Sie sind auf Seite 1von 10

An Approach for Fatigue Load Monitoring without

Load Measurement Devices


Nicolai Cosack, Martin Kühn
Endowed Chair of Wind Energy, Universität Stuttgart, Germany
+49 (0)711 685 68273, cosack@ifb.uni-stuttgart.de

Abstract: loads, the adoption of maintenance cycles


according to the load history and the realisation
of a basic condition monitoring by means of
A novel approach for the monitoring of wind
trend analysis of the load data. Furthermore,
turbine fatigue loads has been developed. The
the load information could also be used to
method works without the installation of
compare wind turbines in a series not only on
additional measurement devices. Instead, an
the basis of their energy production but also on
empirical relationship between standard wind
their load history. Results of such a comparison
turbine control signals and wind turbine loading
might be useful to improve the planning of
is established. Once this procedure has been
future wind farms.
established for one turbine, it should be suitable
for the estimation of fatigue loads for all
Therefore, the present research project
turbines of that type. This paper gives an
investigates methods for using easily available
overview of the basic approach as well as
turbine signals, like rotor speed, electrical
results of an estimation of single stage spectra
power, pitch angle, etc., for fatigue load
and single stage distributions for some
monitoring. The work aims to use these
important load components of a state of the art
standard wind turbine signals for the estimation
wind turbine.
of simple load indicators, such as damage
equivalent loads, but also for the more detailed
1 Introduction fatigue spectra, e.g. multi stage load duration
distributions and rainflow spectra.
Routinely, load measurements at wind turbines
are carried out only for a few prototypes as
temporary measurement campaigns. The 2 System Specification
development of cost effective and reliable
A complete fatigue load estimation system does
systems which can be used for continuous load
not only consist of an estimation procedure, but
monitoring of wind turbines has been proven to
requires also a process control and a
be a difficult task. Besides the high costs of
monitoring scheme. The general layout of such
measurement equipment and the large effort of
a system is shown in Figure 2 and described
post-processing and evaluating the collected
briefly in this section.
data, the expected lifetime of a turbine might
exceed the lifetime of standard measurement
Process Control
devices (e.g. strain gauges) by far and
replacement might not always be easy.
System Identification
Accurate estimation results can only be
Besides information about energy yield,
achieved, if the properties of the individual wind
operating hours and maintenance or repair
turbine are in good agreement with the design
activities, a wind turbine’s performance and
assumptions of the estimation system. Due to
especially its dynamic behaviour is known only
site specific conditions, such as soil properties,
from the prototype testing. This prototype
as well as to fabrication and assembly
testing does not necessarily reflect the real
tolerances, the properties of wind turbines in a
conditions in the series at particular sites. As a
series can differ to some extend. Hence an
consequence, very little or, in most cases,
exact match of the design assumptions is
nothing is known about the load history of more
virtually impossible and system identification is
than a few of the many existing turbines.
required to check that the turbine’s properties
lie within a suitable range. It should be
Measuring loads in a continuous manner offers
performed at least once prior to the initialisation
some significant benefits for the operation of a
of the estimation system and periodically if
wind turbine and can help to produce energy at
significant system degradation over time can be
lower costs. Among the most obvious
assumed.
advantages are the adjustment of turbine
operational control with respect to the current
Figure 2: Layout of a load estimation system

