Kalman filter based strain estimation for fatigue

assessment of an offshore monopile wind turbine

K. Maes1 , G. De Roeck1 , A. Iliopoulos2 , W. Weijtjens3 , C. Devriendt3 , G. Lombaert1

1 KU Leuven, Department of Civil Engineering,

Kasteelpark Arenberg 40, B-3001 Leuven, Belgium

e-mail: kristof.maes@kuleuven.be
2Vrije Universiteit Brussel, Mechanics of Materials and Constructions,
Pleinlaan 2, B-1050 Brussels, Belgium
3Vrije Universiteit Brussel, Acoustics and Vibration Research Group,
Pleinlaan 2, B-1050 Brussels, Belgium

Abstract
This paper presents an application of a Kalman filtering algorithm for strain estimation in the tower of
an offshore monopile wind turbine in the Belgian North Sea. Real measured data obtained from in situ
measurements are used in the validation study. The system model used in the strain estimation is constructed
based on the mode shapes and natural frequencies obtained from a finite element model of the structure and
the damping characteristics that are obtained from a prior output-only operational modal analysis (OMA)
of the tower. A recently developed approach for quantification of the estimation uncertainty introduced by
sensor noise and unknown excitation acting on the structure is applied. The approach is used to determine
the noise statistics for Kalman filtering that minimize the estimation uncertainty. Finally, the quality of the
strain estimates is assessed by comparison with the actual, measured strains. Very accurate strain estimates
are obtained.

1 Introduction

Offshore wind turbines (OWTs) are exposed to continuous wind and wave excitation. This cyclic loading
may lead to high periodic strains at critical locations and structural failure in the end. The continuous
monitoring of the strain response time histories at these locations is important to assess the remaining lifetime
of the structure. The strain response time histories can be directly obtained from sensor data at points
accessible to measurements. For some critical locations which are not accessible for direct measurements,
however, this is not possible. For example, direct measurements of the strains at the mud-line, 30 meter
below the water level, require sensors installed prior to the pile-driving of the monopile foundation. As a
consequence, the sensors cannot be installed on existing OWTs. Experience in the field has also shown that
strain sensors are harder to maintain, are less reliable than accelerometers over long periods in time, and are
more susceptible to installation errors.
When direct measurements of the strains are impossible, response estimation techniques can be used to es-
timate the response at unmeasured locations from a limited set of response measurements (accelerations,
strains, etc.) and a system model. The estimation of unknown quantities from direct measurements on the
structure at other locations is also known as virtual sensing. Various approaches for the estimation of stresses
and strains using response estimation techniques are presented in the literature. Hjelm et al. [1] presented a
strain estimation technique based on natural input modal analysis. This technique was further explored in [2].

1649
1650 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

Other approaches make use of time varying auto-regressive models [3] and model-based Kalman state esti-
mation. The latter approach was introduced in the field of structural dynamics by Papadimitriou et al. in [4],
where acceleration measurements are used as input to the model-based Kalman filter to obtain the strain at
unmeasured locations. The Kalman filter based response estimation was numerically and experimentally
investigated in [5], for the special case of non-zero mean forces applied to the structure. The paper also
explores the advantages of data fusion, i.e. the simultaneous integration of multiple types of measurements,
for example acceleration and strain measurements. It was numerically verified by Jo and Spencer [6] that the
combination of acceleration and strain data in conjunction with the Kalman filter results in better estimates
compared to the ones obtained from the sole use of acceleration or strain data.
This paper presents a full scale verification of a Kalman filtering algorithm, using data obtained from an in
situ experiment on an offshore monopile wind turbine in the Belgian North Sea [7, 8]. The algorithm makes
use of a model of the structure and a limited number of response measurements for the prediction of the
strain in the tower of the turbine during operation. A finite element model of the structure is used in the
strain prediction. The damping characteristics of the structure are obtained from a prior operational modal
analysis (OMA) of the tower. The response used in the strain estimation procedure consists of acceleration
and strain measurements. The noise statistics assumed for Kalman filtering are determined using a recently
developed method for uncertainty quantification. A validation is performed by comparison of the estimated
strains with the actual, measured strains in the tower. The focus in this application is on dynamic strain
estimation.
The outline of the paper is as follows. Section 2 briefly recapitulates the Kalman filter algorithm and its
application for response estimation, as well as the method for uncertainty quantification that is used in the
determination of the noise statistics for Kalman filtering. Next, in section 3, the proposed response estimation
technique is validated using data obtained from a monitoring campaign on a monopile offshore wind turbine.
Finally, in section 4, the work is concluded.

