11 - Performance Analysis of Electrical Signature Analysis-Based Diagnostics Using An Electromechanical Model of Wind Turbine

Renewable Energy xxx (2017) 1e27
Contents lists available at ScienceDirect
Renewable Energy
journal homepage: www.elsevier.com/locate/renene
Performance analysis of electrical signature analysis-based diagnostics

using an electromechanical model of wind turbine
Md Rifat Shahriar a, *, Pietro Borghesani a, Gerard Ledwich b, Andy C.C. Tan a, c
a
School of Chemistry, Physics and Mechanical Engineering, Queensland University of Technology, 2 George St, Brisbane, QLD 4000, Australia
b
School of Electrical Engineering and Computer Science, Queensland University of Technology, 2 George St, Brisbane, QLD 4000, Australia
c
LKC Faculty of Engineering and Science, Universiti Tunku Abdul Rahman, Sungai Long, Cheras, Kajang 43000, Selangor, Malaysia
a r t i c l e i n f o a b s t r a c t
Article history: Electrical signature analysis-based (ESA-based) diagnostics of powertrain faults in wind turbines (WTs) is
Received 2 December 2016 a promising alternative to the more traditional vibration-based condition monitoring. However, the
Received in revised form attempt to identify mechanical faults in electrical signals requires the consideration of the complex
31 March 2017
electromechanical dynamics of the WT. This paper investigates the potential masking effect of power
Accepted 4 April 2017
Available online xxx
electronic switching and wind-induced speed fluctuations on the electrical signatures of typical pow-
ertrain mechanical faults (i.e. rotor imbalance, gear cracks and other localised faults). To identify the
conditions in which these masking effects arise and their severity, an innovative full electromechanical
Keywords:
Wind turbine modelling
model of a WT has been developed, based on the integration of previously proposed models of WT sub-
Wind turbine diagnostics systems, and with the addition of powertrain fault models. This numerical controlled environment al-
Electrical signature analysis lows assessing the impact of power electronics and wind-speed fluctuation on the detectability of
Drivetrain faults powertrain faults by ESA. The results show the criticality of switching-induced noise over the whole
Fault signature masking range of simulated faults, whereas turbulence-induced noise is mainly affecting the detectability of low
frequency signatures. An order-of-magnitude sensitivity analysis is provided for the selected faults and
their interaction with the two masking effects, thus providing valuable indications for the development
of WT ESA-based condition monitoring systems.
© 2017 Elsevier Ltd. All rights reserved.
1. Introduction require an explicit model of the monitored system to perform fault

detection. In this methods, the detection takes the system input and
Due to the depletion of fossil fuel and its detrimental effect on output signals and utilizes the process model to generate features
the environment there is an urgent need to explore other alterna- [4]. These features are then compared with the nominal values to
tive energy sources. Among them wind energy has attracted a lot of produce analytical symptoms. The symptoms can be further uti-
attention worldwide for its abundance and cleanliness. To be able lized by the classification or inference methods for fault diagnosis.
to meet the demand as an alternative to tradition fossil fuel, there is One such model-based fault estimation method is proposed in
an increasing demand for reliable and economical operation of the Ref. [5] where a Takagi-Sugeno fuzzy model is used to represent a
wind turbines (WTs). Especially for the offshore wind farms the WT model and then an unknown input observer is utilized to es-
operation and maintenance (O&M) cost is particularly important as timate the faults. Some other recently developed model-based fault
they show greater failure rates compared to the onshore [1]. A well- diagnosis methods for WTs are reported in Ref. [6]. In contrast,
known approach to reduce the O&M cost is by employing systems signal-model-based methods do not require a process model, but
to monitor and detect component defects ahead of time before a are based on a mathematical model of the measured signal. These
total failure. methods are especially suitable when the measured output signals
A number of methods are available for fault detection and are characterized by harmonic and/or stochastic oscillations. Based
diagnosis which have been extensively reviewed in Refs. [2,3]. on the signal-model, features are calculated from the measured
Among the different methods, the process model-based methods signal [4]. Analytical symptoms are obtained by comparing the
features with those of the healthy system, which are further
exploited for fault diagnosis. The scope of this work is limited to the
* Corresponding author. signal-model-based approaches.
E-mail address: m.rifat@ieee.org (M.R. Shahriar).
http://dx.doi.org/10.1016/j.renene.2017.04.006
0960-1481/© 2017 Elsevier Ltd. All rights reserved.
Please cite this article in press as: M.R. Shahriar, et al., Performance analysis of electrical signature analysis-based diagnostics using an
electromechanical model of wind turbine, Renewable Energy (2017), http://dx.doi.org/10.1016/j.renene.2017.04.006
2 M.R. Shahriar et al. / Renewable Energy xxx (2017) 1e27
For the purpose of monitoring different WT components (e.g. and a torque/pitch controller. The limitation of this interesting
blades, bearing, gearbox, electric generator), the signal-model- modelling study is the simplified electric module: the generator
based approaches use measurements from the corresponding and converter are collectively represented as a first order system,
components which include strain (for blades), oil quality (for without the effects of power electronic switching. An enhanced
gearbox) and vibration/acoustic emission (for gearbox, main benchmark model, based on the FAST model, was also proposed in
bearing and generator). To obtain fault symptomatic features, this Ref. [19]. Due to the incorporation of FAST, the mechanical part of
signal-model-based methods employ time, frequency or joint time- this enhanced benchmark model is more realistic. However, the
frequency domain analysis techniques on the measured quantities electrical part of this enhanced model is inherited from the previ-
[7]. ous benchmark model (power electronic switching is not included).
As mentioned, the present WT monitoring systems require a The absence of a detailed electrical representation of the generator
number of additional sensors to monitor different WT components. and power electronics, coupled with models of the most common
A review of the commercial condition monitoring systems (CMSs) drivetrain faults, makes it impossible to investigate the realistic
can be found in Ref. [8] and a review of recent researches in this effectiveness of ESA-based detection of mechanical faults.
domain can be found in Ref. [9]. However, for WT performance To address this issue, this work presents integration of
monitoring, the traditional methods are now facing competition drivetrain mechanical faults in a full electromechanical model of a
from the newly developed methods that use the already available large scale WT. This will be used to investigate the feasibility and
measurements from the WT control system. In this regard, a shaft limitations of ESA-based diagnosis of mechanical faults. While
torque observer was proposed in Ref. [10], utilizing generator tor- rotor, drivetrain and controllers of the proposed electromechanical
que and speed measurements to estimate the main shaft torque, model are based on the benchmark model [18], the electrical
which can be used for rotor imbalance diagnosis. A gearbox components are modelled based on the FAST-PMSG WT [17]. The
monitoring system, based on the speed measurements, was re- first novelty of the proposed WT model resides in the integration of
ported in Ref. [11], which uses time-frequency-based method to the mechanical fault models in the full WT electromechanical
detect change in gearbox resonance frequency caused by gearbox model. This represents a preliminary step to the second (and main)
defect. In addition, electrical signature analysis-based (ESA-based) novelty of the study: the investigation of the fault detection per-
methods are proposed to detect rotor imbalance [12] and bearing formance of ESA under the influence of power electronic switching
fault [13] in a WT. Compared to the conventional techniques, the and wind-induced speed variation. This represents a significant
existing measurement-based diagnostic techniques (e.g. ESA) are step in the validation of ESA as an economical and yet reliable
more attractive due to the low cost and low effort of sensor alternative to traditional vibration-based condition monitoring
installation and signal acquisition. technology for WT in realistic operating conditions.
A number of investigations have been reported so far to The remainder of the paper is organized as follows: Section 2
demonstrate the potential of ESA-based diagnosis for the detection describes the modelling approach for the WT electromechanical
of wind turbine mechanical faults. However, due to the complexity system, along with a simplified WT model used as reference in the
of the system structure, most of the studies have been performed investigation. Section 3 discusses the development of fault models
on small-scale test-rigs without any feedback control systems and and their integration in the WT model. Section 4 presents the
power electronic converters [12e14]. In particular, no research has theoretical background of ESA-based fault detection. Section 5
yet reported on the robustness of ESA techniques against masking analyses the effect of different influencing factors (converter
effects from the multiple sources of disturbances within the com- switching and wind-induced speed variation) on the generator
plex electromechanical structure of a WT which limits the appli- output. Section 6 extends the analysis to a faulty drivetrain and
cation of ESA. The necessity to investigate multiple operation discusses simulation results of ESA-based fault diagnosis in pres-
regimes, combined with different drivetrain fault typology and ence of the aforementioned masking phenomena. Section 7 dis-
severity, makes an experimental investigation of this phenomenon cusses a possible improvement of the ESA-based technique. Section
on a full-scale real WT highly expensive and impractical. As a result, 8 discusses interaction of drivetrain faults with the WT control
there is a lack of actual condition data of full-scale WT available for system and finally, Section 9 draws the main conclusions of this
mutual comparison. Therefore, the motivation of this work is to study.
develop a controlled environment of numerical simulation models
which would provide the ideal framework for conducting the 2. Wind turbine modelling
performance analysis of ESA-based diagnostics.
To date the most sophisticated mechanical model of horizontal Two different variable speed wind turbine models, with
axis WT is the “NREL offshore 5-MW baseline wind turbine” different levels of modelling complexity, have been utilized in this
developed on the FAST aero-elastic simulator by the National study to perform necessary analysis. These two models, presented
Renewable Energy Laboratory in Colorado [15]. Although FAST- in Fig. 1, share the same modelling for the wind, the blade and pitch
based WT is an excellent tool for the detailed simulation of WT system, and the drivetrain. The two WT models differ in the level of
mechanics, its electrical portion is over-simplified. In order to detail in which they represent the generator and converter dy-
address this issue, previous studies have proposed to couple the namics. The WT model A contains the model of a permanent magnet
FAST-based model with a detailed electrical model of an Induction synchronous generator (PMSG) which sinks its output power in a
generator (IG) and a grid system [16]. A permanent magnet syn- passive load (therefore not grid connected). The same PMSG model
chronous generator-based (PMSG-based) WT model has also been is utilized in the WT model B, which however includes also the IGBT
proposed in combination with the FAST model that includes models based bridge rectifier and inverter connected by a DC link, a har-
of power electronic converters, PMSG and grid [17]. However, these monic filter, a transformer and the grid. As the component models
models were developed to analyse the power flow and load are already available in the literature, their brief description is
transmission along the drivetrain, and not for analysing the diag- provided hereafter with relevant references.
nostic capabilities of ESA. A benchmark model of WT aimed at
investigating fault tolerant control of WTs was proposed in Ref. [18]. 2.1. Modelling of wind and mechanical components
This later model includes a wind module, a simplified blade and
pitch system model, a two-degrees of freedom (DOF) drivetrain, The wind model consists of the following four additive
M.R. Shahriar et al. / Renewable Energy xxx (2017) 1e27 3
Fig. 1. Two modelling approaches for the variable speed WT system with different level of modelling complexity: (a) WT model A, (b) WT model B; (i. wind model, ii. blade and pitch
system, iii. drivetrain, iv. generator, v. electrical components).
components: i) an instantaneous mean wind speed (slow variation

