-MX-2028.00195
[24] Lu, P., Van Eykeren, L., van Kampen, E., and Chu, Q., “Selective-
Reinitialization Multiple-Model Adaptive Estimation for Fault Detec-
tion and Diagnosis,” Journal of Guidance, Control, and Dynamics,
Vol. 38, No. 8, 2015, pp. 1409–1424.
doi:10.2514/1.G000587
[25] Gillijns, S., and De Moor, B., “Unbiased Minimum-Variance Input
and State Estimation for Linear Discrete-Time Systems with Direct
Feedthrough,” Automatica, Vol. 43, No. 5, 2007, pp. 934–937.
doi:10.1016/j.automatica.2006.11.016
[26] Lu, P., and van Kampen, E., “Aircraft Inertial Measurement Unit Fault
Identification with Application to Real Flight Data,” AIAA Guidance,
Navigation and Control Conference, AIAA, Reston, VA, 2015, pp. 1–
17; also AIAA Paper 2015-0859, 2015.
[27] Julier, S., and Uhlmann, J., “A New Extension of the Kalman Filter to
Nonlinear Systems,” Proceedings of AeroSense: The 11th International
Symposium on Aerospace/Defence Sensing, Simulation and Control,
International Soc. for Optics and Photonics, Bellingham, WA, 1997,
pp. 182–193.
[28] Van Der Merwe, R., and Wan, E. A., “The Square-Root Unscented
Kalman Filter for State and Parameter-Estimation,” IEEE International
Conference on Acoustics, Speech, and Signal Processing, IEEE Publ.,
Piscataway, NJ, 2001, pp. 3461–3464.
[29] Mehra, R. K., “Approaches to Adaptive Filtering,” IEEE Transactions
on Automatic Control, Vol. 17, No. 5, 1972, pp. 693–698.
doi:10.1109/TAC.1972.1100100
[30] Soken, H. E., and Hajiyev, C., “Pico Satellite Attitude Estimation via
Robust Unscented Kalman Filter in the Presence of Measurement
Faults,” ISA Transactions, Vol. 49, No. 3, 2010, pp. 249–256.
doi:10.1016/j.isatra.2010.04.001
[31] Goupil, P., “Oscillatory Failure Case Detection in the A380 Electrical
Flight Control System by Analytical Redundancy,” Control Engineer-
ing Practice, Vol. 18, No. 9, 2010, pp. 1110–1119.
doi:10.1016/j.conengprac.2009.04.003
[32] Mulder, J., Chu, Q., and Sridhar, J., “Non-Linear Aircraft Flight Path
Reconstruction Review and New Advances,” Progress in Aerospace
Sciences, Vol. 35, No. 7, Oct. 1999, pp. 673–726.
doi:10.1016/S0376-0421(99)00005-6
[33] De Visser, C. C., “Global Nonlinear Model Identification with
Multivariate Splines,” Ph.D. Thesis, Delft Univ. of Technology, Delft,
The Netherlands, 2011.
[34] Lu, P., van Eykeren, L., van Kampen, E., de Visser, C. C., and Chu, Q.,
“Double-Model Adaptive Fault Detection and Diagnosis Applied to
Real Flight Data,” Control Engineering Practice, Vol. 36, March 2015,
pp. 39–57.
doi:10.1016/j.conengprac.2014.12.007