Mechanical Systems and Signal Processing 54-55 (2015) 491–505

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Modelling of a chaotic load of wind turbines drivetrain$

Andrzej Bielecki a,n, Tomasz Barszcz b, Mateusz Wójcik c
a
AGH University of Science and Technology, Faculty of Electrical Engineering, Automation, Computer Science and Biomedical Engineering,
Chair of Applied Computer Science, Al. Mickiewicza 30, 30-059 Cracow, Poland
b
AGH University of Science and Technology, Faculty of Mechanical Engineering and Robotics, Chair of Robotics and Mechatronics,
Al. Mickiewicza 30, 30-059 Cracow, Poland
c
Jagiellonian University, Faculty of Physics, Astronomy and Applied Computer Science, Reymonta 4, 30-059 Cracow, Poland

a r t i c l e i n f o abstract

Article history: The purpose of this paper is to present a model of the load of the wind turbine gears for
Received 27 April 2013 simulation of real, varying operational conditions for modelling of wind turbine vibration.
Received in revised form The characteristics of the wind, which generates chaotically varying loads on the
28 April 2014
drivetrain components generating load in teeth and bearings of gears during torque
Accepted 27 May 2014
Available online 16 July 2014
transfer, are discussed. A generator of variable load of wind turbines drivetrain is
proposed. Firstly, the module for generation of wind speed is designed. It is based on
Keywords: the approach in which the wind speed was considered as a time series approximated by
Wind turbines the Weierstrass function. Secondly, the rotational speed of the main shaft is proposed as a
Variable load generation
function of the wind speed value. The function depends on a few parameters that are
Model of the main shaft torque
fitted by using a genetic algorithm. Finally, the model of torque of the main shaft is
introduced. This model has been created by using a multi-layer artificial neural network.
The results show that the proposed approach yields a very good fit for the experimental
data. The fit brings about the proper reproducing of all the aspects of the load that are
crucial for causing fatigue and, as a consequence, damaging of gears of the wind turbines.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

We have been witnessing a rapid growth of the wind power generation all over the world for recent years. In the recent
15 years the actual cumulative growth rate was about 28%. The installed wind farms had reached the capacity of 237 GW by
the end of 2011. It is going to increase in future and in 2020 it will have achieved 587 GW, according to the pessimistic
scenario, 759 GW according to the moderate scenario, and 1150 GW according to the optimistic scenario. The decrease of the
amount of CO2 emission in 2020 will be 863 Mio t, 1118 Mio t and 1692 Mio t, respectively [1]. It is estimated that by 2020,
about 12% of the worlds electricity will have been supplied by wind generation [2].
Electric power from the wind energy is different from the one obtained from conventional resources. Unpredictability
and variability of the wind power generation are the fundamental difficulties that power system operators faced [3]. The
ability to balance the power grid, i.e. to match the amount of generated energy with the consumed energy, is the main
problem regarding this variability. Growing amount of the wind power sources creates significant pressure on the power
system control [4–6]. A combination of predicting the wind speed and controlling other units of power generation, in such a

☆
The paper was supported by the Polish Ministry of Science and Higher Education under Grant no. N504 147838.
n
Corresponding author.
E-mail addresses: azbielecki@gmail.com (A. Bielecki), tbarszcz@agh.edu.pl (T. Barszcz), mateusz.wojcik@uj.edu.pl (M. Wójcik).

http://dx.doi.org/10.1016/j.ymssp.2014.05.035
0888-3270/& 2014 Elsevier Ltd. All rights reserved.
492 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

Nomenclature dimH ðFÞ the Hausdorff dimension of the set F.

dimB ðFÞ the box-counting dimension of the set F.
List of mathematical symbols
f λ;s ðtÞ the Weierstrass function having parameters λ
and s.
diam U the diameter of the set U.
Hs ðFÞ the s-dimensional Hausdorff measure of the
set F.

way that they carry the burden of balancing the grid, is one of possible solutions to this problem. This is the main reason for
which the forecasting and modelling tools are being developed with the aim of the efficient reintegration of wind farms into
the power system [7,8]. Especially, forecasting of the power generated by wind farms is taken into consideration [9–12].
However, the model of wind speed variability, introduced in [13,14] and briefly recalled in this paper, see Section 3, was created
for quite other reasons. Studies showed that approximately 80% of all fractures were caused by machinery fatigue and only 20% by
a static overload. That is why the studies concerning variability of operational conditions in wind turbines are crucial for their
engineering. Therefore, modelling of the load of wind turbine gears for simulation of real varying operational conditions for
modelling of the wind turbine vibration is the main goal of this paper. This is a crucial problem for the development of algorithms
for fault detection which will operate successfully in the presence of large and random load changes. Such algorithms have
important practical application because maintenance of the wind turbine generates the largest part of the cost of its operation [15].
A condition monitoring, first of all continuous monitoring of the drivetrain of wind turbines, is a common technique to decrease
this cost [16–22]. Furthermore, condition monitoring of wind turbines, including fault diagnostics, especially at the early stage of a
fault occurrence, is an essential problem in the wind turbines engineering in particular [18,23–25] and in rotating machinery
engineering in general [26–32]. Development of condition monitoring systems requires validation of the applied algorithms for the
fault detection. This can be accomplished by using the data from real faults or from test rigs. Such data are very difficult, or even
impossible, to obtain and they cover only one or a few turbine types. In order to be able to collect practically unlimited amount of
the seeded fault data one can develop a numerical drivetrain model, where it is possible to introduce one or more faults and
various magnitudes. Using mathematical models for this task requires reconstruction of the real conditions in the model.
Modelling of gearboxes and bearings is an important field and it has been investigated by many authors [33–37]. Behaviour of the
drivetrain becomes much more complex when the machine is subjected to a varying load, as it is in the case of the wind turbines
(and many others, e.g. open pit mining). Modelling of such a machinery was researched for a relatively short time, with significant
contributions of [38–40]. Later, it has been followed by several other papers [41–45]. Analytical models for varying load require the
function which generates the load, or torque, to the model. Therefore, the research has been focused on the software module
which could act as a load generator to the drivetrain model.
The model of the chaotic variability of wind speed was proposed in [13,14]. The authors introduced the mathematical model
which is capable of generating series of the wind speed. In order to develop fault detection and validate the methods of fault
detection it is necessary to work out the adequate wind model which represents the most crucial components of the system input
load. Such a model should be an integral part of a cybernetic model for wind turbine. We intended to increase the understanding of
the wind variability. That was the reason why we did not use other – data driven – methods, like, for example, Markov chains,
spectral analysis or simply feeding the recorded wind data into the mechanical drivetrain model [46,7].
The creation of the model of input load of wind turbines drivetrain and the implementation of the simulator of input load were
the main goals of this paper. In order to apply the wind speed model to the model of the drivetrain load, the wind speed should be
converted into the torque on the shaft. Since the model is intended to simulate the load of turbine gears, not of rotor blades, the
wind speed is much more important parameter than its direction. Furthermore, modern wind turbines, including the ones from
which we received the data for simulation, have wind direction sensors that control yaw drives. The sensors enable the turbines to
orient automatically the blades perpendicular to wind direction. Additionally, it weakens the influence that the wind direction has
on turbine gears. Therefore, we could neglect the wind direction in the model without loss of its utility.
The paper is organized in the following way. In Sections 2 and 3 the wind speed model, introduced in [13,14], is briefly recalled.
The new results are presented in Section 4 which is the essential part of this paper. First of all, a module for generating wind speed
values is described – see Section 4.1. Modules for converting wind speed values into a rotational speed and a torque, and generating
drivetrain load are described in detail in Section 4.2. They are the second element of the introduced model of a chaotic load of the
mechanical systems of wind turbines. The comparison of the results obtained from the introduced model with the measured loads
is shown in Section 4.3. The discussion and conclusions are presented in Sections 5 and 6 respectively.

