1168 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO.

4, JULY 2013

Fault-Tolerant Control of Wind Turbines:

A Benchmark Model
Peter Fogh Odgaard, Member, IEEE, Jakob Stoustrup, Senior Member, IEEE,
and Michel Kinnaert, Member, IEEE

Abstract— This paper presents a test benchmark model for wind turbines, but they only cover parts of the wind turbine
the evaluation of fault detection and accommodation schemes. and do not include comparisons of the performance of different
This benchmark model deals with the wind turbine on a system schemes for the detection of faults in the various parts of the
level, and it includes sensor, actuator, and system faults, namely
faults in the pitch system, the drive train, the generator, and the turbine.
converter system. Since it is a system-level model, converter and In [1], a Kalman-filter-based diagnosis system to detect
pitch system models are simplified because these are controlled by faults in the blade root bending moment sensors was presented.
internal controllers working at higher frequencies than the system An unknown input observer was designed for the detection of
model. The model represents a three-bladed pitch-controlled sensor faults around the wind turbine drive train in [2]. In [3],
variable-speed wind turbine with a nominal power of 4.8 MW.
The fault detection and isolation (FDI) problem was addressed by active and passive fault-tolerant control (FTC) schemes were
several teams, and five of the solutions are compared in the second applied to a wind turbine model. More focus has been placed
part of this paper. This comparison relies on additional test data on the electrical conversion system in the wind turbines; some
in which the faults occur in different operating conditions than relevant examples can be found in [4] and [5]. In the former,
in the test data used for the FDI design. an observer-based solution for current sensor fault detection
Index Terms— Benchmark modeling, fault detection, fault is presented, whilst the latter presents a similar solution for
isolation, fault-tolerant control (FTC), wind turbines. voltage sensor fault detection. In [6], a fault detection and
reconfiguration solution for handling faults in a doubly fed
I. I NTRODUCTION wind turbine converter is presented.
Comparing the various detection and accommodation
T ODAY, wind turbines contribute to a large part of the
world’s power production. At the same time, the size of
the standard turbine tends to increase as well. Turbines in
schemes on the wind turbine application is beneficial in the
process to find the best schemes to handle the different faults.
the megawatt size are expensive, and hence their reliability is However, since a wind turbine is a large and complex system,
expected to be high to generate as much energy as possible. such a comparison can only be performed on a limited set of
These wind turbines are expected to produce energy with possible faults. In order to test various detection, isolation, and
very short downtimes. A way to ensure this consists in intro- accommodation schemes on the wind turbine application, this
ducing advanced fault detection, isolation, and accommodation paper presents a benchmark model of a wind turbine at the
systems into the turbines. In the state-of-the-art industrial wind system level, containing sensor, actuator, and systems faults.
turbines, fault detection schemes are simple and are often This benchmark model describes a realistic generic three-
conservative, and so is the fault accommodation mechanism. blade horizontal variable-speed wind turbine with a full-scale
Turbines are turned off even during simple faults to wait for converter coupling. This generic turbine has a rated power at
service. [In addition to the fault detection and isolation (FDI) 4.8 MW. Since this model works at the system level, the fast
systems, condition monitoring is used for the rotating parts.] control loops of the converters are not considered.
Consequently, there is a need to use advanced fault detection, This wind turbine FDI and FTC benchmark model was
isolation, and accommodation schemes in order to improve originally presented in [7]. The purpose of this benchmark
the on-time of the turbine, even though that might result in was to provide a model on which researchers working in the
limited power production for some faults. Some work has been field of fault diagnosis and fault-tolerant control can compare
performed on fault detection, isolation, and accommodation on different methods in their field applied to a wind turbine. Wind
turbines are complicated machines; it was therefore decided to
Manuscript received December 2, 2011; revised September 19, 2012 and keep the benchmark model simple so that non wind-turbine
January 29, 2013; accepted March 26, 2013. Manuscript received in final form experts can use it. Consequently, some simplifications and
April 18, 2013. Date of publication May 22, 2013; date of current version
June 14, 2013. The work of P. F. Odgaard was supported by kk-Electronic assumptions are made. Blades and tower are assumed rigid,
a/s. Recommended by Associate Editor G. E. Stewart. and aerodynamics are described by a static model. The wind
P. F. Odgaard and J. Stoustrup are with Aalborg University, Aalborg 9220, turbine controller included in the model is also simple, leaving
Denmark (e-mail: [email protected]; [email protected]).
M. Kinnaert is with the Department of Control Engineering and System out some typical features; however, it controls the wind turbine
Analysis, Université libre de Bruxelles, Bruxelles 1050, Belgium (e-mail: with acceptable performance. In this paper, the model is
[email protected]). extended with even more faults and test sets, and is described
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. in more detail, so that a better understanding of the benchmark
Digital Object Identifier 10.1109/TCST.2013.2259235 model is provided. Additional test sets are introduced to test
1063-6536/$31.00 © 2013 IEEE
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1169

Blade
Pitch Actuator

Anemometer
Gear
Main Axis Generator Converter
box

Hub Nacelle
Transformer

Fig. 2. Major parts from inside of the nacelle. It can be seen that the blades
Tower
are fixed to the main axis, which in turn is connected to the generator through
the gear box. The generator is electrically connected to the converter, which
in turn is connected to a transformer. The blades are pitched by the pitch
actuators.

turbine is shown in Fig. 1, from which it can be seen that

the nacelle is mounted at the top of a tower. Inside the nacelle
lies the control equipment and the equipment to convert energy
from the rotating mechanical system to electrical energy. The
blades are fixed to the main axis (not seen in the figure) in
Fig. 1. General outline of the wind turbine seen from the outside.
the hub. The anemometer is placed on the top of the nacelle
and measures the wind speed and direction. Fig. 2 represents
the robustness of solutions toward different points of operation the interior of the nacelle, showing that the main axis is
at which faults are introduced. This paper also presents some connected to the generator through a gear box; the generator is
of the best FDI solutions applied to this benchmark model. connected to the grid through the converter and a high-voltage
In these tests, an internal model from kk-electronic a/s is transformer. The general functioning starts with pressure from
used, which includes some additional model features to test the wind on the turbine blades, forcing the wind turbine rotor
further robustness of the proposed schemes. The compared to spin around. Energy is converted from kinetic wind energy
solutions can be seen in [8]–[12]. These solutions are evaluated to mechanical energy through a rotating shaft. This energy
and compared based on the requirements in the benchmark generation can be controlled by changing the aerodynamics
statement. Experiences gained by usage of the methods on the of the rotor. This is achieved by pitching the blades or by
benchmark are provided as well. A number of other solutions controlling the rotational speed of the rotor relative to the wind
have also been applied to this benchmark model. Among these speed. Mechanical energy is converted to electrical energy by a
are [13]–[27]. generator coupled to a converter with rating equal to the rated
The rest of this paper is organized as follows. In Section II, power of the generator. Between the rotor and the generator,
the functionality of wind turbines is described; the fault a gear box is used to transform the rotational speed from the
scenarios are presented in Section III. Next, in Section IV, the rotor to the generator. The converter can be used to set the
wind turbine model is presented. In Section V, the test signals generator torque, which consequently can be used to control
are characterized. The different FDI solutions considered and the rotational speed of the generator as well as the rotor.
compared in this paper are introduced in Section VI. These A more detailed description of the general functioning of the
solutions are compared in Section VII. Conclusions are drawn wind turbine can be seen in [28]–[30].
in Section VIII. The objective of the control system is to follow the power
reference. In case the wind speed is too low for the wind
turbine to reach the power reference, the controller will
II. W IND T URBINE D ESCRIPTION try to optimize the power production. This power control
Wind turbines generate electrical energy from the kinetic should keep the mechanical vibrations at an acceptable level.
energy in the wind. In this test bench model, a specific A system overview can be seen in Fig. 3. This figure shows
kind of turbine is considered. This turbine is a three-blade the relations between the blade and the pitch system, the drive
horizontal-axis turbine with a full converter coupling, even train, the generator and converter, and the controller. The
though, at this system level, the difference is small between variables between these subsystems are defined as follows:
a full converter and a doubly fed induction generator. Both v w is the wind speed acting on the turbine blades; τw
these generator types are variable-speed and pitch-controlled is the torque from the wind acting on the turbine blades;
turbines. More details on wind turbine generator types can τr is the rotor torque; ωr is the rotational speed of the rotor;
be found in [28] and [29]. The physical outline of the wind τg is the generator torque; ωg is the rotational speed of the
1170 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

τr τg
vw Blade & Generator &
Drive Train Zone 1 Zone 4
Pitch System Converter Zone 2 Zone 3
ωr ωg
βr βm , vw,m ωr,m τg,m ,ωg,m ,Pg τg,r 100%

Relative Power [%]

Controller

Pr

Fig. 3. Overview of the benchmark model. It consists of four parts: blade

and pitch systems, drive train, generator and converter, and controller. The
variables in the figure are defined in the text.