The detection of immoderate rotor imbalances continuously for plausibility. Single outliers of
by system identification can serve as a practical input parameters as well as sensor
example. If large imbalances are present but do malfunctions have to be detected in order to
not influence the standard signals used for the avoid corruption of the estimation results.
load prediction, their influence on the estimated Secondly, some estimation procedures can only
loads will not be captured. On the other hand, if handle data within a specific range, especially
a standard signal is significantly influenced by when the estimation is based on empirically
an undetected disturbance, the prediction error derived relations which are not capable of
is likely to be increased. extrapolation of data beyond the calibration
range.
Turbine status If corrupted or unusable data has been found a
In order to give a complete picture, load suitable method of correction and handling of
predictions have to be performed for all turbine these data has to be defined to make sure that
states relevant for the fatigue loading. The no periods of time are neglected for the fatigue
relation between standard signals and loads load estimation.
can be different for the various states such as
normal power production, idling, starts and Transfer function
stops. It can therefore be reasonable to adopt The transfer function is the key part of the
the estimation procedure according to the estimation. Some possible methods like
current turbine state by treating the input data regression models and neural networks are
differently or switching to other estimation described and compared in the next section.
procedures.
Monitoring
Load estimation procedure
Load Monitoring
Statistical analysis The individual short-term fatigue load
Because standard signals are mainly used for estimations would usually be cumulated to give
control of wind turbines they are usually the overall long-term fatigue load. Cumulated
measured with high resolution. These loads are then simply compared against the
continuous high resolution time series can design loads of the turbine and allow for an
either be applied directly for the estimation or assessment of the past operation of the turbine.
divided into sections of equal length, usually 10 With some additional assumptions regarding
minutes, from which suitable statistical the long-term environmental conditions at a site
parameters can then be derived. it is also possible to extrapolate the fatigue
The choice of the input data format depends loads to the design life of the wind turbine.
also on the desired outcome of the estimation However, this method requires choosing
system. If the outcome is not supposed to be a representative loads as basis for the
high resolution time series statistical extrapolation and is therefore more complicated
parameters are probably sufficient in most than the simple load accumulation described
cases as input. above.
If done correctly the life-time extrapolation
Data Verification allows for a better assessment of the turbine
This step serves two main purposes. In the first operation compared to the load accumulation
place, the input data has to be checked especially at the beginning of the monitoring
period. It can be expected that the results from the weighted summation of inputs and a pre-
both methods approach each other after some defined function, the so called activation
time of monitoring, e.g. two to three years. function, which is used to derive the output of
the neuron. Many neurons can be combined in
Condition Monitoring parallel or in line to build a neural network. The
Without much additional effort the above load output of each activation function thereby
estimation system can be extended to serves as input to the following neurons. If the
incorporate a basic monitoring that aims for the network is sufficiently large it is possible to
prediction of the overall turbine condition. approximate very complex relations by a
Besides information from a periodically combination of in fact relatively simple
performed system identification also the result activation functions.
of the data verification can be analysed to One of the advantages of a neural network
safeguard the compliance of the wind turbine approach is that these networks can be trained
with the specification and hence the design to map the relation of input and output data if a
assumptions of the wind turbine. Furthermore, suitable amount of training data is available.
the predicted loads can be analysed with Using one of the many available training
respect to trends to detect wear and tear of the algorithms, the set-up of a neural network is
turbine as well as long-term degradation. often faster and easier than the set-up of a
regression model, especially if complex non-
linear relations have to be mapped.
3 Load Estimation Procedures Furthermore, assumptions regarding the nature
of the relation are required only in that respect
Four different load estimation procedures have
that the network has to be sufficiently large to
been analysed and compared to asses their
be able to approximate the relation with the
suitability for the proposed prediction scheme.
available number of neurons. An a priori
These are classical regression methods, neural
selection of suitable input parameters can
networks, physical model based approaches
improve the training process but is not
and state estimators.
compulsory. Input data which can not be related
to the output will simply be assigned a relatively
Classical regression techniques are often used
low weight during the training process.
to establish relations between two sets of data,
Just as with regression techniques, application
in case that only limited information on the
of the neural network for predictions is only
underlying processes is available [1].
reliable if the input data are within the range
Regression usually requires making some
which has been used for the training of the
assumptions regarding the nature of the
network.
connection between input and target values, for
example such as being linear. Furthermore, in a
preceding step suitable input parameters have Detailed physical models, for example FE-or
to be defined which will be related to the multi body models of structures, can also be
targets. Based on these a priori choices the used to relate input data to load quantities using
input and target values are correlated to each observer based approaches as known from
other and the parameters of the assumed control theory. These models require very
relation are derived. This relation is also specific input data, like accelerations or
referred to as the regression function and can deflections at specific points of the structure.
be interpreted as transfer function which maps Furthermore, detailed knowledge of structural
the input to the target values. properties of the physical system that governs
If the relation between input and target values is the relation has to be available. Alternatively,
very non-linear or depends on several input classical grey-box system identification
parameters the regression approach gets approaches can be applied to set-up suitable
significantly more complex. It is likely that many models [4]. Extrapolation, which is not possible
correlations have to be tested and combined to with regression or neural network techniques, is
a working regression model in this case. The not an issue with this approach. If the model
process can be partly automated by means of reflects the real system behaviour correctly it
more elaborate stepwise regression techniques can be applied at all turbine states.
[2] but still it can not be guaranteed that the
found solution performs well. Another serious State estimation can be applied to classify wind
disadvantage of regression techniques is that turbine operation in terms of fatigue loads. For
they can generally not be used for each identified state a corresponding load is
extrapolation. read from an associated database and
Neural networks have been proven to be a very cumulated to an overall fatigue load. The state
useful alternative to classic regressions [3]. As recognition approach requires an a priori
shown in Figure 3 a single neuron consists of classification of turbine operation into suitable
Figure 3: Single neuron and neural network

