Applied Ocean Research 54 (2016) 73–86

Contents lists available at ScienceDirect

Applied Ocean Research

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

Estimation of wind sea and swell using shipboard measurements – A

refined parametric modelling approach
Najmeh Montazeri a,∗ , Ulrik Dam Nielsen a,b , Jørgen Juncher Jensen a,b
a
Department of Mechanical Engineering, Technical University of Denmark, DK-2800 Kgs Lyngby, Denmark
b
Centre for Autonomous Marine Operations and Systems (AMOS), NTNU, Norway

a r t i c l e i n f o a b s t r a c t

Article history: Shipboard wave estimation has been of interest in recent years for the purpose of decision support. In
Received 15 March 2015 this paper, estimation of sea state is studied using ship responses and a parametric description of direc-
Received in revised form 6 November 2015 tional wave spectra. A set of parameters, characterising a given wave spectrum is estimated through an
Accepted 7 November 2015
optimisation problem using global search basin with proper constraints. The cost function is established
Available online 7 December 2015
based on the difference between the energies of a set of measured ship responses and the correspond-
ing theoretical spectral moments. A partitioning procedure is applied, which is able to separate swell
Keywords:
components from wind seas.
Sea state estimation
Shipboard measurements © 2015 Elsevier Ltd. All rights reserved.
Parametric method
Wind sea
Swell

1. Introduction method is simple instrumentation and hardware, which contains

The effect of waves in rough weather is one of the factors that
may degrade a ship’s operational efficiency. In order to navigate the
ship without undue structural loading, motions, local accelerations
and green water on deck, sensing the waves is important for the
ship master. Estimation of the local directional wave spectra in real
time would be very useful in connection with operational guidance
to predict responses such as the longitudinal hull girder loading
and the sea keeping performance of the ship. Moreover, continuous
monitoring of the wave environment during voyages would also
provide valuable input to long-term ship operational profiles.
Floating wave buoys are one primary tool for wave
measurements, but they are not practical for ship operations
that may require precise sea state information in real time and at
the ship’s location. There are basically two tools for onboard wave
estimation: Wave radar systems and estimation based on ship
responses. Wave information can also be supported by satellite
observations, but this method needs further development to be
applicable for decision support.
The response based approach is focused herein, where the ship
itself is considered as a wave buoy and response measurements are
processed to achieve the sea state. This method has been explored
in many research projects e.g. [1,2]. The main advantage of this
∗ Corresponding author. Tel.: +45 71590871.
E-mail address: najmon@mek.dtu.dk (N. Montazeri).
only an ordinary set of accelerometers and possibly strain gauges
connected to a data logging system.
This paper proposes a simplified parametric approach for
estimation of wave parameters. The difference between this
method and the traditional parametric method, e.g. [3,4], is that
the spectral fitting is not applied frequency by frequency. Instead,
the spectral moments that represent the energy of responses are
employed to form the optimisation problem. As the actual wave
spectrum is sometimes composed of different systems including
wind sea and swell, it is helpful to classify the wave components to
make judgement on the numerical model and to analyse the wave
conditions. This is useful for navigational guidance. Therefore, a
partitioning approach is implemented in the spectral estimation.
The method is applied on simulated responses of an in-service
container ship.
2. Formulation
2.1. Response spectra
The theoretical relationship between cross-spectral density of
the ith and the jth responses of a ship, ij (ω), and the directional
wave spectrum, S(ω, ), is given by:
 