0 3 12.5 25
generator; βr is the reference to the pitch position; βm is the Wind Speed [m/s]

measured pitch position; τw,m is an estimated wind torque

Fig. 4. Illustration of the reference power curve for the wind turbine
based on wind speed measurement; ωr,m is the measured depending on the wind speed. In zone 1, the wind turbine will be idling
rotational speed of the rotor, ωg,m is the measured rotational awaiting higher wind speeds (0–3 m/s). In zone 2, the generated power of the
speed of the generator; τg,m is the measured generator torque; wind turbine will be optimized (3–12.5 m/s). In zone 3, the wind turbine will
be controlled to keep a constant power generation (12.5–25 m/s). In zone 4,
τg,r is the torque reference to the generator; Pr is the power the wind turbine will be parked, preventing damage due to the high wind
reference to the wind turbine; and Pg is the power produced speed (above 25 m/s).
by the generator. The wind turbine controller provides three
pitch references, and all three pitch positions are measured blade β is equal to 0 degrees, and the tip speed ratio is kept
with two sensors to ensure physical redundancy of the pitch constant at its optimal value. The tip speed ratio λ is defined
position measurements. The generator and rotor speeds are as
ωr · R
also measured with two sensors each for the same reason. λ= (1)
These variables are defined as βr1 , βr2 , and βr3 for the pitch vw
reference to blades 1, 2, and 3. β1,m1 and β1,m2 are the two where R is the radius of the blades, and v w is the wind
pitch measurements for blade 1; β2,m1 and β2,m2 are the speed. The optimal power curve is followed by controlling
two pitch measurements for blade 2; and β3,m1 and β3,m2 the generator torque reference τg,r to its optimal value.
are the two pitch measurements for blade 3. The two rotor When the nominal power reference is achieved, the
speed measurements are defined as ωr,m1 and ωr,m2 ; the two controller is switched to control zone 3. In this zone, the
generator speed measurements are defined as ωg,m1 and ωg,m2 . control objective is to maintain the power reference Pr . This
objective is obtained by controlling βr , such that the power
A. Control System Concept coefficient of the wind turbine CP is decreased. In an industrial
control scheme, a proportional integral (PI) controller is often
The controller operates in principle in four operational zones
used to keep ωr at the rated value by changing βr .
governed by the mean wind speed within some time window.
More details on the wind turbine controller can be found in
These control zones are plotted depending on the wind speed
Section IV-E.
in Fig. 4. In zone 1, the turbine is at standstill; zone 2 is
power optimization or partial load; zone 3 is constant power
production; zone 4 is high wind speed. In zone 4, the turbines III. FAULT S CENARIOS
are pitched out of the wind to stop the rotation and the energy In this benchmark model, some faults are considered. These
production of the turbine. Zones 1 and 4 are not considered faults cover sensor, actuator, and process faults in different
any further in this paper and are excluded from the benchmark parts of the wind turbine. In the following, these different
model, since the focus is on detecting faults under normal kinds of faults are listed. These faults have various degrees
operations, which are covered by zones 2 and 3 [31]. An of severity. Some are very serious and must result in a fast
extra control zone is often used between zones 2 and 3, and safe closing down of the wind turbine. The rest are less
providing a smoother transition between the control zones and severe in the sense that the controller can accommodate these
placing lower structural loads on the wind turbine. This will faults while staying in operation, possibly with a reduced
influence the FDI scheme working on the wind turbine, since performance. Both additive and multiplicative faults as well as
there is a correlation between FDI and the controls of the faults resulting in changed dynamics of parts of the system are
wind turbine. However, the proposed control structure gives a considered in this benchmark model. The faults are chosen to
more challenging detection and isolation problem since larger cover different kinds of faults so they require different qualities
transients are present. Therefore this simple control structure in the FDI and FTC schemes. All the described faults originate
is relevant in the benchmark model. from actual faults in wind turbines, but the details of the actual
In Fig. 4, the power curve for the wind turbine is plotted. faults cannot be presented for confidentiality reasons. To sum
This figure shows that for wind speeds between 0 and 12.5 m/s up, the faults are selected to cover different parts of the wind
the turbine is controlled to obtain optimal power production. turbine, different fault types and classes, and different levels
The optimal power is obtained when the pitch angle of the of severity.
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1171

A. Sensor Faults with time. This change will result in two correlated fault
Some of the more relevant faults are considered in signals: ωr and ωg . These changes evolve slowly in reality;
this benchmark model. The first is a fault in the however, it is expected to be too demanding in the benchmark
pitch position measurements. These faults are denoted simulator from a computational point of view to simulate such
β1,m1 , β1,m2 , β2,m1 , β2,m2 , β3,m1 , and β3,m2 . The a fault realistically evolving over a period of months or even
origin of these faults is either electrical or mechanical faults years. Consequently, in this benchmark model, this fault is
in the position sensors, and can result in either a fixed value represented by a small change of the friction coefficient within
or a changed gain factor on the measurements. These two a few seconds. The main point is that the change in the drive
fault types are considered since they have been seen in wind train friction is much slower than the system dynamics and
turbines and it is important to detect and accommodate these the system sample rate. Typically, faults in wind turbine gear
kinds of faults. Especially, the fixed value should be easy to boxes are found using condition monitoring methods relying
detect, and therefore it is important that a fault detection, on additional sensors that measure accelerations, noise levels,
isolation, and accommodation scheme be able to deal with etc., on the gear box; an extensive review of this can be found
this fault. If not handled correctly, these faults will influence in [32]. It would be more cost efficient if such faults could be
the pitch positions because the internal controller in the pitch detected and isolated using standard measurement in the wind
system controls the pitch positions based on these pitch turbine control system; this fault is therefore included in the
position measurements. benchmark.
Secondly, the rotor speed measurement can be faulty. We
denote the deviations in rotor speed measurement caused by D. Severity of Faults
these faults as ωr,m1 and ωr,m2 . The fault signals for the All the faults mentioned above are summarized and listed
two generator speed measurements are denoted as ωg,m1 and in Table I. In this table, fault details such as the type,
ωg,m2 . Both the rotor and generator speed measurements consequence, severity, and developing time are given. Notice
are done using encoders. Encoder faults can be due to both also that the letter in the fault class indicates a sensor (A),
electrical and mechanical failures, which result in either a fixed actuator (B), or system fault (C).
value or a changed gain factor on the measurements. In case In order to deal with these faults in a prioritized order,
of a fixed value fault, the output of the encoder is not updated the severity and consequences of these considered faults,
with new values. The gain factor fault is introduced when the as well as the time of development of the faults, are
encoder reads more marks on the rotating part than actually listed in Table I. Notice that the severity level of all these
present, which can happen as a result of dirt or other false sensor faults is set low because of the physical redun-
markings on the rotating part. dancy of the sensors, so sensor faults are not a problem
B. Actuator Faults when they are detected fast and the sensor systems are
reconfigured. However, when these faults are not handled,
Both the converter and pitch systems can be faulty.
Converter faults are denoted as τg and can result in either they are critical. Notice that the changed dynamics of
the drive train, due to increased friction, is not that severe,
an offset or in changed dynamics of the converter. The cause
but should be detected as an indication of the wear of the
of this fault is internal to the converter, and is due to either a
drive train. However, over the years this wear and tear of the
fault in the converter electronics or an off-set on the converter
drive train will accumulate, leading to a total breakdown of
torque estimate. The converter controller would typically be
the drive train, which clearly is a highly severe fault, since it
able to detect the faults in the electronics, but the converter
is not only highly costly to replace but it also results in a long
torque offset is difficult to detect internally. Yet, from a wind
downtime of the wind turbine as a consequence.
turbine level, it is possible to detect, isolate, and accommodate
a torque offset because it changes the torque balance in the
wind turbine power train. E. Fault Detection and Isolation Requirements
The pitch system, which in this case is hydraulic, has the The FDI requirements are listed in this subsection. The
possibility of faults on all three blades, and these faults are detection time TD for the respective faults is defined in terms
denoted as β1 , β2 , and β3 . Faults in the pitch system of the sampling time for the control system Ts , which, in this
are considered in the hydraulic system, which can result in case, is equal to 0.01 s.
changed dynamics due to either a drop in pressure in the 1) Time of Detection: The maximum detection times for
hydraulic supply system or high air content in the oil. The all faults are provided in the following. The respective fault
former represents faults such as a leakage in hose, a blocked amplitudes are taken into account in the selection of these
pump, or similar others. There will always be some air content detection requirements. The requirements are set such that
in the hydraulic oil used, the content level will vary, and it is faults have not yet developed into critical faults, but also
not possible to control it well. Air is much more compressible such that detection and isolation is challenging considering
than oil, so it changes the dynamics of the hydraulic actuator. the fault amplitude. All the sensor faults are required to fulfill
TD < 10 · Ts . For the converter faults, the requirement is
C. System Faults TD < 3 · Ts . For the pitch system fault, due to dropped pump
The considered system fault is found in the drive train, pressure (fault B3), TD < 8 · Ts is required to hold, and for air
where the friction coefficient in the model changes slowly in the oil (fault B4), TD < 100 · Ts is required. An increased
1172 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