sub-states and the definition of representative example the estimation of high resolution rotor
loads for each sub-state. Once this has been thrust time series with regression models [2] or
achieved, the method is able to give results the estimation of multi stage rainflow spectra
even for turbine states that are difficult to with neural networks [5, 6].
assess with other methods for example
transient situations like starts or stops. Here we focus on the prediction of single stage
spectra based on standard signals and neural
Table 1 summarises the comparison of the networks (Figure 4). The prediction of single
above described estimation methods. In most stage spectra has already been discussed
cases, analysis of standard signals will probably briefly in [5], but only with respect to the rotor
not give such detailed input to employ physical torque. In this work we extend this approach
models for the load prediction. Furthermore, also to other main load components.
information for setting up these models might Because standard signals are usually
not be fully available. measured in a fixed frame of reference behind
Wind turbines operate under constantly the rotor they can be regarded as integral
changing wind conditions and the turbines quantities and include cumulated aerodynamic
themselves are very non-linear systems. and structural dynamic effects from all three
Therefore, an excessive number of possible blades. Here, special focus will be placed on
states and the respective representative loads the suitability of these integral quantities for the
would have to be defined, for state estimation estimation of non-integral loads, such as blade
methods. root bending moments.
Because the presented work aims for an
approach with a broad applicability and which Two main tasks have been performed for the
requires as little information as possible on the set-up of the estimation procedure:
wind turbine itself, regression techniques and
neural networks both seem to be suitable. It is 1. Calculation of statistical parameters and
very likely that the transfer function which single stage loads
relates input and target values is highly non- Because the spectra are estimated for fixed
linear and depends on several input time periods of ten minutes we also use
parameters. This makes the use of classical only statistical parameters of the standard
regression techniques difficult and therefore a signals as input to the estimation
neural network based approach has been procedure. Suitable statistical parameters
chosen for this work. The estimation scheme are derived from standard signals. They
including input and target values is presented in serve then as input to the neural network
more detail in the next section. for the prediction of single stage fatigue
load spectra.