ij (ω) = Hi (ω, )Hj∗ (ω, )S(ω, )d, i, j = 1, b, . . ., N (1)
−

0141-1187/$ – see front matter © 2015 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.apor.2015.11.004
74 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86
where ω is wave frequency,  is relative wave direction and H(ω,
) denotes the complex-valued transfer function, * is the conjugate
notation and N is the number of responses. The left- and right-
hand sides of Eq. (1) are the measured and the calculated response
spectral densities, respectively. The responses are measured with
respect to the moving reference frame of the ship and hence, they
should be considered in encounter frequency domain. The relation
between the encounter frequency and the wave frequency is as
follows:
V 2
ωe = ω − ω cos()
g
where V is the ship speed and g is the gravity acceleration. Many
studies have been performed so far for wave estimation using Eq.
(1), where the errors between the two sides of the equation are
minimised. These studies can be basically categorised as parametric
and non-parametric methods. Both methods have been successful
to some extent (see e.g. [2,3]).
2.2. Parametric modelling
For decision support applications, it is advantageous to describe
the exciting waves by standard models. Because while the ship
is moving, the shape of the wave spectrum is exposed to vari-
ations from one location to another. The interpretation and
evolution of these variations would be easier by spectral param-
eters. The inaccuracy of the spectral shape can be overcome by
increasing the number of parameters. By doing so, the model is
sufficiently flexible to describe various sea states in different loca-
tions.
Additionally, regarding the data storage limit onboard ships,
it is more efficient to archive datasets of the integral parameters
rather than the whole spectra. For this reason, in some of the
non-parametric approaches, a parametric form is eventually fit-
ted to the estimated wave spectrum. This makes the parameters
readily available and improves the fit towards a smooth spec-
tral shape, [5,6]. Another advantage in the parametric approach
is that a full wave spectrum within a wide range of frequen-
cies will be estimated. In other words, the high frequency tail
of the spectrum can be approximately formed such that the
whole spectrum is adjusted to a standard spectral shape, even
though high frequency waves may have been filtered out by the
ship.
The most problematic issue in the parametric method is that
numerical conditioning of the nonlinear minimisation problem is
poor and it is important to start with good initial guesses for the
parameters. Moreover, numerical problems might occur since the
optimisation deals with response magnitudes close to zero in some
combinations of frequencies and directions.
The most common model for one directional wave system, S(ω),
is JONSWAP spectrum that is characterised by the significant wave
height, Hs , the peak frequency of waves, ωp , and the peakedness fac-
tor, . Although this model was initially designed for the North Sea,
it is used almost anywhere, because the bandwidth can be adjusted
by changing the peakedness factor. Since it is important in decision
support that the wave model is general enough to cover various
areas, this spectrum is applied in the present study.
For decision making about speed and heading, it is important
to have the complete directional wave spectrum available as the
short-crestedness of waves can have a significant effect on ship
motions and structural loads. Thus, directional wave spectrum is
used as a product of a frequency-wise spectrum and a cos-2s type
spreading function, which is a function of the mean wave direc-
tion, , and the maximum spreading parameter, smax . The latter
two parameters are also considered as the fit parameters in the
model.
2.3. Energy based cost functions
The area under a response spectrum (Eq. (1)) in a small fre-
quency interval, ıω, represents the response energy contribution
within that band of frequencies. The transformation from wave
frequency to encounter frequency (Eq. (2)) does not change the
amount of energy of the response; i.e. the area under the spectral
density:
¯ ij (ωe )ıωe = ij (ω)ıω
 (3)
(2)
where  ¯ denotes the response spectrum in the encounter fre-
quency domain. By integrating the two sides of Eq. (1) with respect
to corresponding frequencies, another system of cost functions is
derived, which represents the energy balance or the 0th order spec-
tral moment (m0 ) of the measured and the estimated responses:
 ωe  ωh  
h
¯ ij (ωe )dωe =
 Hi (ω, )Hj∗ (ω, )S(ω, ) ddω,
ωe ωl −
l
i, j = 1, 2, . . ., N (4)
This equation has been used in [7] as an additional cost function
to Eq. (1). However, in the present study, Eq. (4) is considered
exclusively. Using only this energy expression makes the opti-
misation simpler since the number of equations in the system is
decreased so that the optimisation is faster. Furthermore, for one
single response where i = j, the left-hand side (the measured part)
of Eq. (4) can be simply calculated using the variance of the corre-
sponding signal. Consequently, the uncertainties due to smoothing
and interpolation within spectral analysis and also probable large
errors at periods close to the transfer function zero can be avoided.
The frequency limits in Eq. (4) should be tuned properly so that
the lower and the upper limits of encounter frequencies on the
left-hand side, ωel and ωeh respectively, precisely represent the cor-
responding wave frequency band between ωl and ωh on the right
hand side. This is not straightforward since the mean direction of
waves is unknown. This issue will be discussed in Section 5.2.