TABLE I
FAULTS C ONSIDERED IN THE B ENCHMARK M ODEL I NCLUDING FAULT T YPES , S EVERITY, AND T IME OF D EVELOPMENT OF THE C ONSIDERED FAULTS

Fault Consequence
Fault Type Symbols Severity Dev. Time
class
β1,m1 , β1,m2 , β2,m1 , False measurement, reconfigure
A1 Sensor fault Fixed value Low Medium
β2,m2 , β3,m1 , β3,m2 system
β1,m1 , β1,m2 , β2,m1 , False measurement, reconfigure
A2 Sensor fault Gain factor Low Medium
β2,m2 , β3,m1 , β3,m2 system

A3 Sensor fault Fixed value ωr,m1 , ωr,m2 False measurement, reconfigure Low Medium
system

A4 Sensor fault Gain factor ωr,m1 , ωr,m2 False measurement, reconfigure Low Medium
system

A5 Sensor fault Fixed value ωg,m1 , ωg,m2 False measurement, reconfigure Low Medium
system

A6 Sensor fault gain Factor ωg,m1 , ωg,m2 False measurement, reconfigure Low Medium
system

B1 Actuator fault Changed dynamics τg Slow torque control, indicates High Fast
serious problems

B2 Actuator fault Offset τg None optimal power production Medium Fast

B3 Actuator fault Changed dynamics β1 , β2 , β3 (Hydraulics) Problems with pump or leakage, High Medium
slow control actions

B4 Actuator fault Changed dynamics β1 , β2 , β3 (Air in oil) Air in oil, slow control actions Medium Slow

C1 System fault Changed dynamics ωr , ωg Increased level of drive train Medium Very slow
vibrations

drive train friction is just expected to be detected with no aerodynamic mapping might also change with time because
specific time requirement. It is recommended that Monte of debris buildup on the blades. For more information on this
Carlo studies be applied to test that the detection scheme specific problem, consult [31], [33], and [34].
can detect the respective faults with these requirements. One
hundred simulation runs should be performed, with different F. Accommodation Requirements
measurement noise realizations. The benchmark model of this paper contains both faults for
2) False Detections: The number of false detections is which the system can be reconfigured to continue power gener-
required to be kept low, and the interval between false ation and very severe faults that require a safe and fast shut-
detections is required to be larger than 100 000 samples on down of the wind turbine. The last group contains the severe
average. The false detections must last for no more than three faults in the two actuators with fault numbers B1 and B3.
samples. This number corresponds to 4.4 faults on average per All remaining faults must be accommodated in some way,
simulated test series; consequently, a Monte Carlo study with and the wind turbine must continue its operation. In all cases,
100 repetitions will allow 440 false detections in total. detection of faults must be reported to the system operator,
3) Missed Detections: The faults included in this bench- and automatic action is required. In case of a single sensor
mark model are of such a size that they all should be detected. fault, system performance must not deteriorate; in the case of
multiple faults, a mild deterioration of the system performance
4) Issues to be Aware of: A major problem in the wind is accepted. Large transients when accommodating the fault
turbine control system in general is that, seen from a control must be avoided.
theoretical point of view, the wind turbine is driven by a In order to evaluate the power
disturbance, namely the wind. The wind speed is, however,  generation during a fault, a
power generation error sum Pe is defined in
measured to some extent with significant measurement noises   2
added as well as a large risk of an offset. This offset can Pe = Pg [n] − Pr [n] (2)
be calibrated, but it is recommended to be considered in the n∈N
FDI system. The measurement noise is modeled as a Gaussian where
 N is the number of samples in a given faulty mode.
white noise, for which the parameters can be found in the list Pg for the faulty case should be as close as possible to
of parameters (Table VII). the fault-free case. In addition, two constraints should not be
It is important to notice the nonlinearities in the aerody- violated, i.e., Pg [n] < 1.1 · 4.8 [MW], and ωg [n] < 1.1 ·
namics of the turbine as well as the switching control structure. ωnom , where ωnom is given in Table VI. The maximum allowed
The FDI system is expected to be robust toward uncertainties power is given by the generator and converter design, and
in this aerodynamic model, partly because exact measures the maximum allowed value of ωg [n] is given by structural
of this mapping on the specific turbines are difficult. This considerations.
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1173

IV. W IND T URBINE M ODEL

In this section, the various model parts are presented
according to the following order: wind model, blade and 0.5

pitch model, drive train model, generator/converter model, 0

and controller. Finally, the numerical values of the different
Cq
parameters are indicated. Notice here that the benchmark −0.5
model implemented in Simulink can be downloaded by
−1
the URL address http://www.kk-electronic.com / wind -
turbine-control / competition - on - fault-detection/wind-turbine- −1.5
benchmark-model.aspx [35].
−2
0
A. Wind Model 5
−50
0
In this benchmark model, a wind model is included that 10 50
Tip Speed Ratio
describes stochastic wind behavior as well as wind shear 15 100
Pitch Angle
effects (namely variations in the wind speed depending of
height) and tower shadow effects (accounting for the passing Fig. 5. Torque coefficient Cq as a function of the tip speed ratio λ and the
of the blade in front of the wind turbine tower). Such a model pitch angle β.
can be found in [36]. This implementation is used in the
benchmark model. The model will be presented in this paper where
without much explanation since the derivation of the parts can
R 2 − r02
be found in the references. This wind model consists of four ψ = 2a 2
parts: the mean wind (slow wind variations) v m (t); a stochastic (R 2 + r02 )sin(θ̄r,i (t))2 + k 2 )
part v s (t); the wind shear v ws (t) (which is the effect of wind (r 2 − R 2 )(r02 sin(θ̄r,i (t))2 + k 2 )
energy lost at the surface of the earth, resulting normally in υ = 2a 2 k 2 0
R 2 sin(θ̄r,i (t))2 + k 2
an increasing wind speed as the distance to the earth surface
α · (α − 1) · r02
increases); and the tower shadow v ts (t). The combined wind m = 1+
model is given by 8 · H2 
(i − 1) · 2π θr (t) + (i−1)·2π
v w (t) = v m (t) + v s (t) + v ws (t) + v ts (t). (3) θ̄r,i (t) = θr (t) + − floor 3
· 2π
3 2π
In this benchmark model, four different wind speeds are
and the function floor(x) is the largest integer not greater than
required: the wind speed at hub height as measured by the
x, r0 is the radius of the blade hub, and k is an aerodynamic
anemometers and three speeds at the three blade tips, respec-
parameter.
tively, denoted as v hub (t), v w1 (t), v w2 (t), and v w3 (t). The last
three are altered by the wind shear and the tower shadow,
B. Blade and Pitch Model
while the first is not. Consequently, v hub (t) only depends on
the first two terms in (3), while the three blade wind speeds This model is a combination of the aerodynamic model and
vary according to their angular position in the rotor plane. the wind and pitch model. τr is found using the aerodynamic
In terms of mean wind, a slowly varying wind sequence model, which requires inputs from the wind and pitch models.
is obtained by processing a set of measured wind data with 1) Aerodynamic Model: The aerodynamics of the wind
a low-pass filter. The stochastic part of the wind model is turbine is modeled as a torque acting on the blades. This
modeled by Kaimal filters (see [30] for details and [37] for aerodynamic torque can be represented by the following
implementation details). An implementation of this model can expression [31]:
be found in the Simulink implementation of this benchmark ρπ R 3 Cq (λ(t), β(t))v w (t)2
model [35]. τr (t) = (5)
2
The wind shear model is given by in which Cq (λ(t), β(t)) is a mapping of the torque coefficients
 3 
2v m (t) R ·α R4 α−1 depending on the tip speed ratio and the pitch angle. The Cq
v ws,i (t) = · χ + · α · · χ 2
mapping used in this model can be seen in Fig. 5. In order
3 · R2 3·H 4 2 · H2
 to model the three blades to have their own pitch angle value,
2v m (t) R 5 (α 2 − α) · (α − 2)
+ · · · χ 3 (4) a simple model is used. The latter assumes that the torque
3 · R2 5 6 · H3 of each blade is equal to a third of the torque given by the
where χ = cos(θr∗ (t)), and θr∗ is the angular position of the three blades. Consequently, the aerodynamic torque equation
three blades, θr1 (t) = θr (t), θr2 (t) = θr (t) + (2/3)π, and θr3 = becomes
θr (t) + (4/3)π. α and H used in (4) are two aerodynamic ρπ R 3 Cq (λ(t), βi (t))v w,i (t)2
parameters. τr (t) = 1≤i≤3 .
6
The tower shadow model is given by This model is valid for small differences between the
m · θ̄r,i (t) pitch angles. A more detailed model has shown similar
v ts,i (t) = · (ψ + υ) behavior.
3 · r2
1174 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