4 Description of the Applied 2. Training of the neural network


Estimation Scheme Prior to the estimation the network has to
be trained to map the relations between the
input data and the single stage spectra.
4.1 Overview Common feed-forward networks with only
Fatigue loads can be described in various one hidden layer have been used in this
formats such as time series, multi stage fatigue work. Training of the network was
spectra or single stage spectra, which can all performed on basis of standard training
serve as targets for an estimation procedure. algorithms such as backpropagation with
The estimation of some of these formats has moment term or resilient backpropagation
been investigated in previous publications, for [3].
Figure 4: Scheme for the estimation of single stage spectra from standard signals using neural
networks

Estimation
Data Requirements Assumptions Coverage Miscellaneous
Method
- Input and target - A priori choice of - Extrapolation not - Standard
values have to be the most relevant possible; only technique for
available input parameters valid for the linear problems
required design ranges - Effort increases
- Assumptions significantly for
Regression
regarding the non-linear
Techniques
nature of the multivariate
regression problems
function - Regression
function available
as equation
- Many data sets of - Network - Extrapolation not - Suitable for non-
input and target architecture and possible; only linear multivariate
values for training training valid for the problems
required algorithms have design ranges - Relevant input
to be chosen parameters are
found
automatically
Neural networks during the training
process
- Relation of input
and target values
‘hidden’ in
network, i.e. not
directly
accessible
- Detailed - Model and input - Coverage of all - Most suitable for
information for the data are operational states specific
set up and sufficiently component
validation of a complex and models, e.g.
Physical models
physical model accurate to allow support structures
necessary for predictions or actuators
- Suitable input
data for the model
- Input and target - Suitable sub- - Depends on the - Accuracy of the
values are states of the defined states results increases
required operational only with the number
- Information on the conditions can be of defined states
State estimation turbine load defined - Capable of load
with database characteristics - Representative estimations
and operational loads for each during transients
behaviour sub-state can be
necessary derived

Table 1: Comparison of estimation methods


Figure 5: Variation of skewness and kurtosis

Suitable statistical parameters and the Frequency domain related parameters


derivation of single stage fatigue loads are Besides knowing the variability of a signal, the
discussed in the following sections. frequency content of the oscillations is also
valuable information and can improve the
4.2 Relevant Statistical Parameters estimation results. Similar to the central
moments in the time domain, the spectral
Most variable-speed pitch-controlled wind moments are basic parameters to characterise
turbines monitor at least electrical power output, the distribution of the signal’s oscillations in the
generator rotational speed, pitch angles and frequency domain.
tower top accelerations. Furthermore wind To derive the spectral moments, time series are
direction deviations and yaw activities are usually transferred to the frequency domain to
recorded. Secondary signals like rotational yield the power spectra density (PSD) of a
accelerations or pitch rates can be derived from signal as shown in Figure 6. Often traditional
these primary data if necessary. Fast Fourier Transform based spectral
estimation methods, for example the
In a pre-processing step adequate statistical periodogram or the correlation method, are
parameters have to be calculated. These applied for this task.
parameters have to be suitable to describe the Other approaches utilise autoregression
turbine activities during the past time period. If models, where the autoregression coefficients
the chosen parameters do not reflect the are derived from the autocorrelation function of
turbine behaviour well or if relevant information a signal. The coefficients can then be related to
has been omitted the accuracy of the estimation the power spectra density. These models are
result is likely to be decreased. Often a good computationally more expensive but can give
description of the corresponding turbine better results than the traditional approaches,
operation is only possible if time and frequency especially if relatively short sequences of a
domain related parameters are included. signal are to be analysed. Furthermore they do
not rely on the assumption, that the signal is
Time domain related parameters composed of harmonically related sinusoids.
The mean values of standard signals can [7]. If the PSD has been derived, the spectral
indicate the average operation point of a turbine moments can be calculated with cyclic
while standard deviations or variances are frequency f and the order n of the spectral
basic measures of the variability of the moment from
operation around this point. ∞
A more detailed description of the variability is λn = ∫ f n ⋅ PSD( f ) df
possible if additional higher order statistical 0
parameters like the skewness and the kurtosis
It should be mentioned that there are also
are taken into account. These parameters are
methods to derive spectral moments directly
calculated from the third and fourth central
from time domain representations of a process
moments of a signal and can be used to
[8, 9]. However, from the above formula it is
describe basic deviations of the signal
obvious, that the contribution of high
distribution from the normal distribution [1]. As
frequencies to the spectral moment increases
illustrated in Figure 5, the skewness indicates
with increasing order of the moment. A
the deviation of the distribution peak from the
comparison of moments of different order can
mean of the distribution, while the kurtosis is
therefore be used to conclude on the
related to the relative flatness of the distribution.
Figure 6: Normalised generator rotational speed signal as time series and power spectral