In addition to Eq. (4), higher order spectral moments of indi-
vidual responses with respect to encounter frequencies are used:
 ωe  ωh    n
h
¯ ii (ωe )dωe V 2
ωen  = ω− ω cos()
ωe ωl −
g
l
 2
× Hi (ω, ) S(ω, ) ddω, n = 1, 2, . . . (5)
3. Spectral partitioning
3.1. Modelling of bimodal wave spectra
In the oceans, swell events may add to the locally generated
wind sea, creating a double (or multiple) peaked spectrum. The dif-
ferent wave systems in a spectrum have different peak frequencies
and usually different directions of propagation. In such a condition,
in order to increase the validity of the normalised spectral shape,
each wave system should be modelled individually. This partition-
ing or separating the wave components is important for evaluation
of ship responses. On the other hand, for onboard estimation of sea
state, assessment of spectral components is favourable for tracing
the evolution of those wave systems all through the operational
period.
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86
A common representation for combined seas is simply to super-
pose two wave spectra:
S(ω, ) = Sw (ω, ) + Ssw (ω, )
where Sw (ω, ) is the wind sea spectrum and Ssw (ω, ) is the swell
spectrum; which itself could be a combination of several swell
systems. However, standard engineering calculations often deal
with spectra that are described by double peaked (bimodal) spec-
tra at most. According to the literature, e.g. [8], decomposition of
wave spectra into maximum two components provides an ade-
quate description of sea state in terms of the integrated parameters,
i.e. the significant wave height, the mean/peak period, and the mean
wave direction.
3.2. System classification of wave spectrum
In general, wind seas are more irregular and short-crested,
respond quickly to wind variations, so they are characterised by
a rather broad spectrum, the energy of which is concentrated
at higher frequencies (between 0.1 and 4 Hz). But swell systems
are more complicated to model since no single self-similar form
exists as in wind driven waves. Swell events consist of rather
regular long-crested waves, which are not strongly affected by
wind. The spectral shape of swell is narrower and more peaked.
In other words, for the same initial spectrum in the storm area,
the shape of the swell spectrum depends on the distance trav-
elled. The swell frequencies usually lie between 0.03 and 0.2 Hz
[9].
Sine a swell system results from a wind sea that propagates
away from the generation area, it is commonly described by the
same spectral model as wind sea. However, modelling such nar-
row peaks by a JONSWAP spectrum implies the peak enhancement
factor to take much higher values the commonly used range
(1–7). A JONSWAP spectrum with a peak enhancement factor
between 7 and 10 is known to be adequate for swell modelling
[10–12].
For the sake of system classification in wind sea or swell, most
methods rely on determination of a separation frequency, ωs , for
a particular spectrum. Wave frequencies higher than ωs are gen-
erated by local winds and wave components at frequencies lower
than ωs correspond to swell. After separating the spectrum into
low-frequency and high-frequency peaks, a JONSWAP model is
fitted around each component, the summation of them is then sup-
posed to describe the whole spectrum. Fig. 1 shows an example of
this spectral partitioning.
In case of ship response-based wave estimation using
non-parametric formulation, partitioning procedure can be imple-
mented, where an algorithm finds a matrix of the locally largest
neighbours in the spectrum. This has been employed for wave esti-
mation using Bayesian method [2,6]. The study in [5] is also capable
of separating wind sea and swell.
In the present study, however, the partitioning is introduced
based on the parametric approach. The procedure is quite different
from the above mentioned common partitioning algorithms in the
sense that neither the spectral shape nor the sea state parameters
are known. Thus, wind sea/swell identification and estimation of
associated parameters are accomplished simultaneously.
In fully developed seas, the peak frequency of the
Pierson–Moskowitz (PM) spectrum, denoted by ωPM , is calculated
by [9]:
g
ωPM = 0.82
Uw
where Uw is usually 10 m elevation wind speed in m/s. This quan-
tity can be calculated during operations as the wind speed can be
75
(6)
Fig. 1. Spectral partitioning and fitting.
continuously measured onboard the ship. A general representation
of the splitting frequency assumes it to be proportional to the peak
frequency of the PM spectrum. So,
g
ωs = (8)
ˇUw
where ˇ is an empirical constant, usually chosen between 1.2 and
1.5. ωs is in rad/s. In partially developed seas, wind generates ran-
dom pressure fluctuations at the sea surface that produce small
waves with shorter periods than fully developed sea. The peak
frequency of the wind sea in JONSWAP spectrum is an empirical
function of both wind speed and fetch length. Since information
about the fetch length is not usually available, a so-called wave age
criterion can be used as [13]:
cp ≤ ˇUw | cos(w − )| (9)
cp is the phase speed of the wind sea, defined as the speed of waves
at the peak and it is calculated by:
g
cp = (10)
ωPM
in deep water. w in Eq. (9) is the wind direction and  is the
mean wave direction as mentioned before. The inequality in Eq.
(9) applies to cover partially as well as fully developed sea states.
In terms of the peak frequencies of wind seas, Eq. (9) becomes:
g
ωp ≥ (11)
ˇUw | cos(w − )|
Another condition commonly used for a system to be identified as
wind sea is that the mean direction of the wave train must be at
most 90◦ apart from the wind direction [13]:
 