1
2) Pitch System Model: The hydraulic pitch system is θ̇ (t) = ωr (t) − ωg (t) (10)
modeled as a closed-loop transfer function between the Ng
measured pitch angle βm and its reference βref . βref is the where Jr is the moment of inertia of the low-speed shaft, K dt
input to the closed-loop transfer function and it is provided is the torsion stiffness of the drive train, Bdt is the torsion
by the wind turbine controller; βm is the output of the transfer damping coefficient of the drive train, Bg is the viscous friction
function which is also the measurement used by the internal of the high-speed shaft, Ng is the gear ratio, Jg is the moment
pitch actuator controller. βm is available for the wind turbine of inertia of the high-speed shaft, ηdt is the efficiency of the
control system and therefore also for FDI and FTC schemes. drive train, and θ (t) is the torsion angle of the drive train.
Notice that the control signal from the internal pitch actuator The fault in terms of a lower drive train efficiency is modeled
controller is not available. In principle, it is a piston servo by substituting another parameter, denoted ηdt2 , for ηdt .
system that can be modeled well by a second-order transfer
function [38] D. Generator and Converter Model
β(s) ωn2 The electrical system in the wind turbine and the electrical
= 2 (6)
βr (s) s + 2 · ζ ωn · s + ωn2 system controllers are much faster than the frequency range
where ζ is the damping factor, and ωn is the natural frequency. used in the benchmark model. On a system level of the wind
A transfer function is attached to all three pitch systems. In turbine, the generator and converter dynamics can be modeled
cases of no fault, the damping factors are assumed equal. by a first-order transfer function
However, in case of a fault in a pitch system, the parameters τg (s) αgc
=
might be different from one pitch system to another. In order to τg,r (s) s + αgc
model the hydraulic power drop and increase of air content, the where αgc is the generator and converter model parameter.
parameters in the transfer function are changed during these The power produced by the generator is given by
faults. Notice here that the hydraulic pressure drop is assumed
to be abrupt, whereas the air content increases slowly. The two Pg (t) = ηg ωg (t)τg (t)
transfer function parameters for the pressure drop case are
where ηg is the efficiency of the generator.
denoted ωn2 and ζ2 , and the two parameters for the increased
air content model are denoted ωn3 and ζ3 .
E. Controller
The change in the sensor gain factor induces a change in
the closed-loop pitch actuator as well as a change in the The wind turbine controller in this simulation model works
measured position. Since the closed-loop pitch actuator is in two regions as previously presented. Zone 2 is denoted the
modeled as a linear system, the fault-induced error on the pitch power optimization and zone 3 is denoted power reference
position is moved from the measurement to the reference to following. In the following text, zones 2 and 3, respectively,
the pitch actuator. The controller is fed by the mean value correspond to control modes 1 and 2.
of the readings of the two sensors. Hence, this sensor fault In this benchmark model, a simple control scheme is used
is modeled as a change in the pitch references, meaning that because the focus is on the fault detection and accommodation
a sensor fault resulting in changed mean value should also of the wind turbine. The state-of-the-art industrial controller
change the pitch reference accordingly contains typically a larger number of control modes, which
provides a smoother transition between power optimization
βi,m1 [n] + βi,m2 [n]
βr,f,i [n] = βr,i [n] − (7) in zone 2 and constant power generation in zone 3. Another
2 typical control feature, which is not included in the benchmark
where i ∈ {1, 2, 3}, and βr,f,i [n] is the new pitch reference in controller, is a drive train damper that is designed to attenuated
which the sensor fault is contained. drive train oscillations. Inclusion of these features in the
benchmark model controller would have made it more compli-
C. Drive Train Model cated, and, on the other, lowered different transient behaviors
In this benchmark model, the drive train is modeled by a and oscillations, which would have made the fault detection
two-mass model. The purpose of the drive train is to transfer and isolation problem easier; consequently, the simple control
torque from the rotor to the generator. It includes a gear box design has been used in the benchmark model.
that increases the rotational speed from the low-speed rotor The controller is implemented in discrete time, with a
side to the high-speed generator side. sample frequency at 100 Hz. Subsequently, all time-dependent
A two-mass drive train model can be represented by variables in the controller are denoted as discrete-time vari-
Bdt ables.
Jr ω̇r (t) = τr (t) − K dt θ (t) − (Bdt + Br )ωr (t) + ωg (t) The controller starts in mode 1.
Ng
The control mode switches from mode 1 to 2 if
(8)
ηdt K dt ηdt Bdt Pg [n] ≥ Pr [n] ∨ ωg [n] ≥ ωnom
Jg ω˙g (t) = θ (t) + ωr (t)
Ng Ng where ωnom is the nominal generator speed. The control mode

ηdt Bdt switches from mode 2 to mode 1 if
− + Bg ωg (t) − τg (t) (9)
Ng2 ωg [n] < ωnom − ω
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1175

where ω is a small offset subtracted from the nominal TABLE II

generator speed to introduce some hysteresis in the switching W IND M ODEL PARAMETERS U SED IN THE B ENCHMARK M ODEL
scheme, thereby avoiding that the control modes are switching α H r0
all the time. 0.1 81 m 1.5 m
In both control modes, the controllers impose the generator
torque reference and the pitch reference τg,r [n] and βr [n], TABLE III
respectively. Notice that all three pitch systems receive the B LADE AND P ITCH M ODEL PARAMETERS U SED IN
same reference in this benchmark model. The control laws THE B ENCHMARK M ODEL
issuing these control signals in the two control modes are
described in the following. R ρ ζ ωn
The fault-tolerant control scheme will depend on the 57.5 m 1.225 kg3 0.6 11.11 rad
m s
nominal control scheme; much research has been conducted in
ζ2 ωn2 ζ3 ωn3
many-input many-output control design for wind turbines. In
this benchmark model, the commonly used industrial control 0.45 5.73 rad
s 0.9 3.42 rad
s
scheme is used.
TABLE IV
1) Control Mode 1: The optimal value of λ is denoted λopt
D RIVE T RAIN M ODEL PARAMETERS U SED IN
and is found as the optimum point in the power coefficient
THE B ENCHMARK M ODEL
(C p ) mapping of the wind turbine. The power coefficient
mapping characterizes the efficiency of the energy transfer
Bdt Br Bg Ng
from wind energy to mechanical energy and it depend on λ
and β. This optimal value is achieved by setting the pitch 775.49 Nms
rad 7.11 Nms
rad 45.6 Nms
rad 95
reference to zero (βr [n] = 0) and the reference torque to the K dt ηdt ηdt2 Jg
converter τg,r as follows: 2.7 · 109 Nm 0.97 0.92 390 kg · m2
rad
 
ωg [n] 2 Jr
τg,r [n] = K opt · (11)
Ng 55 · 106 kg · m2

where
1 Cp
K opt = ρ A R 3 3max (12) G. Model Parameters
2 λopt
The parameters used in the benchmark model are listed in
with ρ the air density, A the area swept by the turbine blades, the following tables. The wind model parameters can be found
and CPmax the maximum value of the power coefficient. in Table II, and the blade and pitch model parameters can
2) Control Mode 2: In this mode, the major control actions be seen in Table III. Table IV shows the drive train model
are handled by the pitch system using a PI controller trying parameters. The generator and converter model parameters can
to keep ωg [n] at ωnom be seen in Table V. The used controller parameters can be
βr [n] = βr [n − 1] + kp e[n] + (ki · Ts − kp )e[n − 1] found in Table VI. The parameters used in the sensor models
are shown in Table VII.
where e[n] = ωg [n] − ωnom . In this case, the converter
reference is used to suppress fast disturbances by
V. T EST S IGNALS D EFINITION
Pr [n]
τg,r [n] = . In the test signal definition described in [7], the defined
ηgc · ωg [n]
faults are present at a predefined time. This paper introduces
six additional test signal sets; they are formed by time-shifting
F. Sensors
the occurrence of the faults defined in the original benchmark
Each sensor is modeled by the sum of the actual variable model. This is done to check that the proposed FDI and FTC
value and a stochastic noise. The mean value and the variance schemes are robust toward different points of operation for the
of the noise are denoted as follows for the various measure- faults in question. In this benchmark model setup, a predefined
ments: wind speed sequence is used. This wind sequence consists of
m w , σw (wind speed - v w ); real measured wind data from a wind park. This wind speed
m ωr , σωr (rotor speed - ωr ); sequence can be seen in Fig. 6.
m ωg , σωg (generator speed - ωg ); In the listing of the possible faults, a subset is chosen for
m τg , στg (generator torque - τg ); the benchmark test sequence.
m Pg , σ Pg (generator power - Pg ); The test includes five sensors faults, three actuator faults,
m β , σβ (pitch angle - β). and one system fault. In the initial test set (test set 1), faults
Notice that for rotor speed, generator speed, pitch angles, are presented in the same order as in Table I. The time shift
and multiple sensors of the same kind are contained in the for the different test sets can be seen below.
wind turbine, and in this simulation model all sensors of the 1) Test set 2: +100 s for all faults.
same kind have the same stochastic parameters. 2) Test set 3: −100 s for all faults.
1176 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

