Vrije Universiteit Brussel
Active Fault Diagnosis on a Hydraulic Pitch System Based on Frequency-Domain
Identification
Vasquez Rodriguez, Sandra Paola; Kinnaert, Michel; Pintelon, Rik
Published in:
IEEE transactions on control systems technology : a publication of the IEEE Control Systems Society.
DOI:
10.1109/TCST.2017.2772890
Publication date:
2019
Document Version:
Submitted manuscript
Link to publication
Citation for published version (APA):
Vasquez Rodriguez, S. P., Kinnaert, M., & Pintelon, R. (2019). Active Fault Diagnosis on a Hydraulic Pitch
System Based on Frequency-Domain Identification. IEEE transactions on control systems technology : a
publication of the IEEE Control Systems Society., 27(2), 663-678. [8125575].
https://doi.org/10.1109/TCST.2017.2772890
Copyright
No part of this publication may be reproduced or transmitted in any form, without the prior written permission of the author(s) or other rights
holders to whom publication rights have been transferred, unless permitted by a license attached to the publication (a Creative Commons
license or other), or unless exceptions to copyright law apply.
Take down policy
If you believe that this document infringes your copyright or other rights, please contact openaccess@vub.be, with details of the nature of the
infringement. We will investigate the claim and if justified, we will take the appropriate steps.
Download date: 10. Aug. 2022
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY - MANUSCRIPT ID TCST-2017-0826 1
Active Fault Diagnosis on a Hydraulic Pitch System
based on Frequency Domain Identification
Sandra Vásquez, Michel Kinnaert, Member, IEEE, and Rik Pintelon, Fellow, IEEE
Abstract—The blade pitch system is a critical subsystem of
variable-speed variable-pitch wind turbines that is characterized
by a high failure rate. This work addresses the Fault Detection
and Isolation (FDI) of a blade pitch system with hydraulic actu-
ators. Focus is placed on incipient multiplicative faults, namely
hydraulic oil contamination with water and air, bearing damage
resulting in increased friction and drop of the supply pressure
of the hydraulic pump. An active model-based FDI approach is
considered, where changes in the operating conditions (i.e. mean
wind speed and turbulence intensity) are accounted for through
the identification of a Linear Parameter-Varying (LPV) model
for the pitch actuators. Frequency domain estimators are used
to identify continuous-time models in a user defined frequency
band, which facilitates the design of the FDI algorithm. Besides,
robustness with respect to noise in measurements and stochastic
nonlinear distortions is ensured by estimating confidence bounds
on the parameters used for FDI. The approach is thoroughly
validated on a wind turbine simulator based on the FAST
software that includes a detailed physical model of the hydraulic
pitch system. This paper presents the design methodology and
validation results for the proposed FDI approach. We show that
an appropriate design of the excitation signal used for active
fault detection allows an early fault diagnosis (except for oil
contamination with water) while ensuring a short experiment
duration and an acceptable impact on the wind turbine operation.
Index Terms—frequency domain identification, hydraulic pitch
system, linear parameter-varying, model-based fault diagnosis,
pitch actuator, wind turbine
I. INTRODUCTION
W
IND power is one of the fastest growing renewable
energy sources in the world. The global wind power
installation per year has increased from 6.5 GW in 2001
to 54.6 GW in 2016, leading to an installed wind capacity
of 486.8 GW at the end of 2016 [1]. Despite the progress
made as a result of this growth, reducing the cost of energy
is still a critical issue in order to make wind power more
competitive over conventional sources. In fact, one of the main
challenges of the wind industry is the reduction of operation
and maintenance (O&M) costs. Because most installation sites
are located in harsh environments and remote places, O&M
costs typically account for 20% to 25% of the overall levelized
cost of energy of wind power [2].
Sandra Vásquez is with the Department of Control Engineering and System
Analysis, Université Libre de Bruxelles (ULB), and with the Department
of Fundamental Electricity and Instrumentation, Vrije Universiteit Brussel
(VUB), Brussels, Belgium, e-mail: savasque@vub.ac.be.
Michel Kinnaert is with the Department of Control Engineering and System
Analysis, Université Libre de Bruxelles (ULB), Brussels, Belgium, e-mail:
Michel.Kinnaert@ulb.ac.be.