w −  ≤ 90 (12)
These conditions can be used as complements in order to
upgrade the estimates of wind seas, which are exposed to filtering
characteristics of ships.
4. Response selection
For wave estimation purposes based on short term responses,
sufficient resolution of the transfer functions (RAOs) is very impor-
tant. In addition, the transfer functions must cover a proper region
of wave frequencies, especially in the area with the highest wave
energy. Otherwise the wave spectrum may be poorly estimated.
(7) Depending upon natural frequencies, some responses might not be
able to capture the high frequency part of the spectrum whereas
some others may do. Therefore, several responses should be used
to make sure that the entire spectrum has been sensed and also
76 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86

RAO Heave RAO Pitch

1.4 0.014
0 deg.
1.2 20 deg. 0.012
40 deg.
60 deg.
1 80 deg. 0.01
100 deg.

RAO [rad/m]
RAO [m/m]

0.8 120 deg. 0.008

140 deg.
160 deg.
0.6 0.006
180 deg.

0.4 0.004

0.2 0.002

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency [rad/s] Frequency [rad/s]

RAO Sway RAO Roll

4 0.14

3.5 0.12

3
0.1
2.5

RAO [rad/m]
RAO [m/m]

0.08
2
0.06
1.5
0.04
1

0.5 0.02

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency [rad/s] Frequency [rad/s]

RAO VBM
1400

1200

1000
RAO [MNm/m]

800

600

400

200

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency [rad/s]

Fig. 2. Response Amplitude Operators (RAO1) at V = 20 kn, T = 14.5 m (The legends are identical in all plots).

to provide a sufficient number of equations in the system that is
not less than the parameters. Selection of the best combination of
responses is very important. It is often seen that shifting between
different responses may influence the estimations remarkably. The
optimum selection of responses may not be identical for all ships
and all operational conditions. This choice should be made based
on prior knowledge about the transfer functions of a particular ship
in different operational conditions.
4.1. Vessel data
linear potential theory. The transfer functions of vertical bending
moments are calculated using an in-house software (ISHIP) based
on strip theory. Fig. 2 shows the amplitudes of the transfer functions
for the operational condition specified in Table 1. In these figures,
180 deg. stands for head sea condition.
Heave and pitch are known as the most reliable responses in
a sense that the transfer functions can be calculated with a good
accuracy based on strip theory, panel method or experiment. So
these motions are popular in the application of wave estimation. It

A CMA-CGM 9415 TEU container ship is considered, which has Table 1

been under investigation for TULCS project.1 The main character- Ship characteristics and operational condition.
istics of the vessel with the typical operational condition are given
Properties Values
in Table 1.
Overall length [m] 349.0
The transfer functions of the motions of the ship are provided
Beam [m] 42.8
by BUREAU VERITAS using HydroStar software which is based on Draft [m] 14.5
Maximum capacity [TEU] 9415
DWT [ton] 113,000
1 Speed [kn] 20
Grant No. 234146.
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86 77