2. Wind speed model mathematical foundations

There are many examples of sets in Rn which have an irregular structure. The topological methods, however, are far insufficient
to describe the irregular structure in detail. The fact that the topological dimensions [47] can be equal only to integral numbers or
infinity is one of the important limitations. Such irregular sets are called fractals. Fractal dimensions, for which nonintegral values
are permitted, are the basic tools that allow us to express the irregular set properties accurately. It should be mentioned that the
idea of fractal dimension has nontrivial applications in the identification and the assessment of bearing degradation in rotating
 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 493

machinery – see [30,48]. The Weierstrass function, whose graph is a fractal set, is used by us for modelling the wind speed as a time
series. In this section the construction and basic properties of this function are briefly recalled. Two basic ideas of fractal dimension
(see [49–51] for details), which are used in fitting algorithm, are described as well.
Let | aU  Rn . The diameter of U is defined as diam U≔supfdðx; yÞ: x; y A Ug, where d is a metric on Rn . In addition, let us
put diam | ¼ 0. For any δ4 0; F  Rn and s 4 0 let us define
( )
1 1
Hsδ ðFÞ≔inf ∑ ðdiam U i Þs : F ¼ ⋃ U i with 0 odiam U i r δ :
i¼1 i¼1

The s-dimensional Hausdorff measure of F is defined as

Hs ðFÞ≔lim Hsδ ðFÞ: ð1Þ
δ-0

This limit exists for any F  Rn , though the limiting value can be equal to infinity.

Lemma 2.1. There is a unique non-negative real number dimH ðFÞ, called a Hausdorff dimension of the set F, such that
(
1 if 0 r s odimH ðFÞ;
Hs ðFÞ ¼
0 if dimH ðFÞ o s o1:

Formally, dimH ðFÞ≔inf fs: Hs ðFÞ ¼ 0g.

Though the Hausdorff dimension is useful in theoretical considerations, it is often difficult to calculate or even estimate it.
Therefore, some alternative definitions of dimension are introduced. The so-called box-counting dimension is one of the widely
used. Let N δ ðFÞ be the smallest number of sets of diameter at most δ which can cover F. The box-counting dimension is defined as
ln N δ ðFÞ
dimB ðF Þ≔lim ð2Þ
δ-0 ln δ
with the assumption that the above limit exists. We have the following relation between the two described fractal dimensions.

Lemma 2.2. Let F  Rn . With the assumption that dimB ðFÞ exists, the following inequality holds:
dimH ðFÞ rdimB ðFÞ:

Remark. The box-counting dimension of a given F can be easily estimated. Let us create several coverings of the set F by
minimal numbers of balls in such a way that in the ith covering the radii of balls are equal to δi . According to (2), let us treat
ln N δi ðFÞ as a linear function of ln δ1i . The straight line, estimated the set of points fðln δ1i ; ln N δi ðF ÞÞg in the best way, can be
obtained by the least squares method. The slope of the line approximates the dimB ðFÞ.

There are many real functions that have fractal graphs. One of them, the Weierstrass function, is utilized in this paper. Let
us suppose λ 4 1 and 1 o s o2. The Weierstrass function f : R-R, which has parameters λ and s, is defined as
1
f λ;s ðtÞ ¼ ∑ λðs  2Þk sin ðλk tÞ: ð3Þ
k¼1

Graphs of the Weierstrass functions for various values of parameters are shown in Fig. 1.
The function has the following properties.

Lemma 2.3. Let f λ;s be the Weierstrass function with parameters λ and s.

1. In each point the Weierstrass function is continuous and non-differentiable.

2. Provided that λ is large enough, dimB ðgraph f λ;s Þ ¼ s.
3. There is a constant c such that s  c=ln λ r dimH ðgraph f λ;s Þ r s.