of the wind speed) vm ðtÞ, ii) a stochastic part vs ðtÞ representing
X3
rpR3 Cq;i ðli ðtÞ; bi ðtÞÞvw;i ðtÞ2
wind turbulence modelled using the Kaimal filters [18], iii) a wind tr ðtÞ ¼ taero ðtÞ ¼ (2)
shear component vws ðtÞ representing the effect of vertical wind i¼1
6
speed gradient (ground boundary effect) and, iv) a tower shadow
component vts ðtÞ which reproduces the wind disturbance experi- where, r is air density and R is the radius of the blades. For a given
enced by the blades when passing near the tower. Descriptions of rotor speed ur ðtÞ, the tip speed ratio li ðtÞ is calculated as the ratio of
the wind shear and tower shadow model can be found in Ref. [18]. the tangential speed of the blade tip ður ðtÞ$RÞ and the wind speed
The wind shear and tower shadow are in phase events, resulting in vw;i ðtÞ.
a combined three-times-per-revolution (3xRev) disturbance in the The rotor of a wind turbine is connected to the generator
rotor torque/speed signal. Wind speeds vw;i ðtÞ experienced at the through a drivetrain (Fig. 2) which is characterized by a trans-
three blade tips are therefore obtained from this wind model, mission ratio Ng , torsional stiffness Kdt , torsional damping coeffi-
which can be represented by the following general expression. cient Bdt and efficiency hdt . The rotor and the generator shaft have
the moment of inertia of Jr and Jg , and viscous friction of Br and Bg
vw;i ðtÞ ¼ vm ðtÞ þ vs ðtÞ þ vws;i ðtÞ þ vts;i ðtÞ (1) respectively. This drivetrain model takes torques tr ðtÞ (rotor) and
tg ðtÞ (generator) as inputs and provides shaft angular speeds ur ðtÞ
where i ¼ 1 to 3 represents the blade number. (rotor) and ug ðtÞ (generator) as the outputs. It is represented by a
The modelling of WT blade and pitch system is performed ac- two-mass model as follows [18].
cording to Ref. [18]. The torque acting on each of the blades is
determined by the aerodynamic profile of the turbine, which de- " # " #
fines the torque coefficient Cq;i ð$Þ. This coefficient depends on the € € q_ r ðtÞ q ðtÞ tr ðtÞ
M$ qr ðtÞqg ðtÞ þ C$ _ þ K$ r ¼ (3)
blade tip speed ratio li ðtÞ and pitch angle bi ðtÞ. The total torque tr ðtÞ qg ðtÞ qg ðtÞ tg ðtÞ
experienced by the rotor can be expressed by the aerodynamic
torque as follows [18]. where,
Table 1
PMSG model specifics.
Parameter Value
Nominal voltage 690 V

Number of poles 4
Stator resistance 0.42 mU
Synchronous inductance 69.04 mH
Back EMF constant 630.774 Vpeak,L-L/krpm
The current output of the generator is fed to a three phase

passive load which in turn defines the voltages vd , and vq .
In WT model B, the PMSG is followed by a back-to-back converter
Fig. 2. Wind turbine drivetrain schematic diagram.
consisting of an IGBT-based active bridge rectifier, a DC link
capacitor and an inverter. A LCL filter keeps the harmonic distortion
of the inverter output at a minimum level to comply with the grid
2 3 code. Finally, the transformer steps up the voltage to the grid
B
6 Bdt þ Br dt 7
Ng voltage level and at the same time provides DC isolation from the
Jr 0 6 7
M¼ ;C ¼ 6
6 h B
! 7 and
7 K grid. A detailed description of these electrical models and associ-
0 Jg 4 hdt Bdt 5 ated parameter values (following the PMSG) can be found in
dt dt þ Bg
Ng Ng2 Ref. [17].
2 3
K
6 Kdt dt 7 2.3. WT controller
6 Ng 7
¼66 h K
7:
7 (4)
4 dt dt hdt Kdt 5 The task of the controller is to generate the necessary reference
Ng 2
Ng signals according to the control modes. The control modes are
dependent on the reference power curve of the WT, which is same
In this work, the torsional stiffness Kdt is modelled as a time for both of the WT models. The power curve has two major oper-
varying function (discussed later in Section 3.3). It provides the ational zones, and hence, the controller has two control modes. In
opportunity to simulate drivetrain faults such as gear defect. torque control mode (TCM, between cut-in (3 m/s) and rated
Numerical values of the model parameters, discussed above, can (12.5 m/s) wind speeds), the WT is operated in a way to maximize
be found in Ref. [18]. the power production through the maximum power point tracking
(MPPT) technique [20]. In pitch control mode (PCM, between rated
2.2. Modelling of generator and electrical components and cut-out (25 m/s) wind speeds), the controller aims at main-
taining the rotor speed at the nominal value by controlling the pitch
To convert mechanical power into electricity, a PMSG is included angle of the blades. In the following subsections, a brief description
in both WT model A and WT model B (Fig. 1). The non-salient PMSG of control operation, in these two control modes, is provided.
is characterized by a sinusoidal back electromotive force and
modelled as a second-order system [20]. The dynamic model of the 2.3.1. Torque control mode (TCM)
PMSG in the dq synchronous reference frame of the rotor, rotating In case of TCM, the controller operates to implement the MPPT
at the speed of ug , can be represented by the following equations technique by controlling the generator torque while keeping the
[20]. pitch reference ðbr Þ at 0 . For MPPT, the required torque reference is
determined by exploiting the quadratic dependency of the turbine
did 1 mechanical torque on the rotor speed [20]. The torque reference
¼ v Rid þ pug Ls iq (5)
dt Ls d ðtg;r Þ is calculated by the following equation [18]:

2
diq 1 ug ðtÞ
¼ vq Riq pug ðLs id þ lÞ (6) tg;r ðtÞ ¼ Kopt (9)
dt Ls Ng
3 where,
tg ¼ pliq (7)
2
1 CpðmaxÞ
Kopt ¼ rAR3 3 (10)
where vd , vq and id , iq represent d-axis and q-axis voltage and 2 l opt
current respectively. Ls is the synchronous inductance, which is the
same for both d and q axes. l is the peak flux of the rotor established A is the area swept by the turbine blades, CpðmaxÞ is the maximum
by the permanent magnets, R is the resistance of each phase of the value of the power coefficient, and lopt is the optimal tip speed
stator windings and p is the number of pole pairs. Specifics of the ratio. In WT model B, this torque reference is utilized by a PI
PMSG model are provided in Table 1. A sign adjustment is required controller for controlling the generator torque through the rectifier.
as the generator torque tg presented by Eq. (7) is negative whereas Substituting Eq. (7) into Eq. (6), and considering only the linear
the drivetrain model considers positive generator torque. The RMS part, a transfer function between the torque and the q-axis voltage
current output of each of the three stator phases can be given by of the generator can be obtained as follows [17]:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
I 2d þ I 2q te ðsÞ 3pl
Is ¼ (8) ¼ (11)
2 vq ðsÞ 2Ls s þ 2R
A PI controller is usually employed to produce the q-axis refer-

ence voltage for the system of Eq. (11), according to the torque
reference tg;r . However, before passing the reference voltage to the
Space Vector Pulse Width Modulation (SVPWM) scheme, the
nonlinear term ðpug ðl þ Ls id ÞÞ is added to this q-axis reference
voltage [17]. Similarly, a PI controller is employed to implement
zero d-axis current control (ZDC) for the generator [20]. The recti-
fier is operated, based on the SVPWM scheme, to control the q-axis
and the d-axis current of the generator. Due to the ZDC, theoreti-
cally the generator current magnitude depends on its q-axis
component only. This q-axis component also defines the torque of
the generator according to Eq. (7).
To ensure continuous power flow from the generator to the grid,
the DC link voltage needs to be maintained at a constant level. This
is performed by controlling the q-axis grid current. The reactive
power output of the WT is controlled by the d-axis grid current. The
control of these grid quantities is performed by the inverter
controller through appropriate reference voltages. A detailed
description of the grid side control operation and the design
technique of the PI controllers can be found in Ref. [17].
2.3.2. Pitch control mode (PCM)