is also common to use vertical motion at a point distant from the
center of gravity.
In addition, due to asymmetric characteristic, roll or sway is usu-
ally included as well to recognise whether the waves enter from
port or starboard side of the ship. However, sway does not have
a restoring force, so the error in calculation of this response may
be quite large particularly in high frequency domain. On the other
hand the assumption that roll motion is linear is sceptical in higher
excitations. It is also clear from Fig. 2 that the transfer functions of
sway and roll are unsmooth or very small in some wave directions.
Those inaccuracies of transfer functions for the latter two motions
may insert considerable uncertainties to the estimations. Never-
theless, the selected responses can be weighted non-uniformly to
compensate the uncertainties. This will be discussed in Section 5.2.
Wave-induced vertical bending moment is another response
that can be used for wave estimation [14]. As seen in Fig. 2, this
response is significant along a wide range of frequencies, so it
is quite beneficial for estimation of wave spectrum as discussed
above. Vertical bending moment is proportional to the wave height
if the ship tends to be wall sided near the operational draft. But in
case of small block coefficient or large wave heights, the assump-
tion of wall sided is not valid any more and the vertical bending
moment becomes a non-linear function of wave height [15].
In general, transfer functions of accelerations have a slower roll
off than motions, meaning that their amplitudes decay slower at
the considered responses, i.e. vertical motion at port (at midship
section and 19.25 m from centreline, where the actual sensors
are located), pitch and sway at the centre of gravity and vertical
bending moment at mid-ship section. For the particular ship and
operational condition, as discussed in Section 4, the above men-
tioned combination sufficiently covers a relatively wide range of
frequencies. As a complement, the rates and accelerations of these
responses are also simulated as discussed before. It is noteworthy
that the roll motion is not used directly but only as a contribution
of vertical motion at the port side.
Since wave records and responses are assumed to be Gaussian
processes, the time series are generated using a set of uncorrelated
standard normal distributed variables [17]. Such long time series
(3600 s) assure that the actual energy of the response records is
preserved and the risk of statistical outliers is minimised. Fig. 3
shows samples of time history simulations. A noise is also added to
the time series as will be discussed in Section 6.
5.2. Optimisation
In order to solve Eq. (4) numerically, the integrals must be dis-
cretised frequency- and directional-wise. So, the cost function can
be expressed as:

X

Y

m0,ij = ııω Hi (ωx , y )Hj∗ (ωx , y )S(ωx , y ) + rij ,

high frequencies and, hence, using a combination of displacements
x=1 y=1
and accelerations (or rates) is recommended in order to improve
the estimates at both low and high frequencies [16]. It is well- i, j = 1, 2, . . ., N (13)
known that the complex valued transfer functions of response rates
where X and Y are the number of frequencies and wave directions,
and accelerations can be evaluated by single and double differenti-
respectively. rij is the residual between the two sides of Eq. (4).
ation, respectively, of displacements. Consequently, corresponding
As explained in Section 2, the variable space of the cost function
equations in the cost function, Eq. (5), will be the second and the
involves five “fit parameters”, which are Hs , Tp , , , and smax for
fourth moments of the response spectra, respectively.
individual wave systems. Estimation of these parameters are car-
ried out by the least square method. The sum of squared residuals,
5. Numerical study SSR, in a normalised form is:

5.1. Generation of response measurements

   
2
N N X Y
1
SSR = [m0,ij − ııω Hi (ωx , y )Hj∗ (ωx , y )S(ωx , y )]
Different wave scenarios characterised in Table 2 are considered m20,ij
i=1 j=1 x=1 y=1
here using JONSWAP model as described in Section 2.2. The wind
(14)
speeds are chosen based on statistical wave plots for the typical
periods and a fetch length of 800 km. The cases A through H repre-
sent unimodal wave spectra and the cases I through L are bimodal In this paper, the number of equations in the matrix of Eq. (14)
sea states with separately located peaks, whereas cases M through P is primarily adjusted to increase the speed of the optimisation and
represent bimodal waves with closely located peaks. “+” and “−” in also to obtain reliable results. Since the sway motion deals with
the directions denote waves from starboard and port sides, respec- hydrodynamic uncertainties, mentioned in Section 4, it is not given
tively and 180 deg. corresponds to headsea condition. The wave the same weight as other responses in such a way that the diagonal
direction is discretised from −180 to 180 deg. with 10 deg. spacing. (i = j) equations of this motion is neglected. In other words, this
The frequency band is from 0.06 to 2 rad/s with non-equal random response is only used for cross-spectral terms to identify whether
sampling frequencies to avoid repeating data [2]. the waves are entered from port or starboard.
The complex valued transfer functions from Section 4 are used Moreover, in the current spectral moment-based method,
to simulate the sensor installations. The ship speed is 20 kn in all numerical problems, i.e. large magnitudes of residuals, can be
cases and assumed to be north head so that the relative direction experienced during the optimisation because the areas under the
is equal to the true wave direction. The transfer functions are cross-spectral curves are relatively small. In addition, smoothing of
interpolated and extrapolated for the corresponding frequency cross-spectra adds uncertainty to the model. Therefore, the cross-
range of waves. 15 sets of 3600 s long realisations are generated for spectral calculations that correspond to off-diagonal components

Table 2
The main parameters of the studied wave cases.