TABLE V 30

G ENERATOR AND C ONVERTER M ODEL PARAMETERS

U SED IN THE B ENCHMARK M ODEL
25

αgc ηgc
50 rad
s 0.98 20

vw [m/s]
TABLE VI 15
C ONTROLLER PARAMETERS U SED IN THE B ENCHMARK M ODEL

K opt Ki Kp ωnom ω Pr 10

1.2171 1 4 162 rad

s 15 rad
s 4.8 · 106 W
5

TABLE VII
S ENSORS M ODEL PARAMETERS U SED IN THE B ENCHMARK M ODEL
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
mw σw m ωr σωr m ωg σωg Time [s]

1.5 ms 0.5 ms 0 rad

s 0.025 rads 0 rad
s 0.05 rad
s Fig. 6. Illustration of the wind speed sequence used in the benchmark model.
m τg στg m Pg σ Pg mβ σβ It can be seen that the wind speed covers the range 5–20 m/s, with a few
0 Nm 90 Nm 0W 1 · 103 W 0° 0.2° spikes at 25 m/s, which is a good coverage of normal operational for a wind
turbine.

period from 2900 to 3000 s.

3) Test set 4: −200 s for all faults. 7) Fault 7: fault type B4, change in the dynamics due to
4) Test set 5: −300 s for all faults. increased air content in the oil on pitch actuator 3. The
5) Test set 6: −400 s for all faults. fault is slowly introduced during 30 s with a constant
6) Test set 7: −500 s for all faults. rate; afterward the fault is active during 40 s, and again
It should be noticed that the schemes described in this paper decreases during 30 s. The fault begins at 3500 s and
are only designed based on test series 1. ends at 3600 s.
In the following, test set 1 is defined and the different 8) Fault 8: fault type B2, an offset on τg of the value
measurement signals are plotted as well. The fault parameters 100 Nm, the fault is active from 3800 to 3900 s.
are chosen such that they result in faults that are of such a size 9) Fault 9: fault type C1, a change in the friction in the
that they influence the wind turbine performance and they are drive train active from 4100 to 4300 s.
detectable and isolatable. On the other hand, they are chosen These faults must be detected and handled according to the
small enough so that they should be a challenge to detect and previously stated requirements. In order to validate the false
isolate. The actual values were found by trial and error on positive rate of the detection scheme, a set of data simulated
the benchmark model before it was originally published. One on an advanced model of the wind turbine is provided for a
fault parameter has been changed in the original model: the fault-free run on the same wind speed sequence.
offset on τg . The orginal value was 2000 Nm, which is now Plots of some of the relevant states and measurements
changed to 100 Nm. The change was made in order to make during this sequence of wind input and defined faults are
detection and isolation of this fault more challenging. It should presented in Figs. 7–14. The power, rotor speed, rotor speed
be noticed that the contributors of the schemes evaluated in measurements, generator speed measurement, and pitch angle
this paper were aware of this parameter change before they measurements can be seen.
finalized their designs. The benchmark model package contains a wind speed
The fault occurrence scenario is the following. sequence, which is a Simulink model with a parameter file.
1) Fault 1: fault type A1, a fixed value on β1,m1 equal to The package can be obtained at [35].
5° in the time period from 2000 to 2100 s.
2) Fault 2: fault type A2, a gain factor on β2,m2 equal to A. Additional Test Scenarios
1.2 in the time period from 2300 to 2400 s. It is relevant to test fault detection, isolation, and accom-
3) Fault 3: fault type A1, a fixed value on β3,m1 equal to modation with multiple occurring faults. The most chal-
10° in the time period from 2600 to 2700 s. lenging situation typically occurs when faults affect interacting
4) Fault 4: fault type A3, a fixed value on ωr,m1 equal to subparts of the wind turbine. An example of such a scenario
1.4 rad/s in the time period from 1500 to 1600 s. is suggested below.
5) Fault 5: fault types A4 and A6, gain factors on ωr,m2 1) Fault A: fault type A2, a gain factor on β2,m2 equal to
and ωg,m1 , respectively, equal to 1.1 and 0.9 in the time 1.2 in the time period from 2300 to 2600 s.
period from 1000 to 1100 s. 2) Fault B: fault type B3, change in the dynamics due to
6) Fault 6: fault type B3, change in the dynamics due to hydraulic pressure drop of the pitch actuator 2; the fault
hydraulic pressure drop of the pitch actuator 2; the fault is assumed to be abrupt and it is present in the time
is assumed to be abrupt and it is present in the time period from 2305 to 2600 s.
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1177

6 2.5
x 10
5

2
4

1.5
3

[rad/s]
1
P [w]

r,m1
2
g

ω
0.5
1

0
0

−0.5
0 500 1000 1500 2000 2500 3000 3500 4000 4500
−1 Time [s]
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]
Fig. 9. Measured rotor speed value from Sensor 1 as output of the simulation
Fig. 7. Wind turbine active power as output of the simulation with the defined with the defined faults present. It can be seen that this sensor signal is fixed
faults present. It can be seen that the different faults decrease the power in the time interval 1500–1600 s.
generation. This is easy to see in the region of full power from approximately
1900 s to the end; the drops in power below the full power level are due to
faults in the wind turbine. 2.5

2
2

1.8 1.5
[rad/s]

1.6
1
r,m1
ω

1.4
ω [rad/s]

0.5
1.2
r

1 0

0.8
−0.5
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]
0.6

Fig. 10. Measured rotor speed value from Sensor 2 as output of the simulation
0.4 with the defined faults present. This sensor has a fault given as a gain factor
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s] of 1.1 in the time interval 1000–1100 s.

Fig. 8. Rotor speed value as output of the simulation with the defined faults
present. The rotor speed is influenced by all the faults as with the generated VI. D ESCRIPTION OF FDI S OLUTIONS
power.
In this section, five solutions to the problems given in the
wind turbine benchmark model are shortly introduced.
3) Fault C: fault type A1, a fixed value on β1,m1 equal to
5° in the time period from 2600 to 3000 s.
A. Gaussian Kernel Support Vector Machine Solution (GKSV)
4) Fault D: fault type B4, change in the dynamics due to
increased air content in the oil on pitch actuator 3. The This scheme is based on a support vector machine built on a
fault is slowly introduced during 30 s with a constant Gaussian kernel, and is presented in [9]. In this design, a vector
rate; afterward the fault is active during 40 s, and again x of features is defined for each fault, which contains rele-
decreases during 30 s. The fault begins at 2605 s and vant measurements, filtered measurements, or combinations of
ends at 3000 s. these. Depending of the fault type, two–four features are used.
5) Fault E: fault type B2, an offset on τg of the value Residuals for all of the defined faults are obtained by
100 Nm, the fault is active from 3800 to 4400 s. projecting the feature vectors on the kernel of the support
6) Fault F: fault types A4 and A6, gain factors on ωr,m2 vector machine. Among a number of tested kernels, a Gaussian
and ωg,m1 , respectively, equal to 1.1 and 0.9 in the time kernel with different variance values was found to be the best
period from 3805 to 4400 s. one for all faults.
1178 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

180 25

160 20

140 15

β2,m2 [ ]
°
ωg,m1 [rad/s]

120 10

100 5

80
0

60
−5
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]

40
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]
Fig. 13. Measured pitch angle of Blade 2 from Sensor 2 as output of the
simulation with the defined faults present. In the time interval from 2300 to
2400 s, a gain factor of 1.2 is introduced.
Fig. 11. Measured generator speed value from Sensor 1 as output of the
simulation with the defined faults present. This sensor has a fault in the time
25
period from 1000 to 1100 s, in which the sensor has a gain factor of 0.9.