In case of PCM, the controller tries to maintain the generator
speed at the rated value. In this respect, a PI controller is employed Fig. 3. Implementation of aerodynamic imbalance.
to produce the reference pitch angle ðbr Þ according to the generator
speed ðug Þ. The pitch system of each blade tries to follow its br . The
pitch angle measurement bi , obtained from the pitch system, is results in a different pitch reference for the faulty blade's pitch
related to its reference br by the following close loop transfer actuator system. During the simulations, while the healthy pitch
function [18]: system receives a pitch reference of br , the faulty one receives the
pitch reference br3f ¼ br þ ðb3 bm3f Þ=2, where bm3f represents
bi ðsÞ u2n the fixed position measurement of the faulty sensor.
¼ (12)
br ðsÞ s2 þ 2zun s þ u2n The cause of aerodynamic imbalance is that the pitch angle of
the faulty blade is different compared to the other two healthy
where un is the natural frequency and z is the damping factor. A blades, making the aerodynamic response of the former different
more detailed description of the pitch system and associated from the latter. This physical phenomenon can be expressed
controller parameters can be found in Ref. [18]. mathematically by Eq. (2) with a different value of Cq;i for the faulty
To suppress fast disturbances, in PCM, the torque reference tg;r blade (due to different bi ), while the wind profile remains the same
is calculated based on the rated power Pr and efficiency hgc [18]. for all the three blades. This clearly indicates that, at the event of
tower shadow and wind shear, torque/speed disturbance contrib-
Pr uted by the faulty blade will have a different magnitude compared
tg;r ðtÞ ¼ (13)
hgc ug ðtÞ to the other two healthy ones. In the torque/speed signal, this re-
sults in an amplitude modulation of the 3xRev phenomenon
(caused by tower shadow and wind shear) by a periodic component
with frequency of 1xRev (rotational speed of faulty blade). Thus,
3. Integration of fault models aerodynamic imbalance, caused by a faulty pitch system, causes
signatures at frequencies n 3xRev±m 1xRev where n; m2Z in
Four different drivetrain faults are incorporated in the WT the speed signal spectrum.
models to investigate their detectability through ESA-based tech-
nique. Implementation of these faults into the WT drivetrain is 3.2. Mass imbalance
discussed in the following sub-sections.
Mass imbalance is caused by a change of mass distribution or
3.1. Aerodynamic imbalance total mass of the turbine blades. In the simulation model imbalance
is added to a nominal perfectly balance situation in which the three
To represent the typical case of aerodynamic imbalance, a blade blades have the same mass m and a centre of gravity located at a
pitch system fault is investigated in this study [18]. Blade pitch distance b from the shaft. Without any loss of generality, we choose
faults can be caused by a faulty sensor measurement of one of the to simulate imbalance as an additional mass mx placed in the centre
blade's pitch angle as illustrated in Fig. 3. Here, blade 1 and 2 pitch of gravity of the reference blade B1 which defines the rotation angle
angle measurement sensors (not shown) are considered healthy, so qr ðtÞ of the rotor shaft. Moreover we choose to express the imbal-
their pitch actuators receive healthy measurements and are able to ance mass as a fraction x of the nominal blade mass m ðmx ¼ x mÞ,
follow br properly. However, one of the two pitch angle measure- as shown in Fig. 4.
ment sensors of blade 3 is considered faulty. The faulty sensor Among the two resulting forces, the centrifugal force of
provides a fixed position measurement. It results in an error in the magnitude mx bu2r can be neglected due to the slow speed of a large
measured pitch angle sent to the pitch actuator, which is eventually scale WT rotor. However, the gravitational force, with amplitude
used for setting the pitch reference of blade 3. Therefore, this fault mx g can cause a periodic variation of the rotor speed with an
smoother power transmission. It is assumed that the proposed

electromechanical model of WT has only one planetary stage and
one parallel stage (high speed stage) with an overall gearbox ratio
of 95:1. It is also assumed that the gear ratio of planetary stage and
parallel stage is 16.814:1 and 5.65:1 respectively. As the gear faults
are more often found on the high speed helical stage, modelling of
time-varying meshing stiffness is performed for this stage only
[22]. This high speed stage is consist of an intermediate speed shaft
(ISS) and a high speed shaft (HSS). The time-varying meshing
stiffness for the gear pair (between ISS and HSS) are therefore
modelled.
Based on the distinctive meshing characteristics of the helical
gears, a time-varying meshing stiffness function was proposed in
Ref. [23] which is calculated based on the gear pair's design data. In
this stiffness function, the maximum stiffness is calculated at the
pitch point and the stiffness function ki ðtÞ of a helical tooth pair is
described as follows:
3 !
t εt2z

ki ðtÞ ¼ kp exp Ca (16)
1:125 εa tz
Fig. 4. Illustration of mass imbalance in a WT rotor.
where, kp is the stiffness value at the pitch point, Ca is an empirical

acceleration/deceleration during the downward/upward motion of coefficient, ε is the total contact ratio, εa is the transverse contact
the unbalanced blade. This periodic speed variation can be ratio of the helical tooth pair and tz is the meshing period for
modelled by introducing an additional torque tim ðqr ðtÞÞ on the passing one transverse base pitch. The parameters kp and Ca
rotor [21]. depend on the gear geometry and their calculation methods are
detailed in Ref. [23].
tim ðqr ðtÞÞ ¼ xmgb cosðqr ðtÞ Þ (14) Adding the contribution of the meshing stiffness ki ðtÞ of each
tooth pair, it is possible to obtain the total time varying torsional
The total torque tr ðtÞ of Eq. (2) is therefore redefined as follows.
stiffness kt ðtÞ:
tr ðtÞ ¼ taero ðtÞ þ tim ðqr ðtÞÞ (15)
X
I
The implementation of the mass imbalance in the WT blade kt ðtÞ ¼ r 2 ki ðtÞ (17)
system is illustrated in Fig. 5. To determine imbalance-induced i¼1
torque, the mass of a blade is considered as 16590 kg (inspired by
Ref. [15]). where I is the total number of tooth pairs meshing in one full
As shown in Eqs. (14) and (15), the mass imbalance causes a revolution of the pinion shaft and r is the equivalent radius of the
sinusoidal torque fluctuation and a consequent analogous variation meshing gear and pinion.
in the rotor speed signal. To simulate a realistic wind turbine gearbox scenario, high
speed stage design data are obtained combining the design pa-
rameters of Refs. [23] and [24] which are utilized to calculate ki ðtÞ
3.3. Gear tooth crack and kt ðtÞ. In Fig. 6 (a) meshing stiffness variation due to mating of
each tooth pair is shown separately for one complete HSS revolu-
In this section, the development of a time-varying meshing tion with a 20 teeth pinion and corresponding kt ðtÞ is provided in
stiffness function (as pointed out in Section 2.1) is described first, Fig. 6 (c) for healthy state. It should be mentioned here that, the
followed by its application to gear tooth fault simulation and shape of kt ðt Þ is in agreement with the results of finite-element-
implementation to the WT drivetrain model. based method and that of multi-body simulation software pre-
A typical large scale WT drivetrain has one or more low speed sented in Refs. [25] and [26] respectively.
planetary stages along with one or more high speed parallel stages. A previous investigation has found that tooth cracks cause a
The planetary stage may have spur or helical gears whereas parallel reduction in meshing stiffness when the defective tooth is in con-
stages usually contain helical gears to get higher contact ratio for tact [27]. Therefore, the tooth fault is modelled in this paper as a
percentage reduction of the stiffness at pitch point. Both ki ðtÞ and
kt ðtÞ are provided in Fig. 6 (b) and (d) respectively for a tooth fault
severity of 5%.
The magnitude of kt ðtÞ is adjusted so that total average stiffness
of the drivetrain corresponds to the value provided in Ref. [18]. The
constant drivetrain torsional stiffness Kdt (in Eq. (3)) is then
replaced by the adjusted time-varying function kt ðtÞ.
The implementation of the gear fault in the WT drivetrain model
is illustrated in Fig. 7. The magnitude of the drivetrain stiffness is
determined based on the shaft angular position and it is utilized to
update the stiffness matrix Kðqg Þ for every shaft position during the
simulation.
Fig. 5. Implementation of mass imbalance in WT model. As the gear defect is simulated as a crack on a single tooth of the
Fig. 6. Time varying meshing stiffness for one shaft revolution: (a) separated meshing stiffness for each tooth pair (healthy), (b) separated meshing stiffness for each tooth pair
(faulty), (c) synthesized meshing stiffness (healthy), (d) synthesized meshing stiffness (faulty).
2
nd
q sin q
td ðqÞ ¼ sinðnd qÞe 2
(18)
where, nd is the defect frequency in events per revolution, and q is a

large positive number resulting in a short-time torque disturbance
ðq > 100Þ. A normalized disturbance function is presented in Fig. 8
(a) for one shaft rotation. The implementation of a localised fault
in the drivetrain of WT model is also presented in Fig. 8 (b). As the
generator shaft angle, wrapped between 0 and 2p, is taken as
reference, Ns (gear ratio between generator and defective shaft) is
Fig. 7. Implementation of HSS pinion tooth fault in the WT drivetrain. multiplied to the generator angle to determine the shaft position.
This shaft angle is utilized to determine the magnitude of the
disturbance function. The defect factor ðdf Þ defines the percentage
high speed shaft (generator shaft), it is expected to produce a one- fluctuation of generator torque ðtg Þ as a result of the localised
per-revolution speed variation in the generator speed signal. defect.
Moreover, it is expected to result in corresponding modulation of The torque fluctuation caused by the drivetrain localised defect
the gear meshing frequencies (GMFs). is expected to result in a similar disturbance in the corresponding
shaft speed signal. In the HSS (generator shaft) speed signal it ap-
pears as a periodic disturbance with order Ns nd and its harmonics.
3.4. Localised fault in the kinematic chain The local defect is simulated on the ISS gear (of the high speed
stage) for which Ns ¼ 20=113 (here HSS gear has 20 teeth and ISS
This sub-section is aimed at modelling a general localised defect gear has 113 teeth). As the defect frequency nd ¼ 4 per ISS rev, the
in the rotating components of a drivetrain. A localised defect (e.g. signature of local defect appears at 0.7080xRev in generator speed
bearing outer race defect) results in a periodic disturbance in the spectrum. However, when the rotor shaft is taken as reference, the
shaft's torque. Inspired by Ref. [28], a periodic angular disturbance defect signature, in generator speed signal appears at gearbox ratio
representing the localised fault is expressed as follows: (95) times Ns nd xRev.
Fig. 8. Localised fault in drivetrain of WT model: (a) disturbance function for one shaft rotation when nd ¼ 4 and q ¼ 100, (b) implementation in the WT model.
4. Principle for ESA-based fault diagnosis

X
∞
ug ðqÞ ¼ ug þ Duh cosðhgq þ jh Þ (22)
4.1. Signatures of mechanical phenomena in generator output
h¼1
Drivetrain mechanical phenomena (e.g. gear meshing, tower where the amplitude of the speed fluctuation due to each har-
shadow, drivetrain faults) result in fluctuations in the generator monics of uh is Duh . This speed fluctuation will result in the
speed signal. These signatures are expected to carry out in the following electrical output:
PMSG electrical output as the output eg ðqÞ is related to the gener-
( )
ator shaft speed ðug Þ and angle ðqÞ by the following expression. X
∞
eg ðqÞ ¼ fp ug þ Duh cosðhgq þ jh Þ cosðpqÞ (23)
h¼1
eg ðqÞ ¼ fpug cosðpqÞ (19) In time domain, this signal has a mixed AM/FM nature, whose
full analytical expression is non-trivial (similar to a non-linear
where f refers to the flux linkage, p is the number of pole-pairs and pendulum). For small fluctuations, this is however well-
the stator electrical frequency ðus Þ equals to the synchronous speed approximated as:
for a PMSG (i.e. us ¼ pug ). In case of a drivetrain fault, the speed
( )
term of Eq. (19) would be dependent on the angular rotation of the X
∞
shaft itself: eg ðqðtÞÞ ¼ fp ug þ Duh cosðhgqðtÞ þ jh Þ cos us t
h¼1
!
X
∞
p Duh
þ sin hgug t þ jh (24)
ug ðqÞ ¼ ug þ uðgqÞ (20) hgug
h¼1
where ug represents the average (nominal) generator shaft speed

where us ¼ pug is the nominal electrical frequency.
and uðgqÞ is a characteristic speed-disturbance function, which is
The spectral content of eg ðqÞ should therefore show a signature
2p=g-periodic in nature:
characteristic of the drivetrain phenomenon, with g-sidebands
around the fundamental order p:

2pk X
∞
uðgqÞ ¼ u g q þ ; k ¼ 1; 2; … (21) F eg ðqÞ ¼ Ah 1ðU±ðp þ hgÞÞ
g (25)
h¼∞
The shaft harmonic order g (often integer) is characteristic of
the specific drivetrain fault/phenomenon which is phase-locked to where F is the complex Fourier series operator (in this case
the shaft speed. transforming from the angular domain q of the shaft to its order
Given the periodic nature of the function uðgqÞ, it is possible to domain U), 1ð,Þ is the indicator function and Ah is the amplitude of
expand it in a Fourier series of harmonics the sideband (or main peak) of the generator current spectrum
uh ðgqÞ ¼ Duh cosðhgq þ jh Þ. (itself depending on the amplitude and periodicity of the speed
Therefore, the instantaneous angular speed in this case would disturbance).
contain a constant part along with the fault related speed variation: Table 2 shows the expected characteristics in the spectral
Table 2
Characteristics of the most typical drivetrain phenomena-induced speed fluctuations and their expected effect on the generator current spectrum (gear ratio ug ⁄ ur ¼ 95,
n ¼ 1; 2; …).
Phenomenon Speed disturbance function ud ðqÞ shape Order Order Sideband location in
(ref. generator shaft) g (ref. rotor shaft) eg ðqÞ order spectrum
gug =ur ðp þ gÞug =ur
Tower shadow Multi-harmonic 3 3 190±3

95 ¼ 0:0316
Gear meshing Multi-harmonic (almost sinusoidal) 20 1900 2090
Mass imbalance Mono-harmonic (sinewave) 1 1 190±1
95 ¼ 0:0105
Aerodynamic imbalance Multi-harmonic 3±n 3±n 190±n
95
Gear tooth crack Multi-harmonic 1 95 190±95 and
2090±95
Localised defect Multi-harmonic 0:708 67:26 190±67:26
representation of eg ðqÞ discussed in the previous chapter.
4.2. Signal processing for ESA-based diagnosis
The common approach to extract the speed information from

the current signal is through amplitude demodulation by utilizing
Hilbert-based analytic signal. The analytic signal itself is based on
the Hilbert transform of eg ðqÞ (i.e. ~
eg ðqÞ). Assuming full separation of
the positive and negative frequency sideband patterns (around p
and p, Bedrosian condition [29]), the Hilbert transform of the
signal eg ðqÞ results in:
( )
X
∞
eg ðqÞ ¼ fp ug þ
~ Duh cosðhgq þ jh Þ sinðpqÞ (26)
k¼1
and the analytic signal ea ðqÞ ¼ eg ðqÞ þ j~

eg ðqÞ can be obtained as:
( )
X
∞
ea ðqÞ ¼ fp ug þ Duh cosðhgq þ jh Þ expðjpqÞ (27)
k¼1
The absolute value of the analytic signal ea ðqÞ (i.e. envelope of

eg ðqÞ) is therefore theoretically proportional to the generator
speed:
( )
X
∞
jea ðqÞj ¼ fpug ðqÞ ¼ fp ug þ Duh cosðhgq þ jh Þ (28)
k¼1
This analytical derivation is exact in the angular/order domain of