3. Wind as a crucial factor of the wind turbines drivetrain load

The wind speed, considered as a time series, shows its very irregular character. Moreover, the output of the wind turbine,
i.e. the generated power, also reveals its irregular character – see the examples shown in Figs. 2 and 3.
That means the chaotic variability of the wind speed generates the overload on the drivetrain components by generating load in
teeth and bearings of gears during the torque transfer [28,52]. The fatigue of mechanical elements is the first step in the process of
damage arising [53,54]. In general, the fatigue of machine elements depends strongly on the range of variation of the stress, which
is imposed on the material, and that is why the stresses in mechanical parts of the wind turbines gears are caused mainly by
irregular changes of wind speed. Therefore, the wind speed parameters, that are the measure of these irregularities, should be
taken into consideration in all the models, that are created in order to simulate varying operational conditions in gears.
Furthermore, it should be stressed that the wind force varies with the square of the wind speed whereas the power in the wind
varies with the cube of the wind speed [8]. That means that small increase in the wind speed produces a large increase in its force
and power. Therefore, a very precise model of the wind speed is necessary in order to simulate properly an input load of the turbine
494 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

5 1.2

4 1

3 0.8
speed [m/s]

speed [m/s]
2 0.6

1 0.4

0 0.2

−1 0

−2 −0.2

−3 −0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time [s] time [s]

1 0.4

0.8 0.3

0.2
0.6
speed [m/s]

speed [m/s]
0.1
0.4
0
0.2
−0.1
0
−0.2
−0.2 −0.3

−0.4 −0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time [s] time [s]

Fig. 1. Weierstrass function with some pairs of parameters λ and s.

6
5.8
5.6
5.4
speed [m/s]

5.2
5
4.8
4.6
4.4
4.2
4
0 50 100 150 200 250 300 350 400 450 50
0
time [s]

Fig. 2. An example of measurement data of wind speed.

Fig. 3. Turbine generator output.

 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 495

drivetrain. Since the model of the wind speed involves magnitude of the wind speed, it can be used to distinguish the high and the
low cycle of the fatigue phenomena. The magnitude of cycles has enormous impact on the durability. It is well known behaviour,
often presented on the S–N or Woehler curve [55]. Such a phenomenon has fundamental impact on the lifetime of mechanical
components, especially gearboxes. There has been several documented failures of wind turbine gearboxes [33]. Proper modelling of
operational point variability can lead to more accurate prediction of the component lifetime. Furthermore, the fractal dimension is a
measure of the high frequency component, which – in the case of the wind turbines – causes high frequency variability of the
turbine torque. The torque variability causes variable stress, produced in mechanical components, especially in gearboxes. It should
be emphasized that focus on a specific type of damage has not been our intention. In fact, in mechanical systems three types of
fatigue occur: a low cycle fatigue, a high cycle fatigue and an overload one, as well as combinations of them. Our goal was to model,
as precisely as possible, the nature of both the wind and the load. As we showed, wind displays fractal behaviour with significant
high frequency content [13]. Therefore, on the basis of this model, the high cycle fatigue can be investigated more profoundly than
that only on the basis of low bandwidth wind models. If the model is sufficiently sophisticated, then the other types of damages
can be investigated as well. If the fractal dimension is higher than 1, the variability of stress can be higher than the one that was
assumed by the designer and the manufacturer. Thus, such operating conditions will cause faster wear of the gearbox and decrease
in the turbine remaining lifetime. Accurate modelling of the fractal dimension of the wind speed time series is crucial for prediction
of the lifetime of wind turbine gearboxes.
The real data that were recorded on a 1.5 MW wind turbine, located in Germany, have been used. The turbine was a typical
design for geared wind turbines. The main rotor with three blades is supported by the main bearing and it transmits the torque to
the planetary gear. The plate, to which the main rotor is connected, is the input of the planetary gear. The planetary gear has three
planets, with their shafts attached to the plate. The planets roll over the stationary ring and they transmit the torque to the sun. The
sun shaft is the output of the planetary gear. Further, the sun drives the two-stage parallel gear. The parallel gear has three shafts:
the slow shaft connected to the sun shaft, the intermediate shaft and the fast shaft. The last one drives the generator [67]. Therefore,
the wind turbine can be viewed as the multiplicator, which multiplies the rotor speed about 100 times in order to drive the
generator.
The data were available by courtesy of the SeaCom GmbH company from Herne, Germany. The measurement system
consisted of signal conditioning unit (PA8000D type form EC Electronics), data acquisition card (USB-6210 type from National
Instruments) and dedicated data acquisition software. The software was developed in the LabView environment and run on
the ARK-3384 embedded computer. The measured wind speed signal was acquired from the wind turbine controller.
In the model of wind speed both the wind acceleration and the fractal dimension of the wind speed graph are the crucial
parameters in the model evaluation [13]. The model is based on Weierstrass function and its parameters were fitted by using
genetic algorithm. The comparison of wind acceleration and fractal dimension of the model and experimental data was
crucial in the calculations of the fitness function values (see [13] for details). Such an approach allowed us to model very
precisely not only the wind speed values but also its variability – see [13] and Table 1.
It should be also emphasized that the mean fractal dimension is equal to 1.3552 with standard deviation 0.0920. It means that,
statistically, it is significantly greater than 1 and that the wind speed variability has not only high values but also chaotic character.
The proposed model of the wind speed, considered as a time series, is the basis of the simulator of the input load for wind turbine
drivetrain. The simulator, that has been designed and implemented by the authors of this paper, is described in the next section.

4. Module generating a variable load

A system module presented in this section has been designed to generate data which simulate the variable load of a wind
turbine. The load of the drivetrain is the torque which acts on the drivetrain and the generator. In general the torque
depends not only on the wind but also on dynamics of the rotor, the gearbox and the generator. The above-mentioned
simulation is a relatively complex task and its full description requires also a deep understanding of the turbine controller.
The controller reacts, among others, to the increased wind speed and after a few seconds it increases the rotational speed
and next the generated power. However, in the approach presented in this paper, the following steps have been followed:

1. We have developed the model which generates 1 s dynamic wind data by using Weierstrass function.
2. We developed the model which connects the wind speed with the rotational speed and the torque by using 15 min
current data. All the considered time constants are much shorter than 15 min so in the first approximation we decided to
neglect the dynamics of the system and to propose the static model.
3. We have validated the quality of our models by applying the static model to the real wind speed data measured onsite and

Table 1
Mean values of errors of approximation parameters.