2 Mathematical formulation

2.1 System model

Consider a linear modally reduced order model of a structure, constructed from a limited number of modes.
When proportional damping is assumed, the continuous-time decoupled equations of motion in modal coor-
dinates are given by:
z̈(t) + Γż(t) + Ω2 z(t) = ΦT Sp (t)p(t) (1)
where z(t) ∈ Rnm is the vector of modal coordinates, with nm the number of modes taken into account in
the model. The excitation force is written as the product of a selection matrix Sp (t) ∈ Rndof ×np , specifying
the force locations, and a time history vector p(t) ∈ Rnp , with np the number of forces. The number of
degrees of freedom is indicated by ndof . Γ ∈ Rnm ×nm is a diagonal matrix containing the terms 2ξj ωj on
its diagonal, where ωj and ξj are the natural frequency and modal damping ratio corresponding to mode
j, respectively. Ω ∈ Rnm ×nm is a diagonal matrix as well, containing the natural frequencies ωj on its
diagonal, and Φ ∈ Rndof ×nm is a matrix containing the mass-normalized mode shapes φj as columns.
The output vector is generally written as:

d(t) = Sd,a Φz̈(t) + Sd,v Φż(t) + Sd,d Φz(t) (2)

where Sd,a , Sd,v , and Sd,d ∈ Rnd ×ndof are selection matrices indicating the degrees of freedom correspond-
ing to the acceleration, velocity, and displacement (or strain) measurements, respectively.
Equations (1) and (2) can be written into state-space form. After time discretization and adding noise, the
I NVERSE METHODS - LOAD IDENTIFICATION 1651

following discrete-time combined deterministic-stochastic state-space description of the system is obtained:

x[k+1] = Ax[k] + Bp[k] + w[k] (3)

d[k] = Gx[k] + Jp[k] + v[k] (4)

where x[k] = x(k∆t) and p[k] = p(k∆t), and d[k] = d(k∆t) (k = 0, . . . , N − 1), ∆t is the sampling
time step, and N is the total number of samples. The state vector x[k] contains the modal displacements and
velocities. The process noise w[k] compensates for modeling errors and ambient excitation, i.e. excitation
which is not accounted for by the excitation force in equation (1). The measurement noise v[k] compensates
for modeling errors, ambient excitation, and measurement errors (see also [9]). The expressions for the
system matrices A, B, G, and J for modally reduced order models and full order models can be found
in [10]. Alternatively, the system matrices can be directly identified from experimental vibration data using
system identification techniques, see e.g. [11, 12].
Finally, consider a vector de (t) ∈ Rnde of nde outputs that are extrapolated from the measured data and the
system model by virtual sensing:

de (t) = Sde ,a Φz̈(t) + Sde ,v Φż(t) + Sde ,d Φz(t) (5)

where the matrices Sde ,a , Sde ,v , and Sde ,d ∈ Rnde ×ndof relate the identified accelerations, velocities and
displacements or strains, respectively, to the degrees of freedom in the model (see also equation (2)). In
the particular case of strain estimation as considered in this paper, the selection matrices Sde ,a and Sde ,v
equal zero. After transformation of equation (5) into its state-space form, using equation (1), and adding
measurement noise, the following (discrete-time) output equation corresponding to the extrapolated output
quantities is obtained:
de[k] = Ge x[k] + Je p[k] + ve[k] (6)
The matrices Ge ∈ Rnde ×ns and Je ∈ Rnde ×np relate to the extrapolated output quantities and therefore
are different from the original matrices G and J in equation (4), that correspond to the measured output
quantities. Note that in case of strain estimation, the direct feedthrough matrix Je equals zero. The measure-
ment noise ve[k] accounts for modeling errors and unmodeled excitation that is not accounted for in the force
vector p[k] .

2.2 Kalman filter algorithm

This section gives an overview of the Kalman filter algorithm [13] and its application for strain estimation in
structural dynamics.
In the derivation of the Kalman filter, the unknown system input is included in the noise processes w[k] and
v[k] . Under this assumption, the system described by equations (3) and (4) becomes:

x[k+1] = Ax[k] + w[k] (7)

d[k] = Gx[k] + v[k] (8)

The noise processes w[k] ∈ Rns and v[k] ∈ Rnd , which now account for all excitation sources, are assumed
to be zero mean and white, with known covariance matrices Q ∈ Rns ×ns , R ∈ Rnd ×nd , and S ∈ Rns ×nd :
    Q S 
w[k]  T T
E w[l] v[l] = T δ (9)
v[k] S R [k−l]
 
Q S
with R > 0, T ≥ 0, and δ[k] = 1 for k = 0 and 0 otherwise. E{·} is the expectation operator.
S R
Finally, it is assumed that an unbiased estimate x̂[0|−1] of the initial state is available, with error covari-
ance matrix Px[0|−1] (i.e. E x[0] − x̂[0|−1] = 0, Px[0|−1] = E (x[0] − x̂[0|−1] )(x[0] − x̂[0|−1] )T ). The
 

estimate x̂[0|−1] is assumed independent on the noise processes w[k] and v[k] for all k.
1652 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

The Kalman filter consists of a two-step recursive filter [14]:

x̂[k|k] = x̂[k|k−1] + K[k] d[k] − Gx̂[k|k−1] (10)
x̂[k+1|k] = Ax̂[k|k] (11)

The first step in equation (10), referred to as the “measurement update”, yields a filtered estimate of the state
vector x[k] , given the measured output d[k] up to time step k. The second step in equation (11), referred to
as the “time update”, yields a one step ahead prediction of the state vector x[k+1] . The expression for the
Kalman gain K[k] ∈ Rns ×nd is found in [14]. In the equations above, the system is assumed to be time-
invariant. The algorithm can, however, also be applied to time-variant systems, resulting in system matrices
A[k] and G[k] depending on the time step k.
When the data vector includes accelerations (J 6= 0), the process noise w[k] and measurement noise v[k] are
inherently correlated [12, 15] (i.e. S 6= 0). Although this correlation is mostly disregarded for the application
of the Kalman filter, the Kalman filter algorithm used in this paper assumes correlated noise processes (see
also [12, 16]).
After applying the Kalman filter algorithm, the estimated state vector x̂[k|k] can be used to estimate the output
de[k] , using the following modified output equation:

d̂e[k|k] = Ge x̂[k|k] (12)

where d̂e[k|k] is the estimated response. For acceleration estimation, an error is introduced, since the direct
feedthrough of the excitation as contained in the noise vectors w[k] and v[k] is disregarded in equation (12).