20
25

15
20

10

15
5
[°]
3,m2
β

10 0
[°]
1,m1

−5
β

−10

0
−15

−5 −20
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]

−10
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Time [s]
Fig. 14. Measured pitch angle of Blade 3 from Sensor 2 as output of the
simulation with fault scenarios. The output of this sensor is fixed in the interval
from 2600 to 2700 s.
Fig. 12. Measured pitch angle of Blade 1 from Sensor 1 as output of the
simulation with the defined faults present. The fault in this sensor results in
a fixed value in the time interval from 2000 to 2100 s.
C. Up-Down Counter Solution (UDC)
Data with and without faults were used for learning the Up-down counters are used in this solution for decision of
model for FDI of the specific faults, and, based on this the FDI based on residuals for each of the faults. The details of
vectors, kernel (structure and parameters) were found. the solutions can be found in [10]. The FDI residuals are based
on residuals obtained by physical redundancy, parity equations,
Kalman filters, and other filters.
B. Estimation-Based Solution (EB) There are two major differences between the used up-down
The general outline of this scheme is that a fault detection counters and straightforward thresholding methods. The first
estimator is designed to determine the presence of a fault, and is that the decision to declare a fault involves discrete-time
an additional bank of N isolation estimators is designed to dynamics and is not simply a function of the current value
isolate the faults, where N is the number of faults consid- of the residual. The second is that a penalty on the residual
ered. As described in [12], it is a preliminary and simplified exceeding the threshold is introduced in this scheme.
implementation of the general method given in [39] and [40].
Specifically, the method in [12] is designed on the basis of
a linear system model and without the use of an adaptive D. Combined Observer and Kalman Filter Solution (COK)
threshold. The estimators used for FDI are designed on the This solution uses a diagnostic-observer-based residual
basis of the provided models including model parameters. generator for residual generation for the faults in the drive
Each isolation estimator is designed on the basis of a particular train, in which the wind speed also is considered as a distur-
fault scenario under consideration. bance. Details on this scheme can be found in [8].
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1179

TABLE VIII
R ESULTS OF E VALUATION FAULTS 1–3

Fault GKSV EB UDC COK GFM

1 T d = 0.02 [s], T d = 0.02 [s], T d = 0.03 [s], T d = 10.32 [s], T d = 0.04 [s],
Td = 0.02 [s], Td = 0.01 [s], Td = 0.02 [s], Td = 10.23 [s], Td = 0.03 [s],
Td = 0.02 [s], Td = 0.02 [s], Td = 0.03 [s], Td = 10.33 [s], Td = 0.04 [s],
F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 0.89, F D = 0, Fd = 0,
Fd = 0 Fd = 0, Fd = 0 Fd = 0, Fd = 0
M D D = 3%, Fd = 1
M Dmin = 0%,
M Dmax = 20%
2 T d = 47.24 [s], T d = 44.65 [s], T d = 69.12 [s], T d = 19.24 [s], T d = 13.70 [s],
Td = 3.23 [s], Td = 0.63 [s], Td = 7.60 [s], Td = 3.43 [s], Td = 0.38 [s],
Td = 95.09 [s], Td = 95.82 [s], Td = 95.72 [s], Td = 49.93 [s], Td = 25.32 [s],
F D = 0, Fd = 0, F D = 22, Fd = 16, F D = 0, F D = 0.97, F D = 3.08,
Fd = 0, Fd = 28, Fd = 0, Fd = 0, Fd = 0, Fd = 1, Fd = 18
M D D = 56%, M D D = 56%, M D D = 67%, Fd = 5
M Dmin = 0%, M Dmin = 0%, M Dmin = 0%,
M Dmax = 100% M Dmax = 100% M Dmax = 100%
3 T d = 0.02 [s], T d = 0.54 [s], T d = 0.04 [s], T d = 10.35 [s], T d = 0.05 [s],
Td = 0.02 [s], Td = 0.51 [s], Td = 0.03 [s], Td = 1.54 [s], Td = 0.03 [s],
Td = 0.02 [s], Td = 0.76 [s], Td = 0.10 [s], Td = 10.61 [s], Td = 0.06 [s],
F D = 0, Fd = 0, F D = 4, Fd = 1, F D = 0, Fd = 0, F D = 1.42, F D = 1.61,
Fd = 0 Fd = 11, Fd = 0, Fd = 1, Fd = 4 Fd = 1, Fd = 5
M D D = 3%, M D D = 3%,
M Dmin = 0%, M Dmin = 0%,
M Dmax = 20% M Dmax = 20%

TABLE IX
R ESULTS OF E VALUATION FAULTS 4–6

Fault GKSV EB UDC COK GFM

4 T d = 0.11 [s], T d = 0.33 [s], T d = 0.02 [s], T d = 0.18 [s], T d = 0.10 [s],
Td = 0.09 [s], Td = 0.27 [s], Td = 0.02 [s], Td = 0.03 [s], Td = 0.03 [s],
Td = 0.18 [s], Td = 0.44 [s], Td = 0.02 [s], Td = 0.46 [s], Td = 0.34 [s],
F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 1, Fd = 1, F D = 2.31, F D = 3.36,
Fd = 0 Fd = 0 Fd = 8 Fd = 0, Fd = 5 Fd = 1, Fd = 18
5 T d = 25.90 [s], T d = 0.01 [s], T d = 2.96 [s], T d = 31.32 [s], T d = 9.49 [s],
Td = 1.24 [s], Td = 0.01 [s], Td = 0.38 [s], Td = 1.54 [s], Td = 0.56 [s],
Td = 87.49 [s], Td = 0.01 [s], Td = 21.08 [s], Td = 91.13 [s], Td = 17.18 [s],
F D = 0, Fd = 0, F D = 117, F D = 0.75, F D = 0.26, F d = 2.42,
Fd = 0, Fd = 95, Fd = 142 Fd = 0, Fd = 3 Fd = 0, Fd = 2, Fd = 1, Fd = 18
(MD)D = 3%, (MD)D = 14%,
M Dmin = 0%, M Dmin = 0%,
M Dmax = 20% M Dmax = 40%
6 (MD)D = 100%, T d = 11.31 [s], T d = 11.81 [s], T d = 23.80 [s], T d = 15.52 [s],
M Dmin = 100%, Td = 0.06 [s], Td = 0.53 [s], Td = 0.33 [s], Td = 0.02 [s],
M Dmax = 100% Td = 55.27 [s], Td = 55.72 [s], Td = 64.95 [s], Td = 61.13 [s],
F D = 2, Fd = 0, F D = 22, F D = 0.03, F D = 3.67,
Fd = 20 Fd = 15, Fd = 25 Fd = 0, Fd = 3 Fd = 1, Fd = 37

This diagnostic observer is designed to decouple the E. General Fault Model Solution (GFM)
disturbance and simultaneously achieve the optimal residual
generation with respect to process and sensor noise models. This solution is an automatic generated solution for FDI;
A Kalman-filter-based scheme is designed for two of the the details can be seen in [11]. Three main steps in the
subsystems. Generalized likelihood ratio test and cumulative design of this proposed method are: 1) a large set of poten-
variance index are used for fault decision based on the tial residual generators are generated; 2) the most suitable
statistical properties of the residual signals. For the fault residual generators are selected and then constructed by use
isolation purpose, a bank of residual generators based on dual of the algorithms presented in [41] (the selection is done by
sensor redundancy is designed. The residual bank is used to means of a greedy selection algorithm); and 3) the diagnostic
isolate the sensor faults, and system faults are isolated by a tests for the selected set of residual generators are designed.
decision table. A compensation strategy for the closed-loop A comparison between the estimated probability distributions
pitch system is used for the FDI system of the pitch sensor of residuals is used for diagnostic tests and evaluated with
faults. faulty and no-fault data.
1180 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

TABLE X
R ESULTS OF E VALUATION FAULTS 7 AND 8

Fault GKSV EB UDC COK GFM

7 (MD)D = 100%, T d = 26.07 [s], T d = 12.93 [s], T d = 34.00 [s], T d = 31.70 [s],
M Dmin = 100%, Td = 3.33 [s], Td = 2.86 [s], Td = 17.22 [s], Td = 0.61 [s],
M Dmax = 100% Td = 52.66 [s], Td = 51.08 [s], Td = 52.93 [s], Td = 180.70 [s],
F D = 1.8, Fd = 1, F D = 2, Fd = 1, F D = 0, Fd = 0, F D = 1.25,
Fd = 5 Fd = 4 Fd = 0 Fd = 1, Fd = 5
8 T d = 0.01 [s], T d = 0.01 [s], T d = 0.02 [s], T d = 0.01 [s], T d = 7.92 [s],
Td = 0.01 [s], Td = 0.01 [s], Td = 0.02 [s], Td = 0.01 [s], Td = 7.92 [s],
Td = 0.01 [s], Td = 0.01 [s], Td = 0.02 [s], Td = 0.01 [s], Td = 7.92 [s],
F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 0, Fd = 0, F D = 0, Fd = 0,
Fd = 0, Fd = 0, Fd = 0, Fd = 0, Fd = 0,
(MD)D = 97%, (MD)D = 97%, (MD)D = 97%, (MD)D = 97%, (MD)D = 97%,
M Dmin = 0%, M Dmin = 0%, M Dmin = 0%, M Dmin = 0%, M Dmin = 0%,
M Dmax = 100% M Dmax = 100% M Dmax = 100% M Dmax = 100% M Dmax = 100%