Rik Pintelon is with the Department of Fundamental Electricity and
Instrumentation, Vrije Universiteit Brussel (VUB), Brussels, Belgium, e-mail:
Rik.Pintelon@vub.ac.be.
In this context, Fault Detection and Isolation (FDI) for
wind turbines has gained increasing attention from industry
and academia. The capability for the early detection and
localization of faults makes FDI systems essential for the es-
tablishment of condition-based maintenance and repair, which
allows for significant cost savings. In the past, the focus of
FDI systems was mainly on the wind turbine drive train (i.e.
main bearing, shaft, gearbox and generator) [3]. However,
reliability field studies have exposed the blade pitch system as
the most critical subsystem for variable-speed variable-pitch
wind turbines. According to [4], this subsystem accounts for
16% of the overall failure rate, and for 20% of the overall
downtime. Besides, the function of the blade pitch system is
fundamental: blade pitching enables wind turbines to enhance
the energy capture, mitigate operational loads and perform
aerodynamic braking [5].
Blade pitch systems can use either electrical or hydraulic
actuators (i.e. based on electric motors or valve-controlled
hydraulic cylinders), with both technologies commonly used
among installed wind turbines [6]. This paper focuses on
hydraulic pitch systems, where oil related issues (leakage
and contamination) and component malfunctions (sensors,
pump, valve, etc.) still affect the proper operation of current
systems [7], [8]. These facts motivate the development of
FDI systems for hydraulic pitch systems. Besides, for cost
efficiency these FDI systems should ideally rely on commonly
available sensors.
Various FDI solutions for the hydraulic pitch system have
been published recently, notably in response to the wind
turbine benchmark model for fault diagnosis and fault tolerant
control presented in [9]. This benchmark model allows the
simulation of sensor and actuator faults in the generator and
the hydraulic pitch system, where FDI should be based on
standard measurements on wind turbines (i.e. wind speed,
rotor speed, blade pitch angles and generator speed, torque and
power). In particular, the hydraulic actuators are approximated
as second order Linear Time-Invariant (LTI) systems. Then,
faults like air contamination in the hydraulic oil and drop in
the supply pressure of the hydraulic pump are simulated as a
change on the parameters of this LTI system.
Different techniques have been considered to address the
FDI on this benchmark model. For example, a data-driven
fault detection (FD) approach based on principal component
analysis is presented in [10]. [11] provides a data-driven FDI
scheme based on Gibbs sampling and Fuzzy/Bayesian net-
works. A state space based set-membership approach for FD
is used in [12]. An FDI approach based on model falsification
using set-valued observers is proposed in [13]. Other solutions
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY - MANUSCRIPT ID TCST-2017-0826 2
for the benchmark model are reported in [14].
However, the proposed approximation of a hydraulic pitch
actuator as an LTI model has important limitations for rep-
resenting its dynamics under healthy and faulty conditions.
Indeed, a detailed physical model for hydraulic pitch actuators
that accounts for friction in the blade bearings and contamina-
tion of the hydraulic oil is presented in [5]. By integrating this
physical model in a wind turbine simulator based on the FAST
software of NREL (National Renewable Energy Laboratory)
[15], two issues that should be considered for FDI on hydraulic
pitch actuators were pointed out: i) the dynamics changes with
the operating conditions (i.e. mean wind speed and turbulence
intensity), and ii) active fault detection is necessary, since the
system input signal (i.e. the reference blade pitch angle) does
not excite well the system in the frequency band where the
faults are the most perceptible.
Therefore, we designed a model-based FDI system for
a hydraulic pitch system that considers these issues by: i)
introducing an additional excitation on the pitch system, and
ii) accounting for operating point changes through the identi-
fication of a Linear Parameter-Varying (LPV) model based on
input/output measurements (i.e. reference and measured blade
pitch angles).