2.3 Uncertainty quantification

This section summarizes the method for quantification of estimation uncertainty introduced by unknown
stochastic excitation and measurement noise, that has been recently presented in [9]. The method allows to
estimate the power spectral density (PSD) of the error on the estimated strains, denoted as Sd̃e d̃e (ω).
The uncertainty quantification approach was implemented in [9] for a joint input-state estimation algorithm
and is here extended for the Kalman fitler algorithm. It is found that the expression for the strain error
PSD Sd̃e d̃e (ω) for the Kalman filter algorithm is identical to the expression obtained for the joint input-state
estimation algorithm in [9], which is given by:

Sd̃e d̃e (ω) = (Hde pS (ω) − Hd̂e d (ω)HdpS (ω))SpS pS (ω)(H∗de p (ω) − H∗dp (ω)H∗d̂ d (ω)) . . .
S S e

+ Hd̂e d (ω)SvM vM (ω)H∗d̂ d (ω) (13)

e

where Hde pS (ω) and HdpS (ω) are the transfer function matrices of the system that relate the Fourier trans-
form of the actual strains de[k] and the measured response d[k] , respectively, to the Fourier transform of the
stochastic forces pS[k] acting on the structure. Hd̂e d (ω) is the steady-state transfer function matrix of the
system inversion algorithm, that relates the Fourier transform of the estimated strains d̂e[k] to the Fourier
transform of the measured response d[k] [9]. As compared to the implementation for the joint input-state
estimation algorithm, the (steady-state) transfer functions Hd̂e d (ω) in this case should be replaced by those
of the Kalman filter algorithm, which are given by:

Hd̂e d (ω) = exp(iω∆t)Ge A−1 (exp(iω∆t)Ins − A + ALss G)−1 ALss (14)

with Ins ∈ Rns ×ns an identity matrix. The steady state gain matrix Lss is calculated by recursively applying
Kalman filter equations, or, alternatively, by solving the Riccati differential equation [17]. The matrices
SpS pS (ω) and SvM vM (ω) in equation (13) are the PSD functions of the unknown stochastic force and the
measurement noise, respectively, that are assumed known. The reader is referred to [9] for an in depth study
of the uncertainty quantification approach.
I NVERSE METHODS - LOAD IDENTIFICATION 1653

3 Monopile wind turbine

In this section, the Kalman filter algorithm is applied for a data set obtained from a monitoring campaign
on a Vestas V90 3 MW wind turbine on a monopile foundation (Fig. 1) [8]. The wind turbine is located at
the Belwind offshore windfarm in the North Sea, 46 km off the Belgian coast. The hub-height of the wind
turbine is on average located at 72 m above the lowest astronomical tide (LAT). The wind turbine is mounted
on a monopile foundation with a diameter of 5 m. The water depth at the turbine location is 24 m w.r.t. LAT
and the monopile has a penetration depth of 21 m.

(a) (b)
Figure 1: (a) Vestas V90 3 MW wind turbine on a monopile foundation and (b) mud-line, hotspot for fatigue.

3.1 Measurement setup

The data considered in this paper has a duration of 10 minutes and corresponds to rotating conditions of the
turbine. The measured wind speed equals ±11 m/s, the rotation speed of the turbine ±16 rpm.
In total, 10 accelerometers and 6 strain gauges have been installed on the wind turbine; 4 accelerometers
at a height of 69 m (LAT), 2 accelerometers and 2 optical fiber bragg grating strain sensors at a height of
41 m (LAT), 2 accelerometers at a height of 27 m (LAT), and 2 accelerometers and 4 optical fiber bragg
grating strain sensors at a height of 19 m (LAT). Fig. 2 shows the sensor configuration. The data at all levels
are transformed in respectively the Fore-Aft (FA) direction, parallel to the wind direction, and the Side-Side
(SS) direction, perpendicular to the wind direction, using the known yaw direction. The acceleration and
strain data obtained are hereafter referred to as aαβ and ǫαβ , respectively, where α indicates the level in
m LAT and β indicates the direction (FA or SS).

y a1 a5 a7 a9
ǫ1 ǫ3
z x a3
ǫ2 ǫ6 ǫ4
a2 a6 a8 a10
ǫ5
a4
h = 69 m h = 41 m h = 27 m h = 19 m
Figure 2: Sensor configuration (levels h relative to LAT).