VII. E VALUATION OF S OLUTIONS the wind speed is low. An active FDI scheme would be better
for these faults, as proposed in [3].
In this section, the five solutions presented in Section VI In the following, some observations on the performance of
are compared on simulations with all seven test series and for the specific schemes are presented. Computational time has
all eight faults. The mean, minimum, and maximum values not been an evaluation criterion, since the proposed schemes
of fault detection time (Td ), the number of false positive have not been optimized in terms of computational time.
detections (FD ), and missed detections (MD) are computed Consequently, no comments are given on the computational
for the different methods. This means that these variables are times for the specific schemes.
computed for a given fault taking all simulations in all test GKSV only detects and isolates the sensor faults 1–5. Faults
series into account. 1, 3, and 4 are detected within the specifications in all test
The following notation is used: T d denotes the mean value series without any false positive detections. For those faults,
of the detection time; Td denotes the minimum value of the the scheme is independent of the time location of the faults and
detection time; Td denotes the maximum value of the detection therefore the point of operation at which the faults occurs. This
time; F D denotes the mean value of the number of false scheme thus provides good robustness toward the changed
positive detections; Fd denotes the minimum value of the point of operation of the faults. It should also be noticed
number of false positive detections; Fd denotes the maximum here that, even though that the scheme is data driven, system
value of the number of false positive detections; MDd denotes knowledge have been used to select the relevant measurements
the mean percentage of missed detection in a test series; for the different faults. It is also critical which data is used
MDmin denotes the minimum percentage of missed detections in the design process, e.g., whether the fault-free data cover
in a test series; and MDmax denotes the maximum percentage the entire operational range, and whether it is actually fault
of missed detections in a test series. In case no information free. In the benchmark case, the general wind speed sequence
on missed detections is provided, no missed detections were is the same in all simulations faulty or fault free. The scheme
present, meaning that all missed detection statistics were zero might have encountered difficulties if the test wind speed were
for the specific method for all test series. The results of the to contain values not included in the wind speed sequence
evaluation are shown in Table VIII for faults 1–3, in Table IX used in the design case. It is expected that this method can be
for faults 4–6, and in Table X for faults 7 and 8. None of the applied to the actuator and system faults as well.
proposed schemes was designed for detection and isolation of EB detects and isolates overall the faults fast for the original
fault 9, so these are not included in the tables. test series and slower as the fault time location, and thereby
In general, all schemes only detected fault 8 in test series the operational point moves away from the test series 1.
1, (which they are designed on). The reason for this is that A large number of false positive detections are present for
a change in the occurrence time of this fault will change the some faults. Clearly, the large number of false positive
ratio between the offset and the requested generator torque, detections could be lowered by a different choice of design
and since the offset is relatively small, this will make it very parameters. The estimator design requires knowledge of both
difficult to detect this fault in the other test series unless the model as well as the used estimator technique. It would
the algorithms were designed for these new generator torque be necessary to design the estimators for the different points
reference levels. This also leads to identical missed detection of operation of the faults in order to ensure a constant
statistics for all schemes for fault 8. Most of the schemes did performance of the scheme.
have problems detecting and isolating fault 2 for the higher UDC detects and isolates almost all faults in all test
test series numbers. The reason for this is that this fault is a series. Most of the faults were well detected and isolated
gain factor on one of the pitch sensors and the mentioned test relatively fast, but with some false positive detection. The
series include a low level of pitch values for these series since tests also showed that this scheme is relatively robust toward
ODGAARD et al.: FAULT-TOLERANT CONTROL OF WIND TURBINES 1181

the operational point at which the faults occur. Knowledge [4] K. Rothenhagen and F. W. Fuchs, “Current sensor fault detection and
of the model was used to find the best and simplest suited reconfiguration for a doubly fed induction generator,” in Proc. IEEE
Power Electron. Specialists Conf., Jun. 2007, pp. 2732–2738.
residual generators for the different faults. The advantage of [5] K. Rothenhagen, S. Thomsen, and F. W. Fuchs, “Voltage sensor fault
this is that the detection and isolation scheme is dedicated to detection and reconfiguration for a doubly fed induction generator,” in
specific faults, but, on the other hand, it also requires time Proc. IEEE Int. Symp. Diag. Electric Mach., Power Electron. Drives,
Sep. 2007, pp. 377–382.
and knowledge in the design process to find the best solution. [6] P. Poure, P. Weber, D. Theilliol, and S. Saadate, “Fault-tolerant power
It is expected that the UDC scheme could be returned to electronic converters: Reliability analysis of active power filter,” in Proc.
lower the number of false positive detections, and this decision IEEE Int. Symp. Ind. Electron., Jun. 2007, pp. 3174–3179.
[7] P. F. Odgaard, J. Stoustrup, and M. Kinnaert, “Fault tolerant control
rule provides a good balancing between fast detections and of wind turbines–A benchmark model,” in Proc. 7th IFAC Symp. Fault
avoidance of false positive detections. Detection, Supervis. Safety Tech. Int. Conf. Image Process., Jul. 2009,
COK detects and isolates most of the faults in all test series. pp. 155–160.
[8] W. Chen, S. X. Ding, A. H. A. Sari, A. Naik, A. Q. Khan, and S. Yin,
This is, however, done slowly and for most of the faults slower “Observer-based FDI schemes for wind turbine benchmark,” in Proc.
than required. The tests did also show a few false positive IFAC World Congr., Aug.–Sep. 2011, pp. 7073–7078.
detections for most of faults. Again, it is a scheme based on [9] N. Laouti, N. Sheibat-Othman, and S. Othman, “Support vector
a model that requires knowledge of the system, model, used machines for fault detection in wind turbines,” in Proc. IFAC World
Congr., Aug.–Sep. 2011, pp. 7067–7072.
observers, and Kalman filters. It would probably improve the [10] A. A. Ozdemir, P. Seiler, and G. J. Balas, “Wind turbine fault detection
performance of this scheme if it were returned to provide faster using counter-based residual thresholding,” in Proc. IFAC World Congr.,
detection even though it would increase the number of false Aug.–Sep. 2011, pp. 8289–8294.
[11] C. Svard and M. Nyberg, “Automated design of an FDI-system for the
positive detections slightly. wind turbine benchmark,” in Proc. IFAC World Congr., Aug.–Sep. 2011,
GFM detects and isolates all faults (except fault 8) in all pp. 8307–8315.
test series slowly and with some false positive detections. [12] X. Zhang, Q. Zhang, S. Zhao, R. M. G. Ferrari, M. M. Polycarpou, and
T. Parisini, “Fault detection and isolation of the wind turbine benchmark:
This scheme performs better than the other schemes, as the An estimation-based approach,” in Proc. IFAC World Congr., Sep. 2011,
operational point of the faults are moving further away from pp. 8295–8300.
the operational points of test series 1. This indicates a scheme [13] P. Pisu and B. Ayalew, “Robust fault diagnosis for a horizontal
axis wind turbine,” in Proc. IFAC World Congr., Aug.–Sep. 2011,
robust toward the point of operation of the faults. The major pp. 7055–7060.
advantage of this scheme is that it requires a very low level of [14] J. Blesa, V. Puig, J. Romera, and J. Saludes, “Fault diagnosis of wind
system knowledge and a simple system model. The detection turbines using a set-membership approach,” in Proc. IFAC World Congr.,
Aug.–Sep. 2011, pp. 8316–8321.
and isolation scheme is automatically generated based on [15] J. Dong and M. Verhaegen, “Data driven fault detection and isolation
all possible residuals. It provides a working solution, which, of a wind turbine benchmark,” in Proc. IFAC World Congr., Sep. 2011,
however, is not as good as the other tested solutions when pp. 7086–7091.
[16] F. Kiasi, J. Prakash, S. Shah, and J. M. Lee, “Fault detection and isolation
faults occurs close to the values in test series 1, in terms of of benchmark wind turbine using the likelihood ratio test,” in Proc. IFAC
detection time, etc., The fact that it is performing better as World Congr., Aug.–Sep. 2011, pp. 7079–7085.
the point of operation of the faults moves away from design [17] S. Simani, P. Castaldi, and M. Bonfe, “Hybrid model-based fault
point indicates that the generic nature of the design process detection of wind turbine sensors,” in Proc. IFAC World Congr., Sep.
2011, pp. 7061–7066.
introduces some robustness in the design, simply because it is [18] S. Simani, P. Castaldi, and A. Tilli, “Data-driven approach for wind
not optimally designed. turbine actuator and sensor fault detection and isolation,” in Proc. IFAC
World Congr., Aug.–Sep. 2011, pp. 8301–8306.
[19] F. Stoican, C.-F. Raduinea, and S. Olaru, “Adaptation of set theoretic
VIII. C ONCLUSION methods to the fault detection of wind turbine benchmark,” in Proc.
IFAC World Congr., Aug.–Sep. 2011, pp. 8322–8327.
In this paper, a benchmark model for testing fault detec- [20] S. Simani and P. Castaldi, “Data–driven design of a fuzzy logic fault
tion and fault accommodation schemes in wind turbines was tolerant control for a wind turbine benchmark,” in Proc. Safeprocess,
presented. The model simulates the actuator, sensor, and Aug. 2012, pp. 108–113.
[21] D. Rotondo, F. Nejari, V. Puig, and J. Blesa, “Fault tolerant control of
system faults in the pitch actuators, drive train, and converter the wind turbine benchmark using virtual sensors/actuators,” in Proc.
system. Various kinds of faults were included in this test IFAC Safeprocess, Aug. 2012, pp. 114–119.
bench model. Seven different test series were presented with [22] P. Casau, P. A. N. Rosa, S. Tabatabaeipour, and C. Silvestre, “Fault
detection and isolation and fault tolerant control of wind turbines
different time locations of the faults, which corresponded to using set-valued observers,” in Proc. IFAC Safeprocess, Aug. 2012,
different operating points at which the faults occurred. Five pp. 120–125.
different FDI schemes designed on one of these test series [23] S. Simani and P. Castaldi, “Adaptive fault-tolerant control design
approach for a wind turbine benchmark,” in Proc. IFAC Safeprocess,
were evaluated on all seven series. Aug. 2012, pp. 319–324.
[24] M. Sami and R. Patton, “Global wind turbine FTC via T-S fuzzy
R EFERENCES modelling and control,” in Proc. IFAC Safeprocess, Aug. 2012,
pp. 325–330.
[1] X. Wei, M. Verhaegen, and T. van den Engelen, “Sensor fault diagnosis [25] M. Sami and R. Patton, “An FTC approach to wind turbine power
of wind turbines for fault tolerant,” in Proc. 17th World Congr. Int. Fed. maximisation via T-S fuzzy modelling and control,” in Proc. IFAC
Autom. Control, Jul. 2008, pp. 3222–3227. Safeprocess, Aug. 2012, pp. 349–354.
[2] P. F. Odgaard, J. Stoustrup, R. Nielsen, and C. Damgaard, “Observer [26] X. Yang and J. Maciejowski, “Fault-tolerant model predictive control
based detection of sensor faults in wind turbines,” in Proc. Eur. Wind of a wind turbine benchmark,” in Proc. IFAC Safeprocess, Aug. 2012,
Energy Conf., Mar. 2009, pp. 1–6. pp. 337–342.
[3] C. Sloth, T. Esbensen, and J. Stoustrup, “Active and passive fault-tolerant [27] J. Kim, I. Yang, and D. Lee, “Control allocation based compensation
LPV control of wind turbines,” in Proc. Amer. Control Conf., Jun. 2010, for faulty blade actuator of wind turbine,” in Proc. IFAC Safeprocess,
pp. 4640–4646. Aug. 2012, pp. 355–360.
1182 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 4, JULY 2013