Measured value Mean error Standard deviation of error

Speed value vi MAPE ¼ 3.8324% 1.2932%

Acceleration vi þ 1  vi MAE ¼0.1192 0.0548
Fractal dimension MAE ¼0.0021 0.0006
Fractal dimension MAPE ¼ 0.7041% 0.1987%
496 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

by comparing the real rotational speed and torque with the model output. It should be stressed that the data, which have
been used for the model validation, were collected with a specialized instrumentation with the time resolution of 1 s.
With such an assumption, the development of the module which connects the wind speed with the torque becomes a task
of developing of a static function between these two values. We have also assumed that the generator is connected to the
grid which is capable of using all the generated powers. Thus, according to the specified assumptions, wind fully determines
this variable load of a wind turbine whose main shaft can be represented as a rotation speed and a torque. In the next
subsection, the module for generating wind speed values is described. After that, the module for converting the wind speed
into the rotational speed and the torque is presented.

4.1. Module for generating wind speed values

The module for the wind speed generation has a four-component input. The inputs are maximum value, maximum
amplitude, mean value, and length. The module returns an array of time series values u(t) that are calculated according to
the proposed wind model [13]:
Kj
ðsj  2Þk
uj ðtÞ ¼ aj t þ bj þ ∑ λj sin ðλkj xðtÞÞ; ð4Þ
k¼1

where
 uj is defined separately for time intervals ½t j ; t j þ 1  that have various lengths,
 the affine part of the model: aj t þ bj represents a long-term trend on the time interval ½t j ; t j þ 1 ,
 x(t) transforms a point of time to a domain of Weierestrass function.
The module that generates wind speed consists of two main components. The first one, the base, is an internal
component that contains blocks of data and it is described in Section 4.1.1; whereas the second one, the generator, described
in Section 4.1.2, generates random blocks from the base and builds time series that corresponds the given input parameters.
The architecture of the wind speed generating module is shown in Fig. 4.

4.1.1. Base
The base is a container that stores blocks of the wind speed data used by the generating procedure (see Section 4.1.2).
Each block is a fragment of time series which is created by using the proposed wind speed model – see Eq. (3).
Blocks in the base are grouped in a specific way. The set of blocks is divided into k ¼ lmax lmin þ 1 groups according to
their domain length. As a result blocks in each group have the same length. Then, each group is divided into
m ¼ ðsmax smin Þ  100 þ 1 subgroups, smax ; smin A ð1; 2Þ, according to fractal dimension of the wind speed series – see Fig. 5 –
so that the fractal dimension (see Eq. (2)) of time series blocks in each group has the same value with the accuracy to the
second decimal place. In such a way the base is created as a tree data structure – see Fig. 6.
The process of the base building is based on Lemma 2.3 point 2 that states that for each s it is possible to find Weierstrass
function with a parameter s that has the fractal dimension equal to s only by increasing the λ parameter value. The algorithm
that creates the base is presented below.

Algorithm 1. Base building.

 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 497

4.1.2. Generating procedure

The generating procedure plays a crucial role, it generates a variable load by using blocks that are stored in the base. The
procedure transforms and joins them together into one time series – see Algorithm 2.

Algorithm 2. Generating data.

Algorithm 3. Function calculating max amplitude value.

function calculateMaxAmplitude(array)
result’0
for i’2 to lengthðarrayÞ do
jresult’maxðresult; jarray½i  array½i 1jÞ
end

Two functions, getRandomBlock and calculateMaxAmplitude, are used in the generating procedure. In GetRandomBlock
function random blocks are chosen from the base in two steps. In the first step, the length is selected with the uniform
distribution of probability which is given on the interval ½lmin ; lmax . Then, the fractal dimension is selected with a normal
distribution with the mean value equal to 1.38 and the standard deviation equal to 0.1. Parameters of the normal distribution
were estimated by using many real time series of wind speed of the length not less than 11 and not greater than 32 – these
values had been obtained experimentally because the wind model was fitted to real data in the best way for such values –
see [13].
498 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

Fig. 4. Architecture of the module.

1 0.8 0.6

0.8 0.6 0.4

0.4
0.6 0.2
magnitude

magnitude

magnitude
0.2
0.4 0
0
0.2 −0.2
−0.2
0 −0.4
−0.4
−0.2 −0.6 −0.6

−0.4 −0.8 −0.8

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
time time time

Fig. 5. The example of time series in blocks with different fractal dimension values: dA ¼ 1.2, dB ¼ 1.4, dC ¼ 1.6.

Fig. 6. The example of the base; the length of blocks is from 11 to 32, and the fractal dimension is from 1.10 to 1.80.

CalculateMaxAmplitude function finds maximum amplitude for a given array – see Algorithm 3.

4.1.3. Generation results

Some examples of the generated wind speed time series are presented in Figs. 7–9, whereas an example of real time
series is shown in Fig. 10. All the generated time series have the same length, equal to 300 s, and mean speed value, equal to
3.35. These examples were generated by using the base which has the following parameters: lmin ¼ 11, lmax ¼ 32, smin ¼ 1:2,
smax ¼ 1:8, xsum_step ¼ 0:2, xsum_count ¼ 10, xmul_step ¼ 3, and xmul_count ¼ 3. Time series presented in Figs. 9 and 10 have the same
mean speed values and maximum amplitude values.

4.2. Module for converting wind speed values into a rotational speed and a torque

The proposed models of the rotational speed and the torque are based on approximation of experimental data. The
approximation functions are continuous ones. According to the assumptions specified at the beginning of Section 4, the
characteristics of generating wind determines fully the variable load.
 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 499

4.5

3.5

2.5

2
0 50 100 150 200 250 300

Fig. 7. An example of variable load time series (solid line – wind speed, dashed line – separator of blocks). Input parameters values: mean_value ¼ 3.35,
max_amplitude ¼ 0.3, max_value ¼4.5, and length¼ 300.

4.5

3.5

2.5

1.5

1
0 50 100 150 200 250 300

Fig. 8. An example of variable load time series (solid line – wind speed, dashed line – separator of blocks). Input parameters values: mean_value ¼3.35,
max_amplitude ¼ 0.8, max_value ¼4.5, and length ¼300.

4.5

3.5

2.5

2
0 50 100 150 200 250 300

Fig. 9. An example of variable load time series (solid line – wind speed, dashed line – separator of blocks). Input parameters values: mean_value ¼ 3.35,
max_amplitude ¼ 0.61, max_value¼ 4.5, and length¼ 300.