3.2 Data acquisition and processing

A sampling rate of 20 Hz is used in the data acquisition. The frequency content below 0.05 Hz and above
5 Hz is removed. As an example, the processed acceleration and strain data at level h = 41 m LAT (FA
and SS) are shown in Figs. 3 and 4. The sample Power Spectral Density (PSD) function is estimated using
Welch’s method, hereby applying a window length of 1024 samples and an overlap of 66%.
1654 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

Acceleration PSD [(m/s ) /Hz]

SS acceleration [m/s ]
2 2

2
−1
1.5 10−2
Acceleration [m/s ]

0.5
2

1 10−3
10
0.5 −4
10−5 0
0 10
−6
−0.5 10−7
−1 10−8
10 −0.5
−1.5
0 100 200 300 400 500 600 0 1 2 3 4 5 −1 −0.5 0 0.5 1
Frequency [Hz] 2
Time [s] FA acceleration [m/s ]
(a) (b) (c)
Figure 3: (a) Time history and (b) estimated PSD of the FA acceleration measured at level h = 41 m LAT,
and (c) time history of the FA acceleration versus time history of the SS acceleration measured at level
h = 41 m LAT.
Strain PSD [(µm/m)2/Hz]
3
40 102

SS strain [µm/m]
101 20
Strain [µm/m]

20 100
10−1
0 0
10
−2
−20 10−3
10 −20
−4
−40 10
0 100 200 300 400 500 600 0 1 2 3 4 5 −40 −20 0 20 40
Time [s] Frequency [Hz] FA strain [µm/m]
(a) (b) (c)
Figure 4: (a) Time history and (b) estimated PSD of the FA strain measured at level h = 41 m LAT, and (c)
time history of the FA strain versus time history of the SS strain measured at level h = 41 m LAT.

3.3 System model

A system model is constructed from a FE model of the wind turbine, developed by OWI/VUB. The FE model
has been updated using a set of experimental modal parameters, which have been obtained through an output-
only system identification procedure using data collected in parked conditions [18, 19]. The displacement
mode shapes corresponding to the first six modes calculated using the updated FE model are shown in Fig. 5.
Table 1 shows the experimentally identified natural frequencies and modal damping ratios [19], as well as the
natural frequencies obtained from the FE model and the MAC-values [20]. For nearly all bending modes of
the tower, a high MAC-value is obtained, indicating a good agreement between the measured and computed
mode shapes. Note that the FE model is symmetric, whereas the actual behavior of the wind turbine is
different in the FA and SS directions and will vary for different wind directions [21].
A reduced-order discrete-time state-space model is constructed from the FE model of the wind turbine,
applying a zero order hold assumption on the force. The model includes the 6 modes listed in Table 1. For
each mode, the mass normalized mode shape and natural frequency are assumed to be known from the FE
model, whereas the modal damping ratio is taken as the experimentally identified value.

60 60 60 60 60 60

40 40 40 40 40 40
Height [m]

Height [m]

Height [m]

Height [m]

Height [m]

Height [m]

20 20 20 20 20 20

0 0 0 0 0 0

−20 −20 −20 −20 −20 −20

−2 0 2 −2 0 2 −2 0 2 −2 0 2 −2 0 2 −2 0 2
m1 m2 m3 m4 m5 m6

Figure 5: First six displacement mode shapes obtained from the FE model, FA-direction (red) and SS-
direction (green). From left to right: FA1, SS1, SS2, FA2, SS3, FA3.
I NVERSE METHODS - LOAD IDENTIFICATION 1655

No. fid ffem [Hz] ξid [%] MAC Description

[Hz]
1 0.361 0.374 1.86 0.99 First FA bending mode
2 0.365 0.374 2.49 0.99 First SS bending mode
3 1.448 1.440 1.38 0.99 Second SS bending mode
4 1.560 1.440 1.14 0.98 Second FA bending mode
5 3.610 3.636 0.56 0.96 Third SS bending mode
6 3.910 3.636 0.92 0.92 Third FA bending mode

Table 1: Experimentally identified modal characteristics and comparison to the modal characteristics ob-
tained from the FE model (fid : identified natural frequency, ffem : natural frequency obtained from FE
model, ξid : identified modal damping ratio, MAC: MAC-value [20]).

3.4 Quantification of estimation uncertainty

This section applies the uncertainty quantification approach outlined in section 2.3 to determine the optimal
noise statistics Q, R, and S that minimize the uncertainty on the estimated strains obtained from the Kalman
filter, when applied for estimating the FA and SS strain at the lowest measurement level (h = 19 m LAT).
Two sensor configurations for strain estimation are compared in the following, hereafter referred to as S1
and S2 (table 2). Sensor configuration S1 consists of six accelerations (a69FA , a69SS , a41FA , a41SS , a27FA ,
a27SS ). Sensor configuration S2 consists of the same six accelerations and in addition contains two strains at
level h = 41 m LAT (ǫ41FA and ǫ41SS ).

Configuration Sensors
S1 a69FA , a69SS , a41FA , a41SS , a27FA , a27SS
S2 a69FA , a69SS , a41FA , a41SS , a27FA , a27SS , ǫ41FA , ǫ41SS
Table 2: Overview of the sensor configurations assumed for strain estimation.