Cases Wind sea Swell Wind speed

Hs (m) Tp (s) (deg.) smax  Hs (m) Tp (s) (deg.) smax  Uw (m/s)

A,B,C,D 3 8 45,90,135,180 10 2 0 – – – – 14
E,F,G,H 0 – – – – 5 15 45,90,135,180 25 4 5
I,J,K,L 3 8 45,−90,135,90 10 2 5 15 −135,90,45,180 25 4 14
M,N,O,P 3 8 45,−90,135,90 10 2 2 12 −135,90,45,180 25 4 14
78 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86

x 10
9
solve the current constrained nonlinear optimisation problem.
Multistart, as the name implies, runs fmincon from multiple
1.6
starting points using a quadratic algorithm. The results of each
fmincon run is stored in a vector, and in the end, the best result is
0.8 selected as the global minimum.
VBM [Nm]

Genetic algorithm is a method based on natural selection, the

0 process that drives biological evolution. This method repeatedly
modifies a population of individual solutions. At each step, it selects
-0.8 individuals at random from the current population to be parents.
This selection is based on their fitness. The parents are then used
to produce children for the next generation by making random
-1.6
changes. The direction and the step length of those changes satisfy
0 100 200 300 400 500 600 700 800 the bounds and linear constraints. Over successive generations, the
Time [s]
population evolves toward an optimal solution that minimises the
cost function.
6 The genetic algorithm differs from the classical derivative-based
optimisation algorithm in two main ways: First, the former method
4
generates a population of points at each iteration and the best point
in the population approaches an optimal solution, whereas the
Vertical Motion [m]

2
classical algorithm generates a single point at each iteration and
0 the sequence of points approaches an optimal solution. Second, in
genetic algorithm, the next population is selected by using random
-2 number generators; whereas in the classical method, the next point
in the sequence is selected by a deterministic computation.2
-4 It is notable that global optimisation is not promising unless
some constraints are applied to the fit parameters. The following
-6 boundaries are generally valid in ocean waves:
0 100 200 300 400 500 600 700 800
Time [s]
0 ≤ Hs ≤ 15 [m] (15)
1.5 6 ≤ Tp ≤ 20 [s] (16)

1 Another constraint is associated with the global steepness limit

of small breaking waves riding on long waves [18]:
0.5 
Pitch [deg.]

0 11.4
-0.5
-1
-1.5
0 100 200 300 400 500
Time [s]
Fig. 3. Sample of time history simulations of vertical bending moment, vertical
motion at port side, and pitch motion.
(i =
/ j) in Eq. (14) are limited to pitch and sway motions in this
study. Besides, the absolute values of the cross-spectral moments
are taken into account on the both measured and estimated sides
to facilitate the optimisation.
Minimisation of Eq. (14) can be solved using gradient-based
methods. For instance “fmincon” from MATLAB is very useful when
linear or/and nonlinear constraints exist. This function begins from
an initial guess, iterates according to a given update scheme, and
finishes when the stopping criteria are met. The final iteration is
a local minimum if the conditions are fulfilled and the Hessian is
positive definite.
It is observed here and in the previous studies that for esti-
mation of main parameters, a large number of local minima may
be present and the calculations are highly dependent on the
initial values. To compensate, a global search basin is required
rather than local minimisation in the least square analysis as also
mentioned in [4]. This can be accomplished by either “MultiStart”
class or “Genetic Algorithm” from MATLAB optimisation toolbox
to achieve the global minima. Both methods are well suited to
Hs 2
< Tp , where Hs [m], TP = [s] (17)
g ωp
The above mentioned restrictions provide proper conditioning for
global optimisation of Hs and Tp . But initialization of the mean wave
direction is also necessary to reach accurate estimates. In real appli-
cation of onboard wave estimation, the range of initial populations
600 700 800 of wave direction can be restricted by using the observations or the
estimates from previous time steps. This is valid since the direction
of waves is not subject to big changes in a short time.
Every wave spectrum is categorised as wind or swell domi-
nated. The dominating peak frequency of the wave spectrum is
extracted by fitting a single JONSWAP model.  and smax are fixed
here for the sake of simplicity. The estimated peak frequency
in this pre-estimation step is compared with the PM peak fre-
quency, which is approximated by using the actual measured wind
speed, Eq. (7). If ωp < ωPM , the waves are assumed to be swell
dominated and if ωp > ωPM , they are wind dominated. This cate-
gorisation should finally conform with the following criterion based
on the relative ratio of zero order spectral moments (m0 ) of swell
and wind sea systems that are split by the separation frequency,
Eq. (8) [19]:
m0ws
= (18)
m0sw
where the subscripts ws and sw stand for wind sea and swell par-
titions, respectively. Those wave fields with  smaller than one
represent swell dominated waves, and those with  greater than
one correspond to wind dominated waves. After identifying the
2
http://se.mathworks.com/help/gads/index.html.
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86 79