4.5

3.5

2.5

2
0 50 100 150 200 250 300

Fig. 10. An example of real wind time series. The mean speed value is equal to 3.35 and max. amplitude ¼ 0.61.
500 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

4.2.1. Model of the main shaft rotational speed

The modelled wind turbine has only two dominating rotational speed levels, 1000 and 1500 rpm. According to the fact
that there are two main levels of rotor rotation speed, the model should consist of two components corresponding to each of
them. Therefore, it has been assumed that the main shaft rotational speed is a function S of the following form:
SðvÞ ¼ a1  arctanðb1  v þ c1 Þ þ a2  arctanðb2  v þ c2 Þ þ d; ð5Þ
where v is a wind speed value. The parameters a1 , b1 , c1 , a2 , b2 , c2 and d have been fitted by using a genetic algorithm. The
genomes consisted of seven chromosomes that correspond to the model parameters. Roulette-wheel selection has been
used as the procedure that chose the population of parents and a multiple-point crossover operator has been applied for
generating progeny population. The level of mutation was equal to 5%. The fitness function E was defined as
E ¼ ∑Mm ¼ 1 jSðvm Þ sm j, where Sðvm Þ is a theoretical value given by the model (5) whereas sm is the experimental data. The
best individual in the paternal population was handed over to the progeny population automatically i.e. the operator of
selection did not verify it. When the best individual remained the same in 200 epochs, half of the population was killed and
new individuals were created randomly. The algorithm was stopped after 1000 epochs. The obtained parameters have the
following values: a1 ¼ 5:98, b1 ¼ 190, c1 ¼ 377, a2 ¼ 4:2, b2 ¼ 1:5, c2 ¼ 11, and d¼894. The approximation of the wind
turbine main shaft rotational speed by using model (5) is visualized in Fig. 11 whereas an example of the experimental data
in comparison with the model data is shown in Fig. 15.

4.2.2. Model of the main shaft torque

According to the results, obtained at the turn of the 80s and 90s of the 20th century, based on Kolmogorov
approximation theorem [56], each continuous mapping can be approximated effectively on a compact set by a multi-
layer artificial neural network with only one hidden layer. The only assumptions are that the activation function of the
hidden neurons is the logistic ones and the activation function of the neurons at the output layer is linear – see [57–61].
Furthermore, the training processes of multi-layer neural network, based on gradient methods, are robust i.e. numerically
stable [62,63]. Therefore, this type of neural networks is an effective tool for finding approximations of a priori unknown
mappings. Thus, according to this approach, the torque can be expressed as a function T which has the following form:
N
  
2
T ðvÞ ¼ ∑  1  A i þB ð6Þ
i¼1 1 þe  2ðvai þ bi Þ

where v is the wind speed value. Formula (6) represents the neural network with N neurons in the hidden layer and the one
in the output layer – see Fig. 12. The parameters ai , bi and Ai , B represent the network weights. The activation function in a
hidden layer has the form f ðsÞ ¼ 2=ð1 þe  βs Þ  1, whereas in the output layer the activation function is given as gðrÞ ¼ r. The
network has been trained by using the Levenberg–Marquardt algorithm with various values of N. The best results have been
obtained for N ¼10. The visualization of the approximation is shown in Fig. 13. An example of the comparison of
experimental values of a torque and the values given by model (6) and calculated on the same domain, wind speed sample
(Fig. 14), is presented in Fig. 17.

4.3. Results

The postulated models were verified by using the measured real values of both shaft rotational speed and torque. The
values of the both mentioned variables and their increments i.e. ai : ¼ vi þ 1  vi and ΔTðvi Þ≔Tðvi þ 1 Þ Tðvi Þ; calculated by
using formulae (5) and (6), were compared with the measured real ones – see Figs. 15–18. Both the values obtained from the
model and the real measured values were taken for the same samples of wind speed. It should be stressed that formulae (5)

1800
1600
rotation speed [rpm]

1400
1200
1000
800
600
400
200
0
0 2 4 6 8 10 12 14 16 18 20
wind speed [m/s]

Fig. 11. The approximation of wind turbine main shaft rotation speed by formula (5). Blue dots represent real measured data. Red line represents the
approximated theoretical model output of the shaft rotation speed/ torque as the function of the wind speed. A specific region of data along 1–4 m/s and
1200–1300 rpm is caused by the dynamic behaviour of the wind turbine controller. Since the proposed method is based on static dependency between the
data, all the data are taken into account. The actual number of data points in the specific region is 670 data points, compared to 14 155 data point in a
region 1000–1100 rpm. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.)
 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 501

Fig. 12. The architecture on a neural network used for calculation of parameters of the model (6).

1.5
torque [kW/rpm]

0.5

0 2 4 6 8 10 12 14 16 18 20
wind speed [m/s]

Fig. 13. The approximation of rotational speed using model (6) with N ¼ 10. Blue dots represent the real measured data. Red line represents the
approximated theoretical model output of the shaft rotation speed/ torque as the function of the wind speed. (For interpretation of the references to colour
in this figure caption, the reader is referred to the web version of this article.)

11
10
9
8
wind speed [m/s]

7
6
5
4
3
2
1
0 50 100 150 200 250 300
time [s]

Fig. 14. The example of a real wind speed (a part of data).

and (6) have been introduced for modelling the variable load in the mechanical parts of wind turbines. Therefore, the
variability of parameters, especially chaotic ones that caused the biggest load and, as a consequence, gears fatigue, is more
crucial than their values. Therefore, the box counting dimensions for real time series (dim-meas) and for time series
obtained from models (dim-mod) were used as a measure of the approximation accuracy. For each set of time series the
mean box counting dimensions were calculated. Moreover, the standard deviations σ and the mean average percentage
errors, MAPE, of the dimensions were calculated. These values for the main shaft rotational speed, given by Eq. (5), are
presented in Table 2 whereas the corresponding results for the main shaft torque model, given by formula (6), are presented
in Table 3. Both the shaft rotational speed and the torque fractal dimension, both theoretical and measured, are equal to
around 1.5 with the standard deviation equal to 0.11 or less. This means that the chaotic character of wind speed variability
[13] causes chaotic variability of the main shaft rotational speed and the torque. In both tables the results were obtained for
wind samples measured every one second per 256 s and 512 s. There were 111 examples for 256 time period and 55 for 512
time period – parameter n in the tables. The data source is the same as the one used in Section 3.
502 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

1800
1600
1400

rotations peed [rpm]

1200
1000
800
600
400
200
0
0 50 100 150 200 250 300
time [s]

Fig. 15. The comparison of real rotation speed (blue line) and result of function (5) (red line) on the same domain – the example wind speed presented in
Fig. 14. It should be stressed that the discrepancy between the model and the data at the neighbourhood of the point t ¼250 s is correctly reconstructed but
it is slightly shifted in time. This shift is caused by the wide distribution of the experimental values around the model function values. (For interpretation of
the references to colour in this figure caption, the reader is referred to the web version of this article.)