The main idea of the FDI approach is to compare the LPV
model obtained for a healthy condition to a local LTI model
identified during the monitoring experiment. The method re-
lies on frequency domain estimators for the identification of
continuous-time models in a user defined frequency band. This
frequency selection can notably simplify the complexity of
the modeling problem. Besides, continuous-time models can
more easily be linked to the physics of the process, which
may ease the design of the FDI algorithms. Moreover, the
estimation of confidence bounds on the parameters used for
FDI is included in order to ensure robustness with respect to
noise in measurements and stochastic nonlinear distortions.
Three types of incipient multiplicative faults are studied:
contamination of the hydraulic oil (high content of air or
water), damages or poor lubrication on the pitch bearings
(causing increased friction), and drop in the supply pressure of
the hydraulic pump (due to wear). The proposed FDI system
is thoroughly validated on a wind turbine simulator based on
the FAST software that includes the detailed physical model
of the hydraulic pitch actuators developed in [5]. Through
this validation, we show how an appropriate design of the
excitation signal and the identification experiments ensures a
short experiment duration and an acceptable impact on the
wind turbine operation.
This paper presents the methodology for the design of
the FDI system and the validation results. It is organized as
follows. Section II provides insight into the operation of the
hydraulic pitch system, the faults under study and the wind
turbine simulator. Next, the general scheme for the proposed
model-based FDI system is discussed in Section III. Further,
Section IV explains the system identification methodology.
The analysis of the faults effects and the design of the
FDI algorithm is presented in Section V. Section VI reports
the results for the FDI system validation, together with the
assessment of the impact that active fault detection has on the
operation of the wind turbine. Finally, Section VII presents
the main conclusions.
II. SYSTEM DESCRIPTION
This work considers a horizontal-axis wind turbine (HAWT)
with a three-bladed rotor and a variable-speed variable-pitch
system. This section presents the working principle of the
hydraulic pitch system, the description of the faults under
study, and the wind turbine simulator used for testing the
proposed FDI system.
A. Hydraulic pitch system
The blade pitch system consists of three identical hydraulic
pitch actuators, each of them with an internal controller. As
shown in Fig. 1, the blade pitch angle (β
p
) is adjusted by
means of a valve-piston mechanism. The piston (c) is attached
to the blade base (a) through a connecting rod, while the blade
base is connected to the blade root via the pitch bearing (b). As
the piston cylinder is fixed to the blade root through a pivot,
β
p
can be set by adapting the piston position (x
p
). Then, x
p
is controlled by means of a hydraulic circuit consisting of a
distribution valve (d), a pump (e), and a tank (f).
Figure 1. Hydraulic pitch actuator. (a) blade base, (b) pitch bearing, (c)
cylinder/piston, (d) distribution valve, (e) pump, (f) tank. β
p
: blade pitch
angle. M
p
, M
b
: pitching and bearing friction moments. x
p
: piston position.
u
v
: control voltage of the valve. Q
A
, Q
B
: flows delivered to the piston
chambers. p
A
, p
B
: pressures in the piston chambers. p
S
, p
R
: supply pressure
and ambient pressure.
In order to control x
p
, appropriate flows (Q
A
and Q
B
) are
delivered to the cylinder chambers (A and B). The control
voltage of the valve (u
v
) governs these flows, which are
also dependent on the pressures in the piston chambers (p
A
and p
B
), the supply pressure (p
S
) and the ambient pressure
(p
R
). The dynamics of this system is also influenced by the
pitching (M
p
) and bearing friction (M
b
) moments, where M
b
is correlated with the pitch angle rate (
˙
β
p
). A proportional
controller is tuned in order to achieve the required dynamic
behavior of the pitch actuator from reference pitch angle
(β
p ref
) to measured pitch angle (β
p
) [5]. The signal β
p ref
,
the pump, and the tank are shared by the three pitch actuators.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY - MANUSCRIPT ID TCST-2017-0826 3
The physical model for the hydraulic pitch actuators (see
Appendix A) was used to simulate data representing the
healthy and faulty operation of the pitch system.
B. Fault description
This work considers the following incipient faults affecting
the dynamics of the hydraulic pitch actuators:
• Contamination of hydraulic oil: both high air content
(F
a
) and high water content (F
w
) are detrimental for
the hydraulic oil health, resulting in increased wear and
corrosion in machinery. These faults also impact the oil
compressibility, affecting the dynamics of the pressures
in the piston chambers (p
A
and p
B
). Under normal
conditions, the volume ratio of air in oil (R) is about
6-9% [16] and the ratio of water in oil (W ) is 0% [17].