[28] F.D. Bianchi, H. De Battista, and R. J. Mantz, Wind Turbine Control Jakob Stoustrup (SM’99) received the M.Sc.
Systems. New York, NY, USA: Springer-Verlag, 2007. degree in electrical engineering and the Ph.D. degree
[29] T. Burton, D. Sharpe, N. Jenkins, and E. Bossanyi, Wind Energy in applied mathematics from the Department of
Handbook. 6 ed. New York, NY, USA: Wiley, 2008. Mathematics, Technical University of Denmark,
[30] I. Munteanu, A. I. Bratcu, N. A. Cutululis, and E. Caenga, Optim. Kongens Lyngby, Denmark.
Control Wind Energy Syst. Toward Global Approach, New York, NY, He has held several positions with the Technical
USA: Springer-Verlag, 2008. University of Denmark, as well as Visiting Profes-
[31] K. E. Johnson, M. J. Pao, L. Y. Balas, and L. J. Fingersh, “Control sorships with the University of Strathclyde, Glasgow,
of variable-speed wind turbines-standard and adaptive techniques for U.K., and the Mittag-Leffler Institute, Djursholm,
maximizing energy capture,” IEEE Control Syst. Mag., vol. 26, no. 3, Sweden.
pp. 71–81, Jun. 2006. Dr. Stoustrup acted as an Associate Editor, Guest
[32] Z. Hameed, Y. S. Hong, Y. M. Cho, S. H. Ahn, and C. K. Song, Editor, and Editorial Board member of several international journals. He is
“Condition monitoring and fault detection of wind turbines and related a former Chair of an IEEE chapter. Since 2008, he has been a Chairman
algorithms: A review,” Renew. Sustain. Energy Rev., vol. 13, no. 1, for the IFAC Technical Committee SAFEPROCESS. He has been a member
pp. 1–39, Jan. 2009. of IFAC Technical Board since 2011. He was a member of the Danish
[33] P. F. Odgaard, C. Damgaard, and R. Nielsen, “On-line estimation and Swedish Research Councils and the European Research Council. He
of wind turbine power coefficients using unknown input observers,” was a Board member of The Danish Academy of Technical Sciences. His
in Proc. 17th World Congr. Int. Fed. Autom. Control, Jul. 2008, main contributions to robust control, fault tolerant control, and plug-and-play
pp. 10646–10651. control have been made through 250-plus peer-reviewed papers. He has carried
[34] P.F. Odgaard, C. Damgaard, and R. Nielsen, “Adaptive power control out industrial cooperation with approximately 100 companies.
of wind turbines based on power coefficient estimations,” in Proc. Eur.
Wind Energy Conf., Apr. 2008, pp. 1–3.
[35] P. F. Odgaard, (2010). Wind Turbine Benchmark Model for Fault
Detection, Isolation and Accommodation [Online]. Available:
http://www.kk-electronic.com/wind-turbine-control/competition-on-
fault-detection/wind-turbine-benchmark-model.aspx
[36] D. S. L. Dolan and P. W. Lehn, “Simulation model of wind turbine 3p Michel Kinnaert (M’10) graduated in mechanical
torque oscillations due to wind shear and tower shadow,” IEEE Trans. and electrical engineering from Université libre de
Energy Convers., vol. 21, no. 3, pp. 717–724, Sep. 2006. Bruxelles (ULB), Brussels, Belgium, in 1983. He
[37] T. Esbensen and C. Sloth, “Fault diagnosis and fault-tolerant control received the M.S. degree in electrical engineering
of wind turbines,” M.S. thesis, Dept. Electroanal Eng., Aalborg Univ., from Stanford University, Stanford, CA, USA, in
Aalborg, Denmark, Jun. 2009. 1984 and the Ph.D. degree from ULB, in 1987.
[38] H. E. Merritt, Hydraulic Control Systems. New York, NY, USA: Wiley, He was with the Belgian National Fund for Scien-
1967. tific Research. He was appointed by ULB, where
[39] X. Zhang, M. M. Polycarpou, and T. Parisini, “Design and analysis of a he is currently a Professor with the Department
fault isolation scheme for a class of uncertain nonlinear systems,” IFAC of Control Engineering and System Analysis. He
Annu. Rev. Control, vol. 3, no. 2, pp. 107–121, 2008. held two Visiting Professor positions at LAGEP,
[40] X. Zhang, M. M. Polycarpou, and T. Parisini, “A robust detection Université Claude Bernard Lyon 1, Lyon, France. He has co-authored, with
and isolation scheme for abrupt and incipient faults in nonlinear M. Blanke, J. Lunze and M. Staroswiecki, the book Diagnosis and Fault
systems,” IEEE Trans. Autom. Control, vol. 47, no. 4, pp. 576–593, Tolerant Control (2nd Ed., Springer, 2006). His current research interests
Apr. 2002. include fault detection and isolation and fault tolerant control with applications
[41] C. Svard and M. Nyberg, “Residual generators for fault diagnosis using to the process industries, power systems and mechatronic systems.
computation sequences with mixed causality applied to automotive Dr. Kinnaert has been the Chairman of the IFAC Technical Committee, and
systems,” IEEE Trans. Syst., Man Cybern., A, Syst. Humans, vol. 40, he is an Associate Editor of Control Engineering Practice.
no. 6, pp. 1310–1328, Nov. 2010.

Peter Fogh Odgaard (M’05) received the M.Sc.

degree in electrical engineering and the Ph.D. degree
in control engineering from Aalborg University,
Aalborg, Denmark, in 2001 and 2004, respectively.
He was a Guest Researcher with the Department of
Mathematics, Washington University, St. Louis, MO,
USA, in 2003. From September 2004 to December
2006, he was a Research Assistant Professor with
Aalborg University. From January 2007 to April
2013, he was a Control Specialist with kk-electronic
a/s, Denmark. In April 2013, he joined Automation
& Control, Aalborg University, as a Research Associate Professor. He has
authored more than 60 peer-reviewed papers, and holds four patents and patent
applications. His current research interests include control of wind turbines,
wind farms and renewable energy generating systems, fault detection and
isolation, fault tolerant control, model predictive control and wavelets.
Dr. Odgaard has been a member of IFAC Technical Committee SAFE-
PROCES and Power and Energy Systems since 2011. He has served as a
Co-Chair of the Danish CSS/RAS joint chapter.