The noise covariance matrices Q, R, and S are calculated as:

   ′  
Q S B  ′T ′T
 0 0
= Cp B J + (15)
ST R J′ 0 RM
where Cp ∈ RnpS ×npS is the covariance matrix of the unknown stochastic forces and RM ∈ Rnd ×nd is
the measurement error covariance matrix, which both act as tuning parameters/matrices [9]. The system
matrices B′ ∈ R2nm ×npS and J′ ∈ Rnd ×npS are obtained by assuming excitation at the degrees of freedom
in the model corresponding to the stochastic forces. The covariance matrix Cp in this case is initially chosen
as 1012 N2 × I6 . This corresponds to stochastic forces with a standard deviation of 106 N acting in the FA
and SS direction at the three highest levels where sensors have been installed. The matrix RM is assumed
diagonal with values (5 × 10−4 )2 (m/s2 )2 and 0.152 µs2 for the diagonal elements corresponding to the
accelerations and strains, respectively. The assigned values characterize the accuracy of the acceleration
and strain sensors. The initial state vector x̂[0|−1] and the corresponding error covariance matrix P[0|−1] are
assigned a zero value.
The uncertainty estimation requires the PSD function of the stochastic forces acting on the structure SpS pS (ω)
as well as the PSD function of the measurement noise SvM vM (ω) (see equation (13)). As in a recent case
study involving the identification of multiple loads on the bridge deck of a footbridge [22], the PSD function
SpS pS (ω) is here estimated from the cross-PSD of the response under ambient loading Sdd (ω) as follows:

SpS pS (ω) = H†dp (ω)Sdd (ω)H∗†

dp (ω) (16)
S S

where ∗ and † denote the Hermitian transpose and Moore-Penrose pseudo-inverse of a matrix, respec-
tively. The response d[k] assumed in the estimation of the force PSD consists of the FA and SS acceleration
1656 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

at the three highest levels (a69FA , a69SS , a41FA , a41SS , a27FA , a27SS ), and the FA and SS strain at the level
h = 41 m LAT (ǫ41FA and ǫ41SS ). The cross-PSD of the response signals, Sdd (ω), is estimated using the
average periodogram method [23], hereby subdividing the 10 minutes of data in 10 blocks of 1 minute each.
Six equivalent stochastic forces are assumed in the form of concentrated FA and SS point loads at the three
highest levels. Figure 6 shows the estimated PSD of the equivalent forces. Especially for the equivalent
forces at levels h = 41 m LAT and h = 27 m LAT (figures 6b and c, respectively), the force PSD is domi-
nated by low frequency components. This is expected, since the structure is mainly excited by wind loading.
At very low frequencies (< 0.25 Hz), the force PSD is characterized by large errors that are mainly due to
measurement noise on the acceleration data and should therefore not be considered. The estimated forces
show a significant contribution of the rotor harmonics. The rotor harmonic frequencies have been indicated
in figure 6. Finally, note that, as expected, the FA force component for each level is generally larger than the
SS component.
12 12 12
10 10 10
Force PSD [N2/Hz]

Force PSD [N2/Hz]

Force PSD [N2/Hz]

10 10 10
10 10 10
8 8 8
10 10 10
6 6 6
10 10 10
4 4 4
10 10 10
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] Frequency [Hz] Frequency [Hz]
(a) (b) (c)
Figure 6: Estimated PSD of the (equivalent) stochastic FA (black) and SS (gray) forces acting (a) at level
h = 69 m LAT, (b) at level h = 41 m LAT, and (c) at level h = 27 m LAT, for data set 1 (rotating
conditions). The natural frequencies obtained from the FE model are indicated by black dashed lines. The
rotor harmonic frequencies for a rotation speed of 16 rpm are indicated by gray dash-dotted lines.

For the calculation of the measurement error PSD SvM vM (ω), the measurement noise present on the ac-
celeration and strain data is assumed white, with a standard deviation of 5 × 10−4 m/s2 and 0.15 µm/m,
respectively, where the measurement errors on the acceleration and strain signals are assumed uncorrelated.
Under these assumptions, the measurement error PSD SvM vM (ω) is obtained as:

SvM vM (ω) = RM /F (17)

Figure 7 shows the PSD of the error on the estimated strains Sd̃e d̃e (ω) for the two different sensor configu-
rations S1 and S2 . For both configurations, the majority of the estimation uncertainty is introduced by the
unknown stochastic excitation, whereas only a minor part of the uncertainty is due to measurement noise. If
this were not the case, the strain error PSD would mainly show a 1/ω 4 behavior (see also [14]). It is also
observed that including strain data (sensor configuration S2 ) results in a significant overall reduction of the
estimation uncertainty.
Str. err. PSD [( µm/m) /Hz]

Str. err. PSD [( µm/m) /Hz]

2

2 2
10 10
0 0
10 10
−2 −2
10 10
−4 −4
10 10
−6 −6
10 10
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] Frequency [Hz]
(a) (b)
Figure 7: PSD of the error on (a) the estimated FA strain and (b) the estimated SS strain at level h =
19 m LAT, for Cp = 1012 I. The error PSD for sensor configuration S1 is indicated by a black solid line,
for sensor configuration S2 by a gray dashed line. The filter cutoff frequency of 0.25 Hz, used for highpass
filtering, is indicated by a vertical dotted line.
I NVERSE METHODS - LOAD IDENTIFICATION 1657