characteristics of the PSD and hence on the to calculations with small values of m.
distribution of energy of the process. Therefore a comparison of single stage spectra
which have been derived from different
Time series which feature significant exponents can indicate the ratio of large to
oscillations with periods that are relatively large small load ranges in the underlying multi stage
compared to the length of the time series are spectra.
often denoted as being trended. Trends in wind
speeds make the comparison of measured wind Besides the rainflow spectra, multi stage load
turbine data more complicated, because they duration distributions are often used to
can have for example a large influence on the characterise the loading of a wind turbine.
variance of some signals while others are left These distributions are a measure of how long
almost unaffected. Therefore trended data sets certain load levels are present. Similar to the
are often excluded when measured data is above definition a single stage distribution can
analysed. be calculated from the time Ti, the
In case of a load estimation system, excluding corresponding load level Li and a reference
data sets is not an option. Instead the time Tref. The exponent p can be used again to
introduction of trend related parameters is indicate the ratio of large and small load levels
preferred. Besides utilising spectral moments, it in the distribution.
is also possible to compare for example the

∑ (L )
total variance of the signal to the variance of the
same but low-pass filtered signal.
p
i ⋅ Ti
Leq ,Tref , p , k = p i

4.3 Output Data Tref


Multi stage load spectra are determined from
The target values of the load estimation
time series by appropriate counting procedures.
procedure are single stage spectra and
An example is the rainflow counting procedure,
distributions on the basis of ten minute time
which is used to count load cycles in a time
periods. They can be assembled to a single
series. The rainflow multi stage spectrum can
value if the same reference numbers and
be further simplified to a single stage spectrum,
exponents have been used.
which under certain assumptions will give the

( )
same fatigue damage. It can be calculated from
∆S eq , N ref , m = m ∑ ∆S eq , N ref , m , k
m
the number of cycles Ni associated with the
load range Si, a reference load cycle number k
Nref and an exponent m. The exponent m is

∑ (L )
usually chosen as inverse slope of the S-N
Leq ,Tref , p =
p
curve and thereby reflects material properties p eq ,Tref , p , k
with respect to fatigue. k