1500

1000
rotation acceleration

500

−500

−1000

−1500
0 50 100 150 200 250 300
time [s]

Fig. 16. The comparison of real rotation increment (blue line) and result of function (5) (red line) on the same domain – the example wind speed presented
in Fig. 14. The note that concerns the discrepancy between the model and the data at the neighbourhood of the point t¼ 250 s, specified at the caption of
Fig. 15, remains valid. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.)

0.8
0.7
0.6
torque [kW/rpm]

0.5
0.4
0.3
0.2
0.1
0
−0.1
0 50 100 150 200 250 300
time [s]

Fig. 17. The comparison of a real torque (blue line) and values given by the model (6) (red line) calculated on the same domain – wind speed sample
presented in Fig. 14. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.)

5. Discussion

The purpose of the described investigations was to create the adequate model of a wind turbine mechanical system input in
order to model the system load. The static behaviour was the main simplification that was introduced regarding the relation
between wind and load. As it was explained at the beginning of Section 4, 15 min current data were analyzed whereas the time
constants between the wind speed and the load were in the range of seconds. Thus, the assumption for simplification can be
justified for this type of data. In such a context the effective reconstruction of variability and irregularity of the input
parameters is more crucial than reconstructing them as a time series. Let us consider the example shown in Fig. 15. At the
neighbourhood of the point t¼250 s the sequence of decrease in speed is correctly reconstructed but it is slightly shifted in
time. This shift is caused by the wide distribution of the experimental values around the model function values, generated by
 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 503

0.6

0.4

torque acceleration
0.2

−0.2

−0.4

−0.6

−0.8
0 50 100 150 200 250 300
time [s]

Fig. 18. The comparison of a real torque increment (blue line) and values given by the model (6) (red line) calculated on the same domain – wind speed
sample presented in Fig. 14. (For interpretation of the references to colour in this figure caption, the reader is referred to the web version of this article.)

Table 2
Results of the main shaft rotational speed approximation by using model (5).

Variables n Dim-model σdim-mod Dim-meas σdim-meas MAPE (%) σ MAPE (%)

vi 111 1.48 0.09 1.54 0.11 6.58 0.05

ai 111 1.48 0.09 1.55 0.10 7.71 0.05
vi 55 1.49 0.06 1.56 0.08 5.15 0.03
ai 55 1.49 0.06 1.59 0.07 7.29 0.07

Table 3
Results of the torque approximation by using model (6).

Variables n Dim-model σmod Dim-meas σdim-meas MAPE (%) σ MAPE (%)

Tðvi Þ 111 1.54 0.07 1.54 0.06 2.28 0.06

ΔTðvi Þ 111 1.56 0.08 1.57 0.07 2.70 0.07
Tðvi Þ 55 1.56 0.05 1.56 0.04 1.78 0.04
ΔTðvi Þ 55 1.58 0.06 1.58 0.05 2.10 0.05

the formulae (5) and (6) – see Figs. 11 and 13. The wide distribution is the effect of, among others, the main shaft inertia which,
on turn, is neglected in the used model due to the omission of the turbine dynamics – see the model assumptions specified at
the beginning of Section 4. The exact behaviour of the turbine depends on the controller which, within the time of about 200 s,
makes the decision to continue idle operation due to high probability of a soon wind gust. The mentioned time shift has no
meaning when the input load is considered but it causes a great value of MAPE of the model. Therefore, due to the fact that the
time series time shift in the context of the load modelling can be neglected, the error of the fitted time series value is not a
good measure of the model accuracy. Therefore, not only the input parameters, i.e. the rotational speed v and the torque T, but
also their increments a≔Δv and ΔT were considered. For the same reasons, the fractal box counting dimension was used as the
measure of the model accuracy because it reflected the input irregularity including chaotic aspect. The applied models allow us
to reflect, very accurately, the box counting dimension of not only the input parameters a and T but also their increments – see
Tables 2 and 3. In particular the torque and its increment, that are the most crucial for the load of mechanical parts, were
modelled with extremely good accuracy – see Figs. 17 and 18 and Table 3.

6. Conclusion

The proposed turbines input load model together with models of turbine gears, see [33,38], will allow us to perform
simulations that are planned to be foundations for developing the methods of fault detection in monitoring and diagnostic
system that have been studied – see [64–67,29]. Up to now, dynamic models of wind turbine drivetrains have been neither
sufficiently stable nor accurate [33,35,38,68]. Nevertheless, the problem is studied intensively. Therefore, the proposed
model of the load can be a helpful contribution for such studies.