the shaft, given the phase-locking of the drivetrain phenomena
with the reference shaft. In order to obtain this angular-domain
description of the signal, computed order tracking (COT) is
applied as a pre-processing tool to the electrical output eg ðtÞ. Using
a reference signal, the COT converts signals sampled at a constant
time interval into signals sampled at a constant angular interval.
One way to perform this angular resampling is to first identify the Fig. 9. Envelope-based ESA technique for drivetrain fault diagnosis.
section of signal that is sampled in one shaft rotation using 1xRev
tachometer signal. Then samples can be obtained at equivalent
angular interval by suitable interpolation method. Implementation rectifier that connects PMSG with the DC link employs VSC prin-
details of this COT technique can be found in Ref. [30]. The COT step ciple to set the generator voltage output according to approxi-
applied to eg ðtÞ not only results in a purely amplitude modulated mately constant DC link voltage. In this setting, the generator
current output has lower THD compared to the phase/line voltage,
eg ðqÞ, but also mitigates the smearing effects of low-frequency
due to the filtering effect of generator inductance. This makes the
wind-induced speed fluctuations on the shaft speed, which smear
phase current of the generator a more suitable candidate for ESA in
the fault-symptomatic sidebands in the raw frequency domain.
this particular case.
In ESA-based fault diagnosis, to achieve better demodulation
performance, the order tracked electrical signal is also band-pass
filtered at around the fundamental electrical order p to keep only 5. Effect of converter switching and wind turbulence on
those sidebands that are related to fault signature. A suitable generator current
demodulation technique, as explained above, can be applied to the
filtered output to extract the fault signature. Applying Discrete In this section, simulation results of WT model B and WT model A
Fourier transform (DFT) to the demodulated signal, the order (whose only difference is the absence of power electronics) are
spectrum is obtained, where the fault signature can then be clearly compared to assess the influence of converter switching. Moreover,
detected. The general technique of ESA-based fault diagnosis is WT model B is used to simulate different wind turbulence scenarios
provided in Fig. 9. to analyses the performance of order tracking in mitigating wind-
In case of gear fault diagnostics, a second demodulation is per- induced rotor speed fluctuations.
formed on the obtained current envelope (itself proportional to In all the analyses presented in this section and the next section
speed), in order to identify tooth-specific anomalies in the gear- (Section 5 and Section 6), the first part of the simulation is dis-
meshing phenomenon. In this case, the signal envelope is ob- carded, since it includes the ramp-up transient of the system from
tained first by AM demodulation and the resulting envelope signal null speed to normal operating conditions. The time duration of
is filtered around the GMF, to keep only the defect signature side- each signal is 350 s. During the simulations, the data acquisition is
bands. It is then followed by a further demodulation step (in this performed at a sampling frequency of 20 kHz and an antialiasing
case either AM or FM), and DFT to reveal the gear defect signature. filter with cut-off frequency of 6 kHz is applied (just above the first
Both electrical outputs of the WT generator (current and 4 kHz harmonics of the converter switching). However, in case of
voltage) can be utilized for drivetrain fault diagnosis through ESA. field implementation of the ESA-based techniques, a down-
However, in the proposed WT model (WT model B), the active sampling step might be necessary to reduce memory
consumption. This could in practice be quite significant, given the ratio (SNR) of the fundamental harmonics in the current signal:
long-time windows necessary to observe a sufficient number of for example, in this present scenario, the calculated SNRs of the
periods of the relatively low frequency fault symptoms, compared current signals are approximately 33 dB and 45 dB for WT model B
to the multi-kHz range of the typical control system sampling rates. and WT model A respectively. Therefore, the widespread spectral
In addition, the down-sampling step would provide the opportu- distribution of the converter noise needs to be considered when
nity for substantial reduction of noise, such as arising from electro- choosing the bandwidth around the fundamental frequency for its
magnetic interference. For order tracking purposes, a 1xRev demodulation: the larger the bandwidth, the more electrical noise
tachometer signal is produced using the rotor shaft angle. Torque components will be captured, thus likely leading to a worse signal
control mode (TCM) simulations have all been run with an average to noise ratio in the demodulated diagnostic signal. This phenom-
wind speed of 10 m/s (only varying around this nominal level in enon will in general pose a challenge to the ESA-based fault
case of turbulence), whereas an average wind speed of 20 m/s has detection especially when the converter-induced noise level is
been selected for the Pitch control mode (PCM) simulations. comparable with the expected fault signature magnitude.
5.1. Effect of converter switching on generator current 5.2. Influence of wind speed variation
This section is aimed at identifying the effect of converter The mechanical fault signatures are usually phase-locked with
switching on electrical measurements (i.e. generator current) by the associated shaft rotation. As a result, fault signature frequency
comparing the outputs of WT model B (with power electronics) changes with rotor speed caused by the wind speed variation. A
with those of WT model A (without power electronics) in case of rapid variation in the wind speed is usually caused by wind tur-
non-turbulent wind profile. As shown by Fig. 10 (a, b) the current bulence (stochastic part of the wind model). Two different wind
signal is less distorted in the WT model A as its generator feeds profiles with considerably different turbulence levels are utilized in
directly to a static load, whereas in WT model B an active rectifier simulating WT model B to investigate the consequence of wind
controls power flow through IGBT switching. The effect of this variability on generator measurements. During simulation, the
switching phenomenon (non-linear load), although partially turbulence levels are achieved by setting the turbulence intensity
filtered by the generator inductance, results in a considerable (TI) parameter of the Kaimal filter at 1 and 3 respectively [18].
distortion of the current signal as evident in Fig. 10 (c, d). In the The effect of the two wind turbulence scenarios on generator
latter, the presence of strong generator-frequency sidebands current spectra is shown in Fig. 11 under both TCM and PCM. Two
around the converter switching frequency (4 kHz) and an overall areas of the spectra are shown: the first (Fig. 11 a, b) corresponds to
increase of the noise level (even at low frequency) are obvious in the neighbourhood of the generator fundamental frequency, i.e. 2
the spectra of the current signals. This reduces the signal-to-noise times the generator shaft speed and 190 times the rotor shaft
Fig. 10. Generator current in WT model B and WT model A in non-turbulent wind conditions: (a) time domain, TCM, (b) time domain, PCM, (c) spectrum, TCM, (d) spectrum, PCM.
Fig. 11. Frequency spectra of generator current signals for different turbulence levels (TI ¼ 1 and TI ¼ 3) for WT model B: (a) zoom around fundamental current harmonics, TCM, (a)
zoom around fundamental current harmonics, PCM, (c) zoom around GMF sideband, TCM, (c) zoom around GMF sideband, PCM.
speed, respectively for TCM and PCM; the second (Fig. 11 c, d) fo- In turn, this is expected to reduce the effectiveness of ESA-based
cuses on the GMF sideband of the current signal, i.e. at 2 þ 20 times diagnostics of low-frequency drivetrain faults, such as mass and
the generator shaft speed and 2090 times the rotor shaft speed (see aerodynamic imbalance. This adverse effect of turbulence-induced
Table 2 for details). In both cases, the main effect recognisable noise on the current signal is slightly less severe in the PCM case,
directly in the current signal spectrum is the smearing of the where the controller attempts to regulate speed. The GMF side-
fundamental generator harmonics, due to the frequency modula- bands (Fig. 12 c, d), are not significantly affected by the turbulence-
tion effect of the wind-induced speed fluctuation. In case of TCM, induced noise, being at a much higher order, and are therefore
for both turbulence scenarios the fundamental us harmonics is clearly visible at 2090xRev in both turbulence scenarios and control
significantly smeared (more for the higher turbulence case) and the modes.
GMF sideband is completely compromised by the speed fluctua-
tion. However, in case of PCM, the controller tries to keep the rotor
speed stable result in alleviating the smearing effect significantly.
As a result, the GMF sideband becomes visible in this case. 6. ESA-based fault diagnosis
The order tracking procedure is alleviating the problem, as
shown in Fig. 12 (corresponding to the signals presented in Fig. 11). The electromechanical models WT model B and WT model A are
In both turbulence cases, the 3xRev sidebands around the funda- utilized in this section to investigate the performance of ESA-based
mental electrical order become clearer after order tracking, with an diagnosis for four different fault types, using the procedure dis-
almost complete correction of the smearing in both cases (Fig. 12 a, cussed in Section 4.2.
b). However, a secondary effect of wind turbulence becomes visible In each of the following sub-sections the specific parameters of
after the correction of its frequency modulation. In particular, the this fault-detection procedure are provided for each fault simula-
amplitude modulation of the speed, which varies according to tion. Three scenarios will be investigated for each fault type.
wind-turbulence, cannot (and should not) be corrected by order
tracking. This results in a low-frequency noise in the speed signal The potential masking influence of power electronic switching
and a consequent narrow-band noise in the current order-spectra will be studied by comparing results obtained with WT model B
around the electrical fundamental frequency. and WT model A, both without wind turbulence.
Fig. 12. Order spectra of generator current signals for different turbulence levels (TI ¼ 1 and TI ¼ 3) for WT model B: (a) zoom around fundamental current harmonics, TCM, (b) zoom
around fundamental current harmonics, PCM, (c) zoom around GMF sideband, TCM, (c) zoom around GMF sideband, PCM.
The wind-turbulence disturbance will be then investigated us- 6.1. Aerodynamic imbalance
ing WT model A only (this model does not include power elec-
tronics), in non-turbulent vs turbulent wind conditions. As mentioned in Section 3.1, an aerodynamic imbalance results
The combined effect of power-electronics and wind turbulence in an amplitude modulation of the 3xRev harmonics of the rotor
will be finally discussed, based on the results of WT model B with speed signal (caused by tower shadow and wind shear) by a 1xRev
and without wind turbulence. component (rotational speed of faulty blade). ESA-capabilities will
therefore be evaluated in this section by the visibility of the
For the sake of brevity, only the severe case of turbulence in- 3xRev ± n 1xRev sideband pattern (characteristic of this fault) in
tensity TI ¼ 3 is used in the following examples. A non-turbulent the demodulated signal obtained from the order-tracked generator
case is represented by TI ¼ 0 in rest of the discussion. current. The order tracked current signal is filtered at around the
fundamental order (190xRev) with a pass-bandwidth of 11xRev
Fig. 13. Order spectra of current envelope signals in WT model A and WT model B for aerodynamic imbalance case bm3f ¼ 3 and non-turbulent wind condition: (a) TCM, (b) PCM.
(±5.5xRev, to preserve the 3xRev peak and two 1xRev sidebands) showed that the turbulence induces low order noise in the speed
and then amplitude demodulated. (i.e. current envelope) thus hiding low order fault signature.
6.1.1. Converter switching effect 6.1.3. Combined result of the two masking effects
The order spectrum of the demodulated current signal in case of Based on the above discussion, It is expected that, under influ-
WT model B (red) and that of WT model A (blue) are compared in ence of both switching- and turbulence-induced noise, detection of
Fig. 13 to demonstrate the effect of power-electronic switching. It is this fault type would only be possible when the severity is high
clear that the switching effect greatly compromises the detect- enough (possibly the PCM case). This is demonstrated in Fig. 15 by
ability of the fault signature especially when the fault severity is comparing results of non-turbulent (blue) and turbulent (red) wind
low (TCM case, Fig. 13 a). A higher detectability in the PCM case conditions for WT model B. However, due to the simultaneous noise
(Fig. 13 b) comes from the fact that, in this control mode the blades effect, only the fault characteristic sidebands, 1xRev and 2xRev are
experience higher wind speeds, thus, increasing the magnitude of clearly visible in both turbulent condition in case of PCM (Fig. 15 b)
the aerodynamic imbalance effect (signal to noise ratio is while the detectability of the signature in turbulent conditions is
improved). not experienced at all in the TCM case (Fig. 15 a).
Summarising, these simulations have shown that the defect
6.1.2. Wind turbulence effect signature of aerodynamic imbalance is not only affected by power
Having discussed the effect of converter switching, it is possible converter electrical noise but is also sensitive to wind-turbulence-
to investigate the effect of the wind turbulence alone by using WT induced speed fluctuations. Therefore the two effects result in an
model A, which does not include power electronics. Fig. 14 com- additive masking of the fault signature for this fault.
pares order spectra of current envelope in non-turbulent (blue) and
turbulent (red) wind conditions. The fault characteristic 6.2. Mass imbalance
3xRev ± n 1xRev sidebands are clearly visible in the non-
turbulent condition for both control modes (blue lines in Fig. 14 Mass imbalance, simulated by increasing the mass of a single
a, b). However, the detectability of the signature in turbulent con- blade, causes a one-per-revolution (1xRev) disturbance in the
ditions is limited to the PCM case (red line of Fig. 14 b) as the fault generator speed. This disturbance is expected to show as an
severity becomes higher at this wind speed. The negative result of amplitude modulation of the current signal, which should present
TCM in turbulent condition (red line of Fig. 14 a) confirms the 1xRev sidebands around the fundamental electrical order of
conclusions drawn in the healthy situation (Section 5.2), which 190xRev. To extract the imbalance-related speed variation, the
Fig. 14. Order spectra of current envelope signals in TI 0 and TI 3 wind cases for aerodynamic imbalance case bm3f ¼ 3 in WT model A: (a) TCM, (b) PCM.
Fig. 15. Order spectra of current envelope signals in TI 0 and TI 3 wind cases for aerodynamic imbalance ðbm3f ¼ 3 Þ in WT model B: (a) TCM, (b) PCM.
signals discussed in this section are subject to the same order- 6.2.2. Wind turbulence effect
tracking and demodulation procedure as in the aerodynamic As for the previous section, the analysis of the effect of wind-
imbalance case (previous section). Given the similar nature of the turbulence is conducted using WT model A. Fig. 17 presents the
two imbalances (in terms of effect on the shaft speed), similar current envelope order spectra in different combinations of tur-
findings are expected for this type of faults. bulent and non-turbulent wind conditions, low and high imbal-
ance, and PCM/TCM. By comparing turbulent (red) and non-
turbulent (blue) simulations it is clear that wind-turbulence is
6.2.1. Converter switching effect significantly compromising ESA-based diagnostics of mass imbal-
The first analysis of ESA-robustness is conducted to identify the ance in both control modes. As expected, the turbulence-induced
masking effect of power electronics. Therefore, mass imbalance noise is comparatively low in PCM, but the speed adjustment ef-
simulations without wind turbulence are performed with WT fect for a low severity fault (red lines in Fig. 17 b) can make the fault
model B (red) and WT model A (blue), and presented for comparison detection impossible even in this control mode.
in Fig. 16. In the low severity case, the 1xRev signature peak is
evident for WT model A (blue lines in Fig. 16 a, b). On the other hand,
when using WT model B, the detection is affected by power elec- 6.2.3. Combined result of the two masking effects
tronics noise in both control modes, shown by the increased noise The simultaneous effect of converter- and wind turbulence-
level (red lines in Fig. 16 a, b). Even under this noise effect, the induced noise is investigated in Fig. 18 using WT model B. As
1xRev signature is fairly detectable in the TCM case as the evident, compared to the converter the wind turbulence introduces
controller permits this speed variation. On the contrary, the low more noise in the low order range, thus affecting the detectability
severity signature is not at all detectable in the PCM case because of low magnitude 1xRev peak. Moreover, as mentioned before, the
the signature (1xRev peak) is attenuated due to the controller detectability of low severity faults is also reduced by the controller
operation which reduces its magnitude below the noise level. effect in PCM (Fig. 18 b).
In case of a high imbalance level (5%) (Fig. 16 c, d), it is possible to Summarising, as for aero-dynamic imbalance, both the elec-
notice also the typical symptoms of aerodynamic imbalance, tronic switching of the converter and wind turbulence affect the
despite not having included any pitch control fault in the simula- ESA-performance significantly.
tion. This can be explained by the fact that a high imbalance causes
speed variations whose magnitude is sufficiently high to signifi- 6.3. Gear defect
cantly affect the tip speed ratio (and the generated torque), thus
resulting in an equivalent aerodynamic imbalance. This would not As discussed in Section 3.3, the gear defect is modelled on the
compromise detection of imbalance but could lead to a wrong HSS (generator shaft) by the reduction of stiffness of one of its
diagnosis of the problem. tooth, representing a tooth crack. As a result of this fault, a 1xRev
Fig. 16. Order spectra of current envelope signals in WT model A and WT model B for non-turbulent wind condition: (a) defect severity 0.1%, TCM, (b) defect severity 0.1%, PCM, (c)
defect severity 5%, TCM, (d) defect severity 5%, PCM.
Fig. 17. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model A: (a) defect severity 0.1%, TCM, (b) defect severity 0.1%, PCM, (c) defect severity 5%, TCM,
(d) defect severity 5%, PCM.
event occurs in the HSS speed, also modulating the GMF. As order xRev in the order tracked current signal. To investigate the effect of
tracking is performed by taking the rotor shaft as the reference, the power electronic switching on this defect signature, the same sig-
fault signature would appear in the order tracked HSS speed at nals obtained with WT model B and WT model A are filtered with a
n 95xRev and m GMF±n 95 xRev. In the current signal the pass-band of 192xRev around the fundamental electrical order
signature appears as sidebands distance by ±n 95xRev and ±m which is then amplitude demodulated. The results are presented in
GMF±n 95 xRev around the fundamental electrical order Fig. 20 for comparison. Regardless of the control mode, the defect
(190xRev) as listed in Table 2. signature at less severe case (5% reduction of stiffness) is
A primary way to detect gear defect signature is by demodu- completely masked by the switching noise.
lating the GMF. For this purpose the two-step demodulation pro-
cedure presented in Section 4.2 is applied. For the second 6.3.2. Wind turbulence effect
demodulation step a band-pass filter is set around the strongest A further investigation is done to measure the influence of wind
GMF harmonics (in this case the 1st GMF harmonics) with a band- turbulence on the ESA-based technique by simulating the same
width accommodating only few (in this case two) signature side- gear defects in the WT model A with both non-turbulent (blue) and
bands to restrict overlapping with the neighbouring GMF har- turbulent (red) cases and the results are presented in Fig. 21. This
monics. The resulting spectra would contain first two harmonics of analysis focuses on the 190±n 95 xRev signature, since the other
gear defect signature i.e. 95xRev and 190xRev. signature would be in any case compromised by the power-
electronics. As the fault signature occurs at higher shaft order, it
6.3.1. Converter switching effect is not expected to be affected by the low frequency noise intro-
The effect of switching noise on this detection method is duced by the wind turbulence. This is reflected in Fig. 21 for both
investigated by comparing results of WT model A and WT model B in control modes and severity cases.
Fig. 19. Regardless of the control mode, the signature sidebands at
190 þ GMF±n 95 xRev are evident in case of WT model A (blue 6.3.3. Combined result of the two masking effects
lines in Fig. 19 a, b) which are completely masked by the switching In Fig. 22, the simultaneous effect of switching and turbulence is
noise in case of WT model B (red lines in Fig. 19 a, b). As a result, the investigated, using WT model B, for the mentioned gear defect
detection of the defect signature in case of WT model B is not cases. Similar detectability is observed at non-turbulent (blue) and
possible at all. On the other hand, in case of WT model A the turbulent (red) wind condition which is expected as the influence
signature is clear in both AM and FM demodulation results as of turbulence is found negligible. Therefore, the only masking effect
shown in Fig. 19 (c, d, e, f) for both control modes. is imposed by the switching, which can considerably hide the less
As mentioned, other signatures of this defect would appear as severe defect signature.
sidebands of the fundamental electrical frequency at 190±n 95 Summarising, while the wind turbulence does not affect gear
Fig. 18. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model B: (a) defect severity 0.1%, TCM, (b) defect severity 0.1%, PCM, (c) defect severity 5%, TCM,
fault detection significantly (due to the high frequency location of fluctuation independently from the effect of converter switching,
the signature), power-electronics still constitute a problem for the Fig. 24 presents the results of the local fault simulations with WT
early detection of this type of faults. model A for both turbulent (red) and non-turbulent (blue) wind
cases. This result confirms the finding of section 6.3, i.e. a defect
6.4. Local defect signature of higher shaft order is not affected by wind turbulence.
Even the less severe defect case (10%) can be identified in both
Local defects, discussed in Section 3.4, are simulated as 4xRev control modes under turbulent wind.
impulsive events on the ISS shaft. Combining this with the gear
ratio between the rotor shaft and the ISS, the local defect is ex- 6.4.3. Combined result of the two masking effects
pected to produce a 67.26xRev signature in the speed signal, order Finally, in Fig. 25, simulation results of WT model B with and
tracked with respect to the rotor shaft. This produces a series of without turbulence are presented to demonstrate the simultaneous
sidebands in the generator's current signal at a distance of effect of the converter switching and the wind-induced speed
67.26xRev from the fundamental electrical order of 190xRev. To fluctuation on the detectability of discussed local defect cases. As
extract this signature, a pass-band of 140xRev is used around the expected, turbulence does not affect the detectability suggesting
fundamental electrical order to filter the order tracked current that the adverse effect on detectability only depends on the
signal which is then followed by amplitude demodulation. switching-induced noise.
6.4.1. Converter switching effect 7. Improvement of ESA-based diagnostics