The covariance of the strain errors d̃e[k|k] is a measure for the overall uncertainty on the estimated strains.
The covariance is immediately obtained from the strain error PSD Sd̃e d̃e (ω) as follows:
Z ωmax 
cov(d̃e[k|k] ) = 2 Re Sd̃e d̃e (ω)dω (18)
ωmin

where ωmin and ωmax are assumed 0.25 × 2π rad/s and 5 × 2π rad/s, respectively. The lower bound
ωmin = 0.25 × 2π rad/s is chosen in agreement with the cutoff frequency used for highpass filtering which
is applied in order to remove the low frequency drift on the estimated strains (see also section 3.6). The
covariance is calculated by numerical integration.
Figure 8 shows the variance of the error on the estimated strains obtained from the Kalman filter for sensor
configuration S2 , as a function of the standard deviation σpS in the calculation of the noise covariance
matrices Q, R, and S. It is observed that for small values of σpS (< 104 N) the error variance depends
significantly on σpS . For larger values of σpS (> 104 N), the error variance becomes low and independent
on the value of σpS assumed in the estimation. As observed from figure 8b, the error covariance obtained for
σpS → ∞ is not necessarily the minimum value. The value σpS = 106 N, that has already been used in the
above, gives rise to low estimation uncertainty. The same value is used for strain estimation in the following.
Strain error var. [( µm/m)2]

Strain error var. [( µm/m)2]

30 30
25 25
20 20
15 15
10 10
5 5
0 0 2 4 6 8 10
0 0 2 4 6 8 10
10 10 10 10 10 10 10 10 10 10 10 10
St. dev. stoch. forces [N] St. dev. stoch. forces [N]
(a) (b)
Figure 8: Variance of the error on (a) the estimated FA strain and (b) the estimated SS strain at level h =
19 m LAT as a function of the standard deviation σpS assumed for the application of the Kalman filter
algorithm, for sensor configuration S2 . The value of σpS corresponding to figure 7 is indicated by a vertical
dotted line.

3.5 Strain estimation results

This section discusses the results of the strain estimation. Figure 9 shows the measured and estimated strain
signals for sensor configuration S1 . A highpass filter with a cutoff frequency of 0.25 Hz has been applied to
the measured and estimated strain signals before plotting the strain time histories. The 0.25 Hz low frequency
limit is chosen to focus on dynamic strains (see also section 3.6). The PSDs have been calculated from the
original unfiltered data. A comparison of the measured and estimated strains shows a good overall estimation
quality. It is observed from the strain PSDs that the measured (unfiltered) strain signals are characterized
by an important low frequency content (frequencies < 0.25 Hz), which is not present in the estimated
strain signals. As such, it is found that the low frequency information cannot be accurately recovered from
acceleration data only (see for example figures 9c and 9f).
It is now verified how the actual estimation errors compare to the strain error PSD as obtained from the
uncertainty quantification approach in section 3.4. Figure 10 compares the PSD of the actual estimation
errors, i.e. the errors on the estimated strains show in figure 9, to the error PSD that was estimated using the
uncertainty quantification approach (see also figure 7). The discrepancy between the actual and estimated
strain error PSD indicates the presence of modeling errors. The uncertainty quantification approach only
accounts for stochastic noise processes, i.e. measurement noise and unknown stochastic excitation, and can
be applied to select the tuning parameters (Q, R, and S) that minimize the uncertainty introduced by those
noise processes (see also section 3.4). When the estimation errors introduced by measurement noise and
unknown stochastic excitation are small with respect to the errors due to model uncertainty, however, the
1658 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

uncertainty quantification will significantly underestimate the actual estimation errors, since modeling errors
are not accounted for.

Strain PSD [(µm/m)2/Hz]

3
30 30 10
2
Strain [µm/m]

Strain [µm/m]
20 20 10
1
10 10 10
0
0 0 10
−1
−10 −10 10
−2
−20 −20 10
−3
−30 −30 10
0 100 200 300 400 500 600 200 220 240 260 0 1 2 3 4 5
(a) Time [s] (b) Time [s] (c) Frequency [Hz]

Strain PSD [(µm/m)2/Hz]

3
20 20 10
2
Strain [µm/m]

Strain [µm/m]
10
10 10 1
10
0
0 0 10
−1
10
−10 −10 −2
10
−3
−20 −20 10
0 100 200 300 400 500 600 200 220 240 260 0 1 2 3 4 5
(d) Time [s] (e) Time [s] (f) Frequency [Hz]

Figure 9: Time history (left), detail time history (middle) and estimated PSD (right) of the FA strain (a–c)
and SS strain (d–e) at level h = 19 m LAT for sensor configuration S1 . The measured strains are shown
in black, the estimated strains in gray.
Str. err. PSD [( µm/m)2/Hz]

Str. err. PSD [( µm/m) /Hz]

3 3
10 10
2

2 2
10 10
1 1
10 10
0 0
10 10
−1 −1
10 10
−2 −2
10 10
−3 −3
10 10
0 1 2 3 4 5 0 1 2 3 4 5
(a) Frequency [Hz] (b) Frequency [Hz]

Figure 10: PSD of the actual (black) and estimated (gray) error on (a) the estimated FA strain and (b) the
estimated SS strain at level h = 19 m LAT, for Cp = 1012 I6 , sensor configuration S1 .

Figure 11 shows the measured and estimated strain signals for sensor configuration S2 . Very accurate strain
estimates are obtained, both for the FA and SS direction. As indicated by Jo and Spencer in [6], the experi-
mental validation shows that better strain estimates are obtained if the Kalman filter is fed with acceleration
and strain data. This is seen from comparison of figures 9 and 11. Even more, it is observed from the PSD of
the estimated strains in figures 11c and 11f that the Kalman filter provides accurate results in the quasi-static
frequency range (frequencies < 0.25 Hz) when strain data are included.