References

[1] Global Wind Energy Council, “Global Wind Energy Outlook 2012”, November 2012.
[2] S. Fan, J.R. Liao, R. Yokoyama, R. Chen, W.J. Lee, Forecasting the wind generation using a two-stage network based on meteorological information, IEEE
Trans. Energy Convers. 24 (2009) 474–482.
504 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505

[3] F. Castellani, A. Garinei, L. Terzi, D. Astolfi, M. Moretti, A. Lombardi, A new data mining approach for power performance verification of an on-shore
wind farm, Diagnostyka 14 (4) (2013) 35–42.
[4] S.O. Faried, R. Billinton, S. Aboreshaid, Probabilistic evaluation of transient stability of a wind farm, IEEE Trans. Energy Convers. 24 (2009) 733–739.
[5] R.D. Fernndez, R.J. Mantz, P.E. Battaiotto, Impact of wind farms on a power system. An eigenvalue analysis approach, Renew. Energy 32 (2007)
1676–1688.
[6] M. Jafarian, A.M. Ranjbar, The impact of wind farms with doubly fed induction generators on power system electromechanical oscillations, Renew.
Energy 50 (2013) 780–785.
[7] M. Lei, L. Shiyan, J. Chuanwen, L. Hongling, Z. Yan, A review on the forecasting of wind speed and generated power, Renew. Sustain. Energy Rev. 13
(2009) 915–920.
[8] J. Wen, Y. Zheng, F. Donghan, A review on reliability assessment for wind power, Renew. Sustain. Energy Rev. 13 (2009) 2485–2494.
[9] D.S. Callaway, Sequential reliability forecasting for wind energy: temperature dependence and probability distributions, IEEE Trans. Energy Convers.
25 (2010) 577–585.
[10] A. Kusiak, H. Zheng, Models for monitoring wind farms power, Renew. Energy 34 (2009) 583–590.
[11] A. Kusiak, Z. Zhang, Short-horizon prediction of wind farm power: a data driven approach, IEEE Trans. Energy Convers. 25 (2010) 1112–1122.
[12] M.C. Mabel, E. Fernandez, Estimation of energy yield from wind farms using artificial neural networks, IEEE Trans. Energy Convers. 24 (2009) 459–464.
[13] T. Barszcz, M. Bielecka, A. Bielecki, M. Wójcik, Wind speed modelling using Weierstrass function fitted by a genetic algorithm, J. Wind Eng. Ind.
Aerodyn. 109 (2012) 68–78.
[14] M. Bielecka, T. Barszcz, A. Bielecki, M. Wójcik, Fractal modelling of various wind characteristics for applications in a cybernetic model of a wind
turbine, in: Lecture Notes in Artificial Intelligence, vol. 7268, 2012, pp. 531–538.
[15] E. Hau, Wind Turbines: Fundamentals, Technologies, Application, Economics, Springer, Berlin, Heidelberg, 2006.
[16] A. Jabłoński, T. Barszcz, M. Bielecka, Automatic validation of vibration signals in wind farm distributed monitoring systems, Measurement 44 (2011)
1954–1967.
[17] A. Jabłoński, T. Barszcz, M. Bielecka, P. Brehaus, Modeling of probability distribution functions for automatic threshold calculation in condition
monitoring systems, Measurement 46 (2013) 727–738.
[18] Z. Hameeda, Y.S. Honga, T.M. Choa, S.H. Ahnb, C.K. Son, Condition monitoring and fault detection of wind turbines and related algorithms: a review,
Renew. Sustain. Energy Rev. 13 (2009) 1–39.
[19] Z. Zhang, A. Verma, A. Kusiak, Fault analysis and condition monitoring of the wind turbine gearbox, IEEE Trans. Energy Convers. 27 (2012) 526–535.
[20] K. Dziedziuch, W.J. Staszewski, T. Uhl, Time–frequency analysis of time-variant systems, Diagnostyka 14 (1) (2013) 37–42.
[21] G. Żywica, B. Bagiński, Ł. Breńkacz, Dynamic state assessment of the water turbine with the power of 600 kW, Diagnostyka 14 (1) (2013) 65–70.
[22] T. Barszcz, M. Strąkiewicz, Novel intuitive hierarchical structure for condition monitoring system of wind turbines, Diagnostyka 14 (3) (2013) 53–60.
[23] A. Kusiak, W. Li, The prediction and diagnosis of wind turbine faults, Renew. Energy 36 (2011) 16–23.
[24] A. Verma, A. Kusiak, Analyzing bearing faults in wind turbines: a data-mining approach, Renew. Energy 48 (2012) 110–116.
[25] T. Barszcz, A. Bielecki, M. Wójcik, Vibration signals processing by cellular automata for wind turbines intelligent monitoring, Diagnostyka 14 (2) (2013) 31–36.
[26] T. Barszcz, A. Bielecki, T. Romaniuk, Application of probabilistic neural networks for detection of mechanical faults in electric motors, Prz. Elektrotech.
8 (209) (2009) 37–41.
[27] T. Barszcz, Jabłoński, A novel method for the optimal band selection for vibration signal demodulation and comparison with the Kurtogram, Mech.
Syst. Signal Process. 25 (2011) 431–451.
[28] P. Borghesani, P. Pennacchi, R.B. Randall, N. Sawalhi, R. Ricci, Application of cepstrum pre-whitening for the diagnosis of bearing faults under variable
speed conditions, Mech. Syst. Signal Process. 36 (2013) 370–384.
[29] Z. Feng, M.J. Zou, Fault diagnosis of planetary gearboxes via torsional vibration signal analysis, Mech. Syst. Signal Process. 36 (2013) 401–421.
[30] J. Hu, L. Zhang, W. Liang, Dynamic degradation observer for bearing fault by MTSSOM system, Mech. Syst. Signal Process. 36 (2013) 385–400.
[31] R. Zimroz, R. Urbanek, T. Barszcz, W. Bartelmus, F. Millioz, N. Martin, Measurement of instantaneous shaft speed by advanced vibration signal
processing – application to wind turbine gearbox, Metrol. Measur. Syst. 18 (2011) 701–712.
[32] J. Dybała, Vibrodiagnostics of gearboxes using NBV-based classifier: a pattern recognition approach, Mech. Syst. Signal Process. 38 (2013) 5–22.
[33] W. Bartelmus, Mathematical modelling and computer simulations as an aid to gearbox diagnostics, Mech. Syst. Signal Process. 15 (2001) 855–871.
[34] J. Bednarz, T. Barszcz, T. Uhl, Rotating machinery diagnostics based on NARX models, Comput. Assist. Mech. Eng. Sci. 14 (2007) 557–567.
[35] F. Chaari, R. Hbaeib, T. Fakhfakh, M. Haddar, Dynamic response simulation of planetary gears by the iterative spectral method, Int. J. Simul. Model. 4
(2005) 35–45.
[36] L. Miller, Gearbox Dynamics, WNT, Warsaw, 1986 (in Polish).
[37] R. Hein, C. Orlikowski, Hybrid model of geared rotor system, Diagnostyka 14 (2) (2013) 17–22.
[38] W. Bartelmus, F. Chaari, R. Zimroz, M. Haddar, Modelling of gearbox dynamics under time-varying nonstationary load for distributed fault detection
and diagnosis, Eur. J. Mech. A: Solids 29 (2010) 637–646.
[39] R.G. Parker, S.M. Vijayakar, T. Imajo, Non-linear dynamic response of a spur gear pair: modelling and experimental comparisons, J. Sound Vib. 237
(2000) 355–435.
[40] Y. Zhan, V. Makis, A.K.S. Jardine, Adaptive state detection of gearboxes under varying load conditions based on parametric modeling, Mech. Syst. Signal
Process. 20 (2006) 188–221.
[41] F. Chaari, W. Bartelmus, R. Zimroz, T. Fakhfakh, M. Haddar, Gearbox vibration signal amplitude and frequency modulation, Shock Vib. 19 (2012)
635–652.
[42] T. Figlus, A. Wilk, Application of analysis of envelope's spectrum for gearbox diagnosing, Int. Rev. Mech. Eng. 6 (2012) 1350–1355.
[43] M.T. Khabou, N. Bouchaala, F. Chaari, T. Fakhfakh, M. Haddar, Study of a spur gear dynamic behavior in transient regime, Mech. Syst. Signal Process. 25
(2011) 3089–3101.
[44] M.T. Khabou, T. Hentati, M.S. Abbes, F. Chaari, M. Haddar, Nonlinear modeling and simulation of spur gear with defected bearings, Multidiscip. Model.
Mater. Struct. 8 (2012) 197–212.
[45] L.F. Villa, A. Renones, J.R. Pern, L.J. De Miguel, Statistical fault diagnosis based on vibration analysis for gear test-bench under non-stationary conditions
of speed and load, Mech. Syst. Signal Process. 29 (2012) 436–446.
[46] G. D'Amico, F. Petroni, F. Prattico, First and second order semi-Markov chains for wind speed modeling, Physica A: Stat. Mech. Appl. 392 (2013) 1194–1201.
[47] R. Engelking, Dimension Theory, North Holland Mathematical Library, vol.19, North Holland Publ. Co, Amsterdam, 1978.
[48] M. Tabaszewski, Optimization of a nearest neighbors classifier for diagnosis of condition of rolling bearings, Diagnostyka 15 (1) (2014) 37–42.
[49] M. Barnsley, Fractals Everywhere, Academic Press Inc., Boston, 1988.
[50] K. Falconer, The Geometry of Fractal Sets, Cambridge University Press, Cambridge, 1985.
[51] K. Falconer, Fractal Geometry. Mathematical Foundations and Applications, Wiley & Sons, Chichester, 1990.
[52] J. Bystryk, P.E. Sullivan, Small wind turbine power control in intermittent wind gusts, J. Wind Eng. Ind. Aerodyn. 99 (2011) 624–637.
[53] W. Bartelmus, R. Zimroz, Vibration condition monitoring of planetary gearbox under varying external load, Mech. Syst. Signal Process. 23 (2009)
246–257.
[54] W. Bartelmus, R. Zimroz, A new feature for monitoring the condition of gearboxes in non-stationary operation conditions, Mech. Syst. Signal Process.
23 (2009) 1528–1534.
[55] R.I. Stephens, A. Fatemi, R.R. Stephens, H.O. Fuchs, Metal Fatigue in Engineering, Wiley and Sons, New York, 2001.
[56] A.N. Kolmogorov, On the representation of continuous function of many variables by superposition of continuous functions of one variable, Dokl.
Akad. Nauk SSSR 114 (5) (1957) 953–956.
 A. Bielecki et al. / Mechanical Systems and Signal Processing 54-55 (2015) 491–505 505