The effect of converter switching on the detectability of local
defects is investigated using WT model A and WT model B with non- An alternative method to estimate generator shaft speed is by
turbulent wind. The obtained results for WT model A (blue) and WT taking time derivative of phase of the analytic signal ea ðtÞ (obtained
model B (red) are present in Fig. 23 for two different severity levels from eg ðtÞ, similarly as Eq. (27)), which, in fact, results in FM
and both control modes. According to the results, even a 10% demodulation of the electrical signal eg ðtÞ. Recalling Eq. (19), and
disturbance in the generator torque can be effectively masked by expressing the generator current as a function of time, the
the switching-induced noise and remain undetected. A moderately following expression is obtained:
higher torque variation (e.g. the 30% severity case) can be detected
in the current envelope.
eg ðqðtÞÞ ¼ fpug ðqðtÞÞcosðpqðtÞ (29)
In this case the demodulation is performed before order

6.4.2. Wind turbulence effect tracking, by extracting the phase of the analytic signal (rather than
In order to investigate the effect of turbulence induced speed its amplitude):
Fig. 19. Results of GMF demodulation-based gear defect detection in WT model A and WT model B for non-turbulent wind condition and 20% defect severity: (a) GMF with signature
sidebands in order tracked current spectra, TCM, (a) GMF with signature sidebands in order tracked current spectra, PCM, (c) AM demodulation of filtered current envelope, TCM,
(d) AM demodulation of filtered current envelope, PCM, (e) FM demodulation of filtered current envelope, TCM, (f) FM demodulation of filtered current envelope, PCM.
orders as discussed in Table 2 of Section 4.1. The overall technique of

argðea ðtÞÞ ¼ arg fpug ðqðtÞÞexpðjpqðtÞ ¼ pqðtÞ (30) estimated shaft speed-based fault diagnosis is presented in Fig. 26.
In this section, the detection capability of all the four faults are
Thus, an estimation of the shaft angular speed, of a synchronous
re-investigated using the FM-based estimation of the generator
generator with p pole pairs, can be obtained as follows.
shaft speed for the lowest severity cases of each fault type, using
1 d WT model B to observe the simultaneous effect of switching and
ðargðea ðtÞÞÞ ¼ q_ ðtÞ ¼ ug ðtÞ (31) turbulence. As shown in Fig. 26, the current signal is FM demodu-
p dt
lated to estimate the shaft speed, which is, then order tracked to
The estimated shaft speed, being affected by any drivetrain remove effect of mean shaft speed variation. The order spectra
mechanical phenomena, is expected to produce signature peaks in obtained after DFT is utilized for fault signature identification. The
its spectrum. Order tracking of the estimated shaft speed is then obtained simulation results are presented in Fig. 27 for the WT
necessary to mitigate the smearing of the peaks in the speed model B under both TCM (blue) and PCM (red) with the highest
spectrum, thus obtaining the angular domain speed signal ug ðqÞ, in level of wind turbulence (TI 3). In case of aerodynamic imbalance
which fault-induced harmonics are visible as peaks at characteristic (Fig. 27 a), this angular speed-based technique performs similarly
Fig. 20. Order spectra of current envelope signals in WT model A and WT model B for non-turbulent wind condition: (a) defect severity 5%, TCM, (b) severity 5%, PCM, (c) defect
severity 20%, TCM, (d) defect severity 20%, PCM.
Fig. 21. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model A: (a) defect severity 5%, TCM, (b) defect severity 5%, PCM, (c) defect severity 20%, TCM, (d)
defect severity 20%, PCM.
Fig. 22. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model B: (a) defect severity 5%, TCM, (b) defect severity 5%, PCM, (c) defect severity 20%, TCM, (d)
defect severity 20%, PCM.
Fig. 23. Order spectra of current envelope signals in WT model A and WT model B for non-turbulent wind case: (a) defect severity 10%, TCM, (b) defect severity 10%, PCM, (c) defect
severity 30%, TCM, (d) defect severity 30%, PCM.
Fig. 24. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model A: (a) defect severity 10%, TCM, (b) defect severity 10%, PCM, (c) defect severity 30%, TCM,
Fig. 25. Order spectra of current envelope signals in TI 0 and TI 3 wind cases in WT model B: (a) defect severity 10%, TCM, (b) defect severity 10%, PCM, (c) defect severity 30%, TCM,
to the envelope-based technique (Fig. 15), i.e. detection is not

possible for the TCM case. Similar detectability is also observed for
mass imbalance detection (Fig. 27 b vs Fig. 18 a). In case of both gear
defect (Fig. 27c) and local defect (Fig. 27 d), superior performance of
this technique is observed, especially in PCM case, compared to the
envelope-based technique (Figs. 22b and 25b respectively).
The superior result of frequency demodulation for the gear and
localised faults suggests a lower sensitivity of this technique to the
switching noise, which is the main issue affecting the diagnostics of
these two faults. This finding will be further investigated in future
studies.
8. Interaction of drivetrain faults with WT control system
In this section, interaction of the control system of WT model B

with the drivetrain faults has been investigated. As mentioned in
Section 2.3, the controller provides the torque and pitch references,
for WT operation, according the control modes. In this section, ef-
fects of the drivetrain faults on these control signals (torque and
pitch references) are investigated and the corresponding responses
of the systems under control are analysed. This investigation helps
to identify possible attenuation of magnitude of the fault signature
by the WT sub-systems. It also provides a preliminary investigation
on the possibility of detecting faults through the control signals.
8.1. Interaction of drivetrain faults with torque control system
The torque control system involves a complex cascade of control