3.6 Low frequency strains

The focus in this application is on dynamic strain estimation. There are two problems when the proposed
methods are applied to predict the low frequency strains, induced by the thrust-loading. The first concern is
the conversion from accelerations to strains. The integration inevitably requires a double integration, which
risks to blow up the low frequency noise in the measured accelerations, resulting in large errors in the low
frequency components of the predicted strains. This is observed in section 3.5 for sensor configuration S1 ,
where it was found that accurate low frequency strains could not be retrieved from acceleration data only.
The near-static strains below 0.25 Hz obtained from the estimation are predominantly due to measurement
noise, which have therefore been removed by filtering. Note that the use of accelerometers with better noise
properties can push this lower frequency bound further down.
The second problem is more fundamental to the concept. All techniques applied in this section rely upon
modal decomposition. A low number of modes is included, that contribute to the dynamic response in the
I NVERSE METHODS - LOAD IDENTIFICATION 1659

Strain PSD [(µm/m)2/Hz]

3
30 30 10
2

Strain [µm/m]
Strain [µm/m]
20 20 10
1
10 10 10
0
0 0 10
−1
−10 −10 10
−2
−20 −20 10
−3
−30 −30 10
0 100 200 300 400 500 600 200 220 240 260 0 1 2 3 4 5
(a) Time [s] (b) Time [s] (c) Frequency [Hz]

Strain PSD [(µm/m)2/Hz]

3
20 20 10
2
Strain [µm/m]

Strain [µm/m]
10
10 10 1
10
0
0 0 10
−1
10
−10 −10 −2
10
−3
−20 −20 10
0 100 200 300 400 500 600 200 220 240 260 0 1 2 3 4 5
(d) Time [s] (e) Time [s] (f) Frequency [Hz]

Figure 11: Time history (left), detail time history (middle) and estimated PSD (right) of the FA strain (a–c)
and SS strain (d–e) at level h = 19 m LAT for sensor configuration S2 . The measured strains are shown
in black, the estimated strains in gray.

frequency range from 0.25 to 5 Hz. At low frequencies, a much higher number of modes is required to
describe the quasi-static behavior of the turbine. The resulting extrapolation can yield erroneous results,
as the static deflection patterns of the turbine might not be accurately described by superposition of a low
number of dynamic modes.
In this section, the lower frequency bound was set to 0.25 Hz to avoid both problems, yet try to capture
as many dynamics as possible, including wave induced vibrations. For the Belgian coast, wave periods
exceeding 5 s are fairly common (wave frequency ≤0.2 Hz), however, necessitating a lower value of the
frequency bound. This can be achieved through further improvement of the current methods or the use of
sensors with better noise properties.

4 Conclusions

A Kalman filtering algorithm is applied for strain estimation in the tower of an offshore Vestas V90 3 MW
wind turbine on a monopile foundation, using measured vibration data in rotating conditions. Focus in this
paper is on the application of a recently developed algorithm for uncertainty quantification to determine the
noise statistics in the Kalman filtering. It is found that the uncertainty quantification approach allows to
select the noise statistics that minimize the uncertainty introduced by sensor noise and unknown stochastic
excitation. When this uncertainty is small compared to the uncertainty introduced by modeling errors, the
estimated uncertainty underestimates the actual estimation uncertainty, however. The Kalman filter algorithm
is evaluated using acceleration data only, as well as a combination of acceleration and strain data. Accurate
results are obtained when using only acceleration data. Adding strains allows for a further improvement of
the estimation accuracy.

Acknowledgements

The research presented in this paper has been performed within the framework of the project G.0738.11
Inverse identification of wind loads on structures, funded by the Research Foundation Flanders (FWO),
Belgium. The financial support of FWO is gratefully acknowledged. The authors affiliated to KU Leuven
are all members of the KU Leuven - BOF PFV/10/002 OPTEC - Optimization in Engineering Center. The
authors would also like to acknowledge the Offshore Wind Infrastructure Application Lab, Parkwind and
Belwind for providing the data that has been used throughout this paper.
1660 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016

References

[1] H.P. Hjelm, R. Brincker, J. Graugaard-Jensen, K. Munch, Determination of stress histories in struc-
tures by natural input modal analysis, In Proceedings of IMAC XXIII International Modal Analysis
Conference, Houston, Texas, USA, 2006, Houston (2006), pp. 838–844.

[2] M. López Aenlle, L. Hermanns, P. Fernández, A. Fraile, Stress estimation in a scale model of a
symmetric two story building, In Á. Cunha, L.F. Ramos, E. Caetano, P. Lourenço, editors, Proceedings
of the 5th International Operational Modal Analysis Conference, IOMAC 2013, Guimarães, Portugal,
May 2013.

[3] E. Yazid, S.L. Mohd, P. Setyamartan, Time-varying spectrum estimation of offshore structure response
based on a time-varying autoregressive model, Journal of Applied Sciences, Vol. 12, No. 23 (2012),
pp. 2383–2389.

[4] C. Papadimitriou, C.-P. Fritzen, P. Kraemer, E. Ntotsios, Fatigue predictions in entire body of metallic
structures from a limited number of vibration sensors using Kalman filtering, Structural Control and
Health Monitoring, Vol. 18 (2011), pp. 554–573.