[57] G. Cybenko, Approximation by superposition of a sigmoidal function, Math. Control Signals Syst. 2 (1989) 303–314.
[58] K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural Netw. 4 (1991) 251–258.
[59] K. Hornik, Some new results on neural network approximation, Neural Netw. 6 (1993) 1069–1072.
[60] K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural Netw. 2 (1989) 359–366.
[61] V. Kurkova, Kolmogorov's theorem and multilayer neural networks, Neural Netw. 5 (1992) 501–506.
[62] A. Bielecki, J. Ombach, Shadowing property in analysis of neural network dynamics, J. Comput. Appl. Math. 164–165 (2004) 107–115.
[63] A. Bielecki, J. Ombach, Dynamical properties of a perceptron learning process – structural stability under numerics and shadowing, J. Nonlinear Sci. 21
(2011) 579–593.
[64] T. Barszcz, M. Bielecka, A. Bielecki, M. Wójcik, Wind turbines states classification by a fuzzy-ART neural network with a stereographic projection as a
signal normalization, in: Lecture Notes in Computer Science, vol. 6594, 2011, pp. 225–234.
[65] T. Barszcz, A. Bielecki, M. Wójcik, ART-type artificial neural networks applications for classification of operational states in wind turbines, in: Lecture
Notes in Artificial Intelligence, vol. 6114, 2010, pp. 11–18.
[66] T. Barszcz, A. Bielecki, M. Wójcik, M. Bielecka, ART-2 artificial neural networks applications for classification of vibration signals and operational states
of wind turbines for intelligent monitoring, in: Advances in Condition Monitoring of Machinery in Non-Stationary Operations, Lecture Notes in
Mechanical Engineering, 2014, pp 679–688.
[67] T. Barszcz, R.B. Randall, Application of spectral kurtosis for detection of a tooth crack in the planetary gear of a wind turbine, Mech. Syst. Signal Process.
23 (2009) 1352–1365.
[68] W. Bartelmus, F. Chaari, R. Zimroz, T. Barszcz, M. Haddar, Modelling of transmission used in wind turbines, in: Proceedings of the Ninth International
Conference on Condition Monitoring and Machinery Failure Prevention Technologies, 2012.