blocks acting on mechanical and electrical quantities, as discussed
in the initial sections of this works. This brief section will start with
a first qualitative analysis of the effect of the non-linear control
block providing a “mechanical” torque reference (as discussed in
Section 2.3.1) to the generator torque (electrical) control system.
Then the effect of the more complex electrical control system will
be discussed and finally numerical simulation results will be pre-
sented to validate these qualitative considerations.
To qualitatively evaluate the effect of the first “mechanical”
control block on the fault detection, it is possible to assume a Fig. 26. Estimated shaft speed-based ESA technique for drivetrain fault diagnosis.
constant wind speed (below the rated level) and a small periodic
perturbation of generator torque/speed due to the fault. Under this
assumptions, the non-linear relationship of Eq. (9) can be linearized The response of the electrical generator system, represented by
into a proportional speed controller: Eq. (11) (Section 2.3.1), is plotted in Fig. 28 (a) and the corre-
sponding PI controller's response is presented in Fig. 28 (b). As
Kopt 2 discussed in Section 2.3, the torque control is aimed at maximizing
tg;r ðtÞ ¼ 2
ug þ Gp ug ðtÞ ug (32) the power production of the WT by controlling its generator torque.
Ng
For the analysis of this part, the simulation results were obtained in
where ug is the average generator shaft speed, ðug ug Þ represents TCM, when the controller was performing torque control while
Kopt ug setting the pitch reference as null.
the fault-induced perturbation and Gp ¼ 2 Ng2
represents the To investigate the case of low frequency fault signature, Fig. 29
equivalent proportional control gain. As a proportional controller, represents the order spectra of the generator speeds, torque ref-
the effect of the term Gp on the mechanical transfer function of the erences and generator torque signals for both healthy and mass
wind turbine simply adds an additional damping to the term Bg imbalance (0.1%) case. The 1xRev signature of mass imbalance is
(damping of the generator shaft). Therefore a magnitude compar- evident in the generator speed (Fig. 29 a). The fault signature is also
ison of the two terms Gp and Bg would provide a qualitative but contained in the torque reference (Fig. 29 b) as it is calculated from
highly informative indication on the significance of the controller in the generator speed by Eq. (9). Therefore, detection of imbalance is
attenuating the effect of fault-induced torque fluctuations on the possible by the torque reference signal. Moreover, as expected, the
generator speed. By using the numerical values adopted for the signature appears in the generator torque signal (Fig. 29 c). Com-
mechanical model (from Ref. [18]) and considering a range of parison of magnitudes of the fault signature, between Fig. 29 (b)
speeds ug up to the rated speed of 162 rad/s, the maximum value of and (c), ensures that the fault signature magnitude is not signifi-
the proportional coefficient Gp is approximately three orders of cantly attenuated by the generator system.
magnitude below the mechanical damping Bg .Therefore, signatures To investigate the case of a comparatively higher frequency
of the drivetrain faults are not expected to be affected by the torque signature, Fig. 30 represents the order spectra of the generator
control system unless they are affected by the response of the speed, torque reference and generator torque output for both
electrical system (generator and its controllers) or the converter healthy and gear defect (20% reduction of stiffness) cases. The
induced noise (discussed in Section 5.1). signature of the gear defect (harmonics of 95xRev, as discussed in
Fig. 27. Detection of drivetrain faults simulated on WT model B under turbulent wind (TI 3) for both control modes: (a) aerodynamic imbalance ðbm3f ¼ 3 Þ, (b) mass imbalance
(0.1%), (c) Gear defect (5% reduction of stiffness), (d) local defect (10% torque variation).
Fig. 28. System response: (a) torque response of generator, (b) PI controller.
Section 6.3), is evident in the speed signal (Fig. 30 a). This signature current signal-based gear fault detection (by GMF demodulation),
also appears in the torque reference (Fig. 30 b), which ensures the which has been discussed in Section 6.3.
possibility of detecting the fault by this control signal. In the
generator torque output (Fig. 30 c), the signature (95xRev) also 8.2. Interaction of drivetrain faults with pitch control system
appears with the same magnitude as torque reference, which en-
sures that it is also unaffected by the generator system. The responses of the pitch system and the associated PI
It should be mentioned here that, the GMF (1900xRev) is highly controller (as discussed in Section 2.3.2) are presented in Fig. 31.
affected by the converter induced noise as evident by Fig. 30 (c). The combined system (the pitch system with its PI controller) is
Therefore, detection of gear fault by the GMF sidebands could not expected to behave as a low-pass filter. In PCM, the controller
be demonstrated. This noise effect is also encountered in case of operates to keep the generator speed constant at the rated value by
Fig. 29. Comparing order spectra between healthy and mass imbalance (0.1%) case: (a) generator speed, (b) torque reference, (c) generator torque.
controlling pitch angle of the pitch system. Therefore, a spectral the fault signature is not transferred to the order spectrum of the
component of the pitch reference, if contained also in the spectrum actual pitch angle (Fig. 33 c). As a result, unlike the mass imbalance
of the measured angle of the pitch system will result in reduction of case, the attenuation of the fault signature magnitude in the speed
magnitude of that component in the speed signal (compared to the signal is not expected here. This is verified by comparing the equal
TCM case when pitch reference is null). magnitude of the gear fault signature (95xRev) in the speed signal
The same faults (mass imbalance and gear defect), discussed (Fig. 33 a) with that of TCM case (Fig. 30 a).
above, have also been investigated here to reveal their effect on the
pitch control signal and at the same time, identify possible atten- 9. Conclusions
uation by the pitch system. For this analysis, the simulation results
were obtained in PCM, when the controller was actively controlling This paper investigates the feasibility of ESA-based mechanical
the pitch angles of the blades. fault diagnosis in wind turbines under the effect of wind turbulence
In case of mass imbalance (0.1%), the order spectra of generator and converter switching. For this purpose, a full electromechanical
speed, pitch angle reference and pitch angle measurement (of the model of a wind turbine has been developed, integrating for the
pitch system) are provided in Fig. 32 and compared with that of first time models of drivetrain faults. Thanks to the completeness of
healthy case. The imbalance signature (1xRev) is clearly visible in the electro-mechanical model, the masking effects of power con-
the pitch reference (Fig. 32 b). As depicted in Fig. 32 (c), the order verter switching and wind-induced speed fluctuations on ESA-
spectrum of pitch measurement, in faulty condition, contains the based diagnostics are also investigated. Simulation results of four
1xRev component (signature of mass imbalance). As a result, the typical drivetrain faults demonstrated how their detectability can
magnitude of the 1xRev component in the generator speed (Fig. 32 be limited by these factors. In particular, converter switching affects
a) is reduced compared to the similar fault severity case in TCM all the defect signatures, due to its broad-band nature, whereas
(Fig. 29 a). This magnitude reduction has already been mentioned wind-turbulence masking is limited to the low frequency signa-
in Section 6.2 while discussing the current signal-based detection tures. The ability to model different level of fault severity also make
of mass imbalance. the developed simulation tool a powerful instrument to assess the
To investigate the effect of pitch control system on the detect- relative significance of the masking effects and the limitations they
ability of gear defect, Fig. 33 presents the order spectra of the impose in pushing forward the detection of incipient faults. A
generator speed, pitch reference and pitch angle measurement (of partial improvement in mitigating the masking effect, was identi-
the pitch system) for both healthy and gear defect cases. The order fied in the possibility of substituting amplitude with frequency
spectrum of the pitch reference clearly shows the signature of a demodulation, which shows higher robustness to power elec-
gear defect (harmonics of 95xRev) as shown in Fig. 33 (b). However, tronics noise. Finally, the possible attenuation of the magnitude of
due to the lower bandwidth of the pitch system transfer function, low frequency fault signatures by the pitch system has been
Fig. 30. Comparing order spectra between healthy and gear defect (20% reduction of stiffness) case: (a) generator speed, (b) torque reference, (c) generator torque.
Fig. 31. System response: (a) pitch system, (c) PI controller.
demonstrated. The feasibility of detecting drivetrain faults using chain) might also affect the performance of ESA-based methods
control signals has also been demonstrated. and would require experimental analysis. Moreover, the fault
To further detail the effectiveness of ESA-based diagnostics, models presented in this paper would benefit from future experi-
there are a number of issues that would benefit from a coupling of mental research to perform a detailed calibration of their magni-
this model with future experimental studies. Apart from the con- tude. This would allow a more precise quantification of the
verter and the wind induced speed disturbance, other sources of detectability of incipient faults with ESA, and the related estimation
noise (e.g. electro-magnetic interference in the measurement of fault severity and progression. These issues represent a future
Fig. 32. Comparing order spectra between healthy and mass imbalance (0.1%) case: (a) generator speed, (b) reference pitch angle, (c) measured pitch angle.
Fig. 33. Comparing order spectra between healthy and gear defect (20% reduction of stiffness) case: (a) generator speed, (b) reference pitch angle, (c) measured pitch angle.
field of investigation, where the proposed electromechanical model

WT model B can be utilized as a simulation tool to validate diag- jea ðqÞj ¼ fpfu þ A cosðqÞg A.8
nostic methodological improvements. After removing the mean value from jea ðqÞj:
jea ðqÞj hjea ðqÞj it ¼ fpA cosðqÞ A.9

Acknowledgements
where h$it is a mean operator.
Computational resources and services used in this work were The envelope signal (after mean removal) is proportional to the
provided by the HPC and Research Support Group, Queensland mass imbalance signature, in terms of magnitude, as demonstrated
University of Technology, Brisbane, Australia. by Eq. (A.9). Therefore, utilizing the envelope signal, it is possible to
detect mass imbalance.
Appendix A. Analytical proof of ESA-based diagnostic method