F
a
occurs when R ≥ 18% (i.e. entrained air) [5], whereas
F
w
arises when W ≥ 0.5% (i.e. emulsified water) [17].
F
a
and F
w
are simulated by changing the effective bulk
modulus (B
eq
) in the piston chambers (R and W are
adjusted in (8) of Appendix A).
• Damages or poor lubrication on the pitch bearing: cor-
rosion, deformations and inadequate lubrication on the
pitch bearing result in a high bearing friction (F
f
), which
causes a poor performance of the pitch system. This fault
is simulated by increasing the bearing friction coefficient
(µ), which increments the bearing friction moment (M
b
)
(µ is adjusted in (9) of Appendix A). Under normal
conditions µ = 0.003 (for a bearing type ball bearing
with cage) [18], whereas F
f
occurs when µ ≥ 0.012 [5].
• Drop in the supply pressure: wear causes a hydraulic
pump to gradually lose its ability for developing full
pressure (F
p
), resulting in a reduced piston speed. This
fault is simulated by decreasing the supply pressure (p
S
)
to less than 75% of its rated value (p
N
S
) [19].
Table I summarizes the definition for the healthy and faulty
scenarios to be considered later. Since the pump and the
hydraulic oil are shared, F
a
, F
w
, and F
p
affect all three pitch
actuators equally.
Table I
DEFINITION OF HEALTHY/FAULTY SCENARIOS
Scenario R W µ
P
S
P
N
S
H : healthy 6 % 0 % 0.003 1
F
a
: high air content 18 % 0 % 0.003 1
F
w
: high water content 6 % 0.5 % 0.003 1
F
f
: high bearing friction 6 % 0 % 0.012 1
F
p
: drop in supply pressure 6 % 0 % 0.003 0.75
C. Wind turbine simulator
The proposed FDI system was tested on a simulator based
on the characteristics of a 1.25 MW real wind turbine. This
simulator is developed in the MATLAB/Simulink environment
(Fig. 2). Thanks to a DLL, the FAST software of NREL [15] is
integrated for the simulation of the coupled dynamic response
of the wind turbine. Besides, realistic wind simulations (i.e
Figure 2. Wind turbine simulator based on FAST. β
p P C
: pitch control signal.
β
p MS
: excitation signal for active fault detection. β
pi
, β
p ref
: blade pitch
angles and their reference. M
pi
, M
bi
: pitching and bearing friction moments.
ω
g
, ω
g ref
: generator rotation speed and its reference. P
g
: generator power.
stochastic, full-field, turbulent wind) are generated with the
software TurbSim of NREL [20].
The hydraulic pitch system is active when the wind speed
(v
w
) is between its nominal (v
w N
) and its cut-out (v
w co
) value
(i.e. the so-called power limitation zone). The pitch control
module moderates the aerodynamic input power so that, in
combination with the generator control, it ensures that the
generator power and speed (P
g
and ω
g
) are maintained at
their rated values when v
w
surpasses v
w N
. To this end, a
proportional–integral controller adjusts the pitch control signal
(β
p P C
) in a range from 0° to 25° based on the filtered
measurement of the generator speed (ω
g
) and its reference
(ω
g ref
). This control law is adapted based on the filtered
measurement of the blade pitch angles (β
pi
). Its design follows
the methodology described in [21].
The FDI system module requires the measured blade pitch
angles (β
pi
) and their reference (β
p ref
). Besides, a designed
excitation signal (β
p MS
) is added on top of the pitch control
signal (β
p P C
) in order to enhance the detection of faults. This
topic will be further developed below.
III. MODEL-BASED FDI SYSTEM
The proposed FDI system considers each hydraulic pitch
actuator as a black-box (Fig. 3): only the reference (β
p ref
)
and measured blade pitch angle β
p
are available. The pitching
(M
p
) and bearing friction (M
b
) moments are considered as
disturbances. The aim is to identify a linear model for each
pitch actuator and perform the fault diagnosis based on the
parameters of this model.