∑ (∆S i
m
⋅ Ni ) 5 Test and Validation
∆S eq, N ref , m, k = m i
N ref 5.1 General
From the above formula it is also obvious, that if
The above described approach for the
the single stage spectrum is calculated with a
estimation of single stage spectra and
large exponent, large load ranges will have a
distributions from statistical parameters of
relatively higher impact on the result compared
standard signals has been developed and
tested on the basis of simulated wind turbine • the overall error calculated from the
data first. In a second step, the method is logarithmic sum of all estimations
validated with measured data from real wind ∆Seq,Nref,m
turbines. Only results on the basis of measured • the linear average of all estimation
data are given in this section. errors ∆Seq,Nref,m,k
• the standard deviation of all estimation
Since 2005 the Endowed Chair of Wind Energy errors as measure of the average error
is involved in a long-term measurement of a single estimation
campaign that is still going on at the prototype
of the Multibrid M5000 turbine in Bremerhaven. Exact values are listed in Table 2, while Figure
The M5000 is a state of the art multi-megawatt 7 illustrates the data further.
turbine with a rated power output of 5 MW, a
rotor diameter of 116 m and a hub height of The mean error is relatively small for all
102m. estimations and decreases even further if the
For the validation of the presented approach, individual estimations are cumulated to an
about 150h of measured data have been overall long-term spectrum. In this case the
extracted reflecting typical site conditions with maximum deviation is about 4% only in the
wind speeds between 4 and 22m/s and case of the out-of-plane blade root bending
turbulence intensities ranging from 3 to 17%. moment and an exponent of m=10. These
Only measured time series where the turbine accurate results seem at least to some extend
was in normal power production mode are to be due to the averaging of prediction results.
considered. The standard deviation of errors indicates that
larger differences between predicted and real
Time and frequency domain related statistical loads can be expected for a single estimation.
parameters of electrical power, generator Especially for large exponents an error of 10 to
rotational speed, generator rotational 13% is likely.
acceleration, pitch angle and tower top
accelerations have been calculated. Single The most accurate predictions have been
stage loads of rotor torque, rotor thrust and derived for the in-plane and edgewise blade
blade root bending moments have been root bending moments respectively. These
analysed using two exponents (4 and 10) for quantities are dominated by deterministic
each load quantity. The blade root bending oscillations which are due to the self weight of
moments are estimated as in-plane and out-of- the blade and the rotation of the rotor. As the
plane moments in a fixed coordinate system as weight is constant and the rotational speed is a
well as edge- and flapwise moments in a direct input to the procedure a very good result
pitching coordinate system. can be expected. This holds true even for the
standard deviations of the estimation error
For each load quantity and each exponent a which are also higher for all other loads. The
single neural network has been trained. About somewhat lower standard deviation for spectra
120h of data have been used to train the calculated on the basis of the exponent of 4
networks. The trained neural networks have indicates that these quantities can be predicted
then been utilised to predict the single stage better compared to the spectra based on larger
spectra on basis of the derived statistical exponents.
parameters. This has been done for the
remaining 30h of data, which have not been In general, the estimations of out-of-plane and
included in the training process. flapwise bending moments comprise the largest
errors. This is probably due to the fact, that all
The estimation results of single stage spectra input signals are measured in a fixed frame of
∆Seq,Nref,m are given in the next section for two reference behind the rotor. They can therefore
exponents, followed by the results for the single be regarded as integral quantities and include
stage distributions ∆Leq,Tref,p. In both cases, the cumulated aerodynamic and structural dynamic
estimations have been compared with results effects from all three blades. Utilisation of these
from standard time series counting procedures signals for the prediction of aerodynamically
and the errors have been computed. driven blade root loads results hence in
relatively large scatter of the estimation results.
5.2 Estimation of Single Stage
Spectra
For each load the following quantities have
been computed, in order to assess the
accuracy of the estimation results:
Figure 7: Comparison of single stage spectra estimation errors

Figure 8: Comparison of single stage distribution estimation errors

5.3 Estimation of Single Stage


Distributions 6 Conclusion and Outlook
A method for the estimation of single stage
The same estimation errors as described above
fatigue loads from standard wind turbine signals
have also been derived for the single stage
has been developed. As the exact relationships
distributions. Results are given in Table 3 and
between standard wind turbine signals and load
Figure 8.
quantities are in general unknown, an empirical
approach has been applied. Here, statistical
In general, the single stage distributions can be
parameters of the standard signals are related
predicted with higher accuracy compared to the
to the main load components by trained neural
spectra. The estimation errors for the
networks. The focus is placed on the estimation
cumulated distributions are between -1.3 and
of the most important load components, such
0.5% in any case.
as blade root bending moment, rotor thrust,
Unlike to the spectra, the largest standard
rotor torque and tower bending moments, which
deviations of the estimation errors occur for
are characteristic for the overall turbine load
integral load quantities, i.e. the rotor torque and
level.
the rotor thrust. The most likely reason for this
Measured data from a state-of-the-art multi
is the relatively large variation in the 10 minute
megawatt wind turbine has been used for the
mean value of these loads compared to the
validation of the approach. Results show
variations in the other quantities.
relatively small estimation errors only,
especially if they are cumulated to an overall
long-term load.
Mean error from Standard deviation of
Error from ∆Seq,Nref,m
∆Seq,Nref,m,k error from ∆Seq,Nref,m,k
Load quantity
% % %
m=4 m = 10 m=4 m = 10 m=4 m = 10
Rotor torque 0.5 -1.2 1.1 2.3 6.7 13.2
Rotor thrust 1.8 0.6 3.7 2.8 7.9 9.9
In-plane bending moment 0.0 1.1 0.1 1.3 1.6 1.8
Out-of-plane bending moment 2.7 4.0 5.4 6.4 10.9 11.7
Edgewise bending moment 0.7 0.0 0.8 0.3 1.7 1.6
Flapwise bending moment 2.0 2.1 3.8 8.0 11.2 12.2
Table 2: Single stage spectra estimation errors