References
To demonstrate validity of the proposed ESA-based method, an [1] J. Carroll, A. McDonald, D. McMillan, Failure rate, repair time and unscheduled
analytical proof is provided here showing how the fault signature O&M cost analysis of offshore wind turbines, Wind Energy 19 (6) (2016)
can be extracted from the generator electrical output. As an 1107e1119.
[2] Z. Gao, C. Cecati, S.X. Ding, A survey of fault diagnosis and fault-tolerant
example of the drivetrain faults, the case of mass imbalance has techniques Part I: fault Diagnosis with model-based and signal-based ap-
been considered here. For this analytic proof, envelope-based ESA, proaches, IEEE Trans. Ind. Electron. 62 (6) (2015) 3757e3767.
shown in Fig. 9, is considered. [3] Z. Gao, C. Cecati, S.X. Ding, A survey of fault diagnosis and fault-tolerant
techniques - Part II: fault Diagnosis with knowledge-based and hybrid/
According to the discussion of Section 3.2, in presence of mass
active approaches, IEEE Trans. Ind. Electron. 62 (6) (2015) 3768e3774.
imbalance, the generator shaft speed, in the angular domain, can be [4] R. Isermann, Fault-diagnosis Applications, Springer, 2011.
expressed as follows: [5] X. Liu, Z. Gao, M. Chen, Takagi-sugeno fuzzy model based fault estimation and
signal compensation with application to wind turbines, IEEE Trans. Ind.
uðqÞ ¼ u þ A cosðqÞ þ u0 ðqÞ A.1 Electron. PP(99) (2017) 1.
[6] H. Sanchez, T. Escobet, V. Puig, P.F. Odgaard, Fault diagnosis of an advanced
wind turbine benchmark using interval-based ARRs and observers, IEEE Trans.
where q is the shaft angle, u is the average shaft speed, A is the Ind. Electron. 62 (6) (2015) 3783e3793.
rquez, A.M. Tobias, J.M. Pinar Pe rez, M. Papaelias, Condition
amplitude of the fault induced speed variation and u0 ðqÞ is a result [7] F.P. García Ma
monitoring of wind turbines: techniques and methods, Renew. Energy 46 (0)
of other drivetrain phenomena. Here, A cosðqÞ is the fault signature (2012) 169e178.
of mass imbalance. [8] C.J. Crabtree, D. Zappala, P.J. Tavner, Survey of commercially available condi-
The generator electrical output, in presence of mass imbalance, tion monitoring systems for wind turbines, in: Durham University School of
Engineering and Computing Sciences and the SUPERGEN Wind Energy
can be expressed as: Technologies Consortium, 2014.
[9] W. Yang, P.J. Tavner, C.J. Crabtree, Y. Feng, Y. Qiu, Wind turbine condition
eg ðqÞ ¼ fpðu þ A cosðqÞ þ u0 ðqÞÞcosðpqÞ A.2 monitoring: technical and commercial challenges, Wind Energy 17 (5) (2014)
673e693.
[10] N. Perisi
c, P.H. Kirkegaard, B.J. Pedersen, Cost-effective shaft torque observer
which can be rewritten as: for condition monitoring of wind turbines, Wind Energy 18 (1) (2015) 1e19.
[11] P.F. Odgaard, J. Stoustrup, Gear-box fault detection using time-frequency
A A based methods, Annu. Rev. Control 40 (2015) 50e58.
eg ðqÞ ¼ fpucosðpqÞ þ fp cosfðp þ 1Þqg þ fp cosfðp 1Þqg [12] W. Yang, P.J. Tavner, C.J. Crabtree, M. Wilkinson, Cost-Effective condition
2 2
monitoring for wind turbines, Ind. Electron. IEEE Trans. 57 (1) (2010)
þ fpu0 ðqÞcosðpqÞ 263e271.
[13] W. Qiao, Recovery Act: Online Nonintrusive Condition Monitoring and Fault
A.3 Detection for Wind Turbines, Faculty Publications from the Department of
Applying band-pass filter on eg ðqÞ, to eliminate the last term Electrical Engineering, Univ. of Nebraska-Lincoln, 2012.
[14] W. Yang, C. Little, R. Court, S-Transform and its contribution to wind turbine
(related to drivetrain phenomena) of Eq. (A.3): condition monitoring, Renew. Energy 62 (0) (2014) 137e146.
[15] J. Jonkman, S. Butterfield, W. Musial, G. Scott, Definition of a 5-MW Reference
A A Wind Turbine for Offshore System Development, National Renewable Energy
egf ðqÞ ¼ fpucosðpqÞ þ fp cosfðp þ 1Þqg þ fp cosfðp 1Þqg Laboratory, 2009.
2 2 [16] M. Singh, E. Muljadi, J. Jonkman, V. Gevorgian, I. Girsang, J. Dhupia, Simulation
A.4 for Wind Turbine Generatorsewith FAST and MATLAB-simulink Modules,
National Renewable Energy Laboratory (NREL), Golden, CO, 2014.
The Hilbert transform of egf ðqÞ can be expressed as: [17] D.S. Ochs, R.D. Miller, W.N. White, Simulation of electromechanical in-
́
teractions of permanent-magnet direct-drive wind turbines using the FAST
A A aeroelastic simulator, Sustain. Energy IEEE Trans. PP(99) (2013) 1e8.
egf ðqÞ ¼ fpusinðpqÞ þ fp sinfðp þ 1Þqg þ fp sinfðp 1Þqg
~ [18] P.F. Odgaard, J. Stoustrup, M. Kinnaert, Fault-tolerant control of wind turbines:
2 2 a benchmark model, control systems technology, IEEE Trans. 21 (4) (2013)
A.5 1168e1182.
[19] P.F. Odgaard, K.E. Johnson, Wind turbine fault detection and fault tolerant
Then the analytic signal, ea ðqÞ ¼ egf ðqÞ þ j~
egf ðqÞ can be calcu- control - an enhanced benchmark challenge, in: American Control Conference
lated as: (ACC), 2013, 2013, pp. 4447e4452.
[20] B. Wu, Y. Lang, N. Zargari, S. Kouro, Power Conversion and Control of Wind
Energy Systems, John Wiley & Sons, 2011.
A
ea ðqÞ ¼ fpuexpðjpqÞ þ fp expfjðp þ 1Þqg [21] M.R. Shahriar, P. Borghesani, A.C.C. Tan, Speed-based diagnostics of aero-
2 dynamic and mass imbalance in large wind turbines, in: Advanced Intelligent
A Mechatronics (AIM), 2015 IEEE International Conference on, 2015, pp.
þ fp expfjðp 1Þqg A.6 796e801.
2 [22] H. Luo, C. Hatch, M. Kalb, J. Hanna, A. Weiss, S. Sheng, Effective and accurate
approaches for wind turbine gearbox condition monitoring, Wind Energy 17
Eq. (A.6) can be further simplified as: (5) (2014) 715e728.
[23] Y. Cai, Simulation on the rotational vibration of helical gears in consideration
ea ðqÞ ¼ fpfu þ A cosðqÞgexpðjpqÞ A.7 of the tooth separation phenomenon (a new stiffness function of helical
involute tooth pair), J. Mech. Des. 117 (3) (1995) 460e469.
Taking the absolute value of ea ðqÞ: [24] F.M. Alemayehu, S. Ekwaro-Osire, Loading and design parameter uncertainty
in the dynamics and performance of high-speed-parallel-helical-stage of a of cracks and pitting defects on gear meshing, Proc. Inst. Mech. Eng. Part C J.
wind turbine Gearbox1, J. Mech. Des. 136 (9) (2014) 091002. Mech. Eng. Sci. 226 (11) (2012) 2805e2815.
[25] J. Hedlund, A. Lehtovaara, A parameterized numerical model for the evalua- [28] A. Bourdon, H. Andre, D. Re
mond, Introducing angularly periodic disturbances
tion of gear mesh stiffness variation of a helical gear pair, Proc. Inst. Mech. in dynamic models of rotating systems under non-stationary conditions,
Eng. Part C J. Mech. Eng. Sci. 222 (7) (2008) 1321e1327. Mech. Syst. Signal Process. 44 (1e2) (2014) 60e71.
[26] C. Zhu, S. Chen, H. Liu, H. Huang, G. Li, F. Ma, Dynamic analysis of the drive [29] Y. Xu, D. Yan, The Bedrosian identity for the Hilbert transform of product
train of a wind turbine based upon the measured load spectrum, J. Mech. Sci. functions, Proc. Am. Math. Soc. 134 (9) (2006) 2719e2728.
Technol. 28 (6) (2014) 2033e2040. [30] K. Fyfe, E. Munck, Analysis of computed order tracking, Mech. Syst. Signal
[27] A.F.d. Rincon, F. Viadero, M. Iglesias, A. de-Juan, P. Garcia, R. Sancibrian, Effect Process. 11 (2) (1997) 187e205.

11 - Performance Analysis of Electrical Signature Analysis-Based Diagnostics Using An Electromechanical Model of Wind Turbine

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

11 - Performance Analysis of Electrical Signature Analysis-Based Diagnostics Using An Electromechanical Model of Wind Turbine

Uploaded by

Copyright:

Available Formats

Renewable Energy xxx (2017) 1e27

Contents lists available at ScienceDirect

Performance analysis of electrical signature analysis-based diagnostics

1. Introduction require an explicit model of the monitored system to perform fault

components: i) an instantaneous mean wind speed (slow variation

Nominal voltage 690 V

The current output of the generator is fed to a three phase

A PI controller is usually employed to produce the q-axis refer-

2.3.2. Pitch control mode (PCM)

smoother power transmission. It is assumed that the proposed

where, kp is the stiffness value at the pitch point, Ca is an empirical

where, nd is the defect frequency in events per revolution, and q is a

4. Principle for ESA-based fault diagnosis

where ug represents the average (nominal) generator shaft speed

Tower shadow Multi-harmonic 3 3 190±3

representation of eg ðqÞ discussed in the previous chapter.

4.2. Signal processing for ESA-based diagnosis

The common approach to extract the speed information from

and the analytic signal ea ðqÞ ¼ eg ðqÞ þ j~

The absolute value of the analytic signal ea ðqÞ (i.e. envelope of

This analytical derivation is exact in the angular/order domain of

6.4.1. Converter switching effect 7. Improvement of ESA-based diagnostics

In this case the demodulation is performed before order

 orders as discussed in Table 2 of Section 4.1. The overall technique of

to the envelope-based technique (Fig. 15), i.e. detection is not

8. Interaction of drivetrain faults with WT control system

In this section, interaction of the control system of WT model B

8.1. Interaction of drivetrain faults with torque control system

The torque control system involves a complex cascade of control

Fig. 31. System response: (a) pitch system, (c) PI controller.

ﬁeld of investigation, where the proposed electromechanical model

jea ðqÞj hjea ðqÞj it ¼ fpA cosðqÞ A.9

Appendix A. Analytical proof of ESA-based diagnostic method

You might also like

orders as discussed in Table 2 of Section 4.1. The overall technique of