[5] R.P. Palanisamy, S. Cho, H. Kim, S-H Sim, Experimental validation of Kalman filter-based strain
estimation in structures subjected to non-zero mean input, Smart Structures and Systems, Vol. 15,
No. 2 (2015), pp. 489–503.

[6] H. Jo, B.F. Spencer, Multi-Metric model-based structural health monitoring, In J. P. Lynch, K-W
Wang, and H. Sohn, editors, Proceedings of SPIE 9061, Sensors and Smart Structures Technologies for
Civil, Mechanical, and Aerospace Systems 2014, San Diego, California, USA, March 2014.

[7] R. Shirzadeh, C. Devriendt, M. Ahmadi Bidakhvidi, P. Guillaume, Experimental and computational

damping estimation of an offshore wind turbine on a monopile foundation, Journal of Wind Engineering
and Industrial Aerodynamics, Vol. 120 (2013), pp. 96–106.

[8] C. Devriendt, F. Magalhães, W. Weijtjens, G. De Sitter, Á Cunha, P. Guillaume, Structural health

monitoring of offshore wind turbines using automated operational modal analysis, In F.-K. Chang,
editor, Proceedings of the 9th International Workshop on Structural Health Monitoring, IWSHM 2013,
Stanford, CA, USA, September 2013, Stanford (2013), Vol. 1, pp. 2415–2422

[9] K. Maes, A.W. Smyth, G. De Roeck, G. Lombaert, Joint input-state estimation in structural dynamics,
Mechanical Systems and Signal Processing, Vol. 70–71 (2016), pp. 445–466.

[10] E. Lourens, C. Papadimitriou, S. Gillijns, E. Reynders, G. De Roeck, G. Lombaert, Joint input-response

estimation for structural systems based on reduced-order models and vibration data from a limited
number of sensors, Mechanical Systems and Signal Processing, Vol. 29 (2012), pp. 310–327.

[11] P. Van Overschee, B. De Moor, Subspace algorithm for the stochastic identification problem, Auto-
matica, Vol. 29, No. 3 (1993), pp. 649–660.

[12] E. Reynders, G. De Roeck, Reference-based combined deterministic-stochastic subspace identification

for experimental and operational modal analysis, Mechanical Systems and Signal Processing, Vol. 22,
No. 3 (2008), pp. 617–637.

[13] R. Kalman, A new approach to linear filtering and prediction problems, Journal of Basic Engineering,
Transactions of the ASME, Vol. 82D (1960), pp. 35–45.

[14] K. Maes, A. Iliopoulos, W. Weijtjens, C. Devriendt, G. Lombaert, Dynamic strain estimation for fa-
tigue assessment of an offshore monopile wind turbine using filtering and modal expansion algorithms,
Mechanical Systems and Signal Processing, Vol. 76–77 (2016), pp. 592–611.
I NVERSE METHODS - LOAD IDENTIFICATION 1661

[15] Y. Niu, M. Klinkov, C.-P. Fritzen, Online force reconstruction using an unknown-input Kalman filter
approach, In Proceedings of the 8th International Conference on Structural Dynamics, EURODYN
2011, Leuven, Belgium, July 2011, Leuven (2011), pp. 2569–2576.

[16] P. Van Overschee, B. De Moor, Subspace identification for linear systems. Kluwer Academic Publish-
ers, Dordrecht, The Netherlands (1996).

[17] G.F. Franklin, J.D. Powell, M. Workman, Digital control of dynamic systems. Addison-Wesley, Menlo
Park, CA, USA (1998).

[18] A. Iliopoulos, R. Shirzadeh, W. Weijtjens, P. Guillaume, D Van Hemelrijck, C. Devriendt, A modal

decomposition and expansion approach for prediction of dynamic responses on a monopile offshore
wind turbine using a limited number of vibration sensors, Mechanical Systems and Signal Processing,
Vol. 68–69 (2016), pp. 84–104.

[19] C. Devriendt, W. Weijtjens, M. El Kafafy, G. De Sitter, Monitoring resonant frequencies and damping
values of an offshore wind turbine in parked conditions, IET Renewable Power Generation, Vol. 8,
No. 4 (2014), pp. 433–441.

[20] R.J. Allemang, D.L. Brown, A correlation coefficient for modal vector analysis, In Proceedings of the
1st International Modal Analysis Conference, Orlando, FL, USA, 1982, Orlando (1982), pp. 110–116.

[21] W. Weijtjens, T. Verbelen, G. De Sitter, C. Devriendt, Foundation structural health monitoring of an

offshore wind turbine – a full-scale case study, Structural Health Monitoring (2015).

[22] K. Maes, K. Van Nimmen, E. Lourens, A. Rezayat, P. Guillaume, G. De Roeck, G. Lombaert, Veri-
fication of joint input-state estimation for force identification by means of in situ measurements on a
footbridge, Mechanical Systems and Signal Processing, Vol. 75 (2016), pp. 245–260.

[23] E. Reynders, System identification methods for (operational) modal analysis: review and comparison,
Archives of Computational Methods in Engineering, Vol. 19, No. 1 (2012), pp. 51–124.
1662 P ROCEEDINGS OF ISMA2016 INCLUDING USD2016