Mean error from Standard deviation of


Error from Leq,Tref,p
Leq,Tref,p,k error from Leq,Tref,p,k
Load quantity
% % %
p=4 p = 10 p=4 p = 10 p=4 p = 10
Rotor torque -0.1 -0.3 0.8 -0.1 7.5 2.6
Rotor thrust -0.4 -1.3 1.4 0.2 5.7 4.8
In-plane bending moment 0.4 0.5 0.4 0.6 1.4 1.8
Out-of-plane bending moment 0.0 0.4 -0.1 0.5 2.1 2.1
Edgewise bending moment -0.1 0.2 0.0 0.5 1.6 1.9
Flapwise bending moment 0.0 0.6 -0.1 1.0 2.3 2.1
Table 3: Single stage distribution estimation errors

Further validation of the method with data from Analyse von Standardsignalen; Akida 2006;
onshore wind turbines is planned. Emphasis will Aachen, 2006
be placed on the robustness of the method and [3] Gurney, K.; An Introduction to Neural
also on transferability of the established Networks; UCL Press; London; 1997
estimation procedure to other turbines of the [4] Marrant, B., van Holten, T.; System
same type. Application and improvement of the Identification for the Analysis of Aeroelastic
approach with data from the German offshore Stability of Wind Turbine Blades; EWEC;
test site “alpha ventus” will begin as soon as London; 2004
measured data is available. [5] Cosack, N., Kühn, M.; Prognose von
Ermüdungslasten an Windenergieanlagen
mittels Standardsignalen und neuronaler
Acknowledgements Netze; Dresdner Maschinenelemente
Kolloquium; Dresden; 2007
The research is part of the Project ”Verification
[6] Cosack, N., Kühn, M.; Estimating fatigue
of Offshore Wind Turbines” (OWEA) at the
spectra from standard wind turbine signals;
German offshore test site ”alpha ventus” and is
DEWEK; Bremen; 2008
supported by the Federal Ministry for
[7] Kay, S.M., Marple, S.L.; Spectrum Analysis
Environment, Nature Conservation and Nuclear
A Modern Perspective; Proceedings of the
Safety (BMU). The authors further acknowledge
IEEE; Vol. 69; No. 11; November 1981
the cooperation with the Multibrid GmbH,
[8] Petrucci, G., Zuccarello B.; Fatigue life
Bremerhaven, and the possibility to validate the
prediction under wide band random loading;
presented approach on the basis of measured
Fatigue and Fracture of Engineering
data.
Materials and Structures; Vol. 27; 2004
[9] Holm, S.; Spectral Moment Matching in the
References Maximum Entropy Spectral Analysis
Method; IEEE Transaction on Information
[1] Fahrmeier, L., Künstler, R., Pigeot, I., Tutz, Theory; Vol. IT-29; No. 2; March 1983
G; Statistik; 5. Auflage; Springer; 2004
[2] Cosack, N., Kühn, M.; Überwachung von
Belastungen an Windenergieanlagen durch

Das könnte Ihnen auch gefallen