Figure 3. Hydraulic pitch actuator as a black-box. β
p
, β
p ref
: blade pitch
angle and its reference. M
p
, M
b
: pitching and bearing friction moments.
The general scheme of the proposed model-based FDI
system is presented in Fig. 4. Two operating phases are
distinguished: Configuration and Monitoring.
The configuration phase has to be completed before the
monitoring can be done. This includes the following tasks:
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY - MANUSCRIPT ID TCST-2017-0826 4
• Design of excitation signal: signal class, amplitude and
excited frequency band are chosen for β
p MS
. This
signal is applied only when the system identification
experiments are performed (both for the LPV model
identification and for the monitoring).
• LPV model for healthy condition: an LPV model
G
H
(jω
k
, q) is identified for each pitch actuator. This
model accounts for the dependence of the dynamics
on the operating conditions (i.e. mean wind speed and
turbulence intensity) expressed as scheduling parameters
q. The subscript H denotes a healthy scenario (as shown
in Table I).
The monitoring phase takes place regularly, with a frequency
established by the user. The following tasks are performed:
• Local LTI model for monitoring: a continuous-time LTI
model G
m
(s) is identified for each pitch actuator, and
the scheduling parameters q
m
are measured during the
monitoring experiment. The subscript m stands for mon-
itoring.
• Interpolated LTI model for healthy condition: for each
pitch actuator, a continuous-time LTI model G
H int
(s)
corresponding to healthy condition is derived from
G
H
(jω
k
, q) and q
m
. The subscript int denotes an in-
terpolation.
• Fault diagnosis: the models G
H int
(s) and G
m
(s) are
compared. A change in the model parameters denotes the
presence of a fault. Confidence regions for the parameters
are considered in order to provide robustness with respect
to noise in measurements and stochastic nonlinear dis-
tortions. When a fault is detected, the most likely fault
is chosen from a set of characterized faults. Further, a
warning or alarm message is generated depending on the
estimated fault severity.
Figure 4. General scheme of the model-based FDI system for the hydraulic
pitch system.
The design methodology for this FDI system is discussed
in more detail in Sections IV and V.
IV. FREQUENCY DOMAIN IDENTIFICATION
This section presents the proposed methodology for the
frequency domain identification of the hydraulic pitch actu-
ators. First, the main signals of a pitch actuator are analyzed.
Then, the strategy for the identification of local LTI models
and the design of the excitation signal are presented. Next,
the selection of scheduling parameters and the construction
of the LPV model are discussed. Finally, the identification
procedures are illustrated on one pitch actuator in healthy
condition.
A. Signals analysis
To establish a system identification strategy, the main sig-
nals of a hydraulic pitch actuator (Fig. 3) in healthy condition
were analyzed. The wind turbine was simulated for a mean
wind speed (¯v
w
) of 14.7 m/s and a turbulence intensity (T.I.)
of 10%. The excitation signal β
p MS
was not applied (i.e.
β
p ref
= β
p P C
, see Fig. 2). Figure 5 presents these main
signals in the frequency domain after applying a Hamming
window (length 200 s, sampling frequency of 800 Hz).
Figure 5. Hydraulic pitch actuator signals in the frequency domain.
Here are some remarks on these signals:
• Blade pitch angle (β
p
): β
p
fairly follows its reference
(β
p ref
) for frequencies below 0.6 Hz approximately.
• Pitch control signal (β
p P C
): this signal is the result of
the feedback action that regulates the generator speed
(ω
g
). The energy of β
p P C
is mainly concentrated in the
frequency band from 0 to about 0.4 Hz. However, to
detect faults it is necessary to excite the system from
0.1 to 30 Hz (see Subsection IV-E). Further, the non-
periodic nature of β
p P C
induces leakage on the frequency
response measurements.
• Pitching moment (M
p
): this disturbance has its energy
concentrated at 0 Hz and 0.35 Hz. The effect of M
p
is an increased variability of the frequency response
measurements at frequencies around 0.35 Hz.
• Bearing friction moment (M
b
): this signal is correlated
with the pitch angle rate (
˙
β
p
), and it brings an important