dominated system, a more accurate estimation of the parame- a deviation from the mean direction is applied when splitting the
ters is obtained through the partitioning approach as described in response spectra.
Section 3.
The wave spectrum should be split into low frequency and ⎧
⎪
⎪ ωe (0 )
high frequency parts using the separation frequency which deter- ⎪
⎪
mines the integral bounds in Eq. (4) for each partition. Clearly, the ⎨ ωe (0 − )
left-hand side of Eq. (4) is integrated with respect to encounter ωe(s,wind) = min (21)
⎪
⎪ ωe (0 + )
frequencies, whereas the separation frequency is determined in ⎪
⎪
wave frequency domain. Thus, Eq. (2) is employed to calculate the ⎩
ωe(s,swell) , if any
corresponding separation frequency of encounter. As mentioned
before, this transformation is not straightforward since the mean
wave direction is unknown. So, a range of mean wave directions is This deviation is assumed a bit larger than swell ( = 35 deg.)
considered to initiate the encounter splitting frequency, according since wind generated waves are more short-crested as explained
to which the response spectra are split. The optimisation is then in Section 3.2. In the last step, the significant wave height of the
carried out under the condition that the difference between the secondary peak is adjusted by fitting the whole measured (gen-
estimated mean wave direction and the initial value is less than erated) response spectra to a double peaked wave model fixing
20 deg. Then the optimum solution is the one with minimum SSR the other estimated parameters. This can avoid the total energy of
among the whole range. the waves to be overestimated since the wind sea part may have
Since bimodal wave systems are likely to be omnidirectional, an overlap with the swell part particularly for waves with closely
the above procedure is applied first on the dominant system located peaks. A flowchart overview of the optimisation procedure
and then on the secondary system. For the swell part, displace- is shown in Fig. 4.
ments and rates are taken into the matrix equation. The separation
encounter frequency used to split the response spectra is calculated
as:
⎧
⎪
⎪ ωe (0 )
⎪
⎪
⎨ ωe (0 − )
ωe,(s,swell) = max (19)
⎪
⎪ ωe (0 + )
⎪
⎪
⎩
ωe(s,wind) , if any

0 is the initial mean direction and implies a deviation from

the mean direction due to short-crestedness of waves. The mag-
nitude of this interval depends on the spreading factor, but a fixed
value of = 25 deg. is considered here for simplicity. In case of wind
dominated waves, the separation frequency of the wind sea part,
ωe(s,wind) , is also considered to avoid any gap between the two parti-
tions in the encounter frequency domain. It is also of concern that in
following sea condition, the triple-valued function problem in Eq.
(2) might cause error in the estimattions since a range of encounter
frequencies may be ignored. The above mentioned interval can also
compensate this problem.
For the wind sea partition, estimation is more complicated than
swell. As mentioned before, the responses of a large ship with a high
inertia filter out the high frequency part of the spectrum and using
the same conditions as swell, Eqs. (15)–(17), may be insufficient. To
overcome the problem, additional constraints are introduced. For
instance in JONSWAP models, an upper bound is considered for the
steepness of wind generated waves [20]:

Hs
Tp < 15.7 (20)
g

As wind sea classification is made using a criterion related to

Pierson Moskowitz spectrum, the peakedness factor is fixed to a
constant value of 1 as in many researches, e.g. [21]. Eq. (12) is
also applied using the measured wind direction. According to this
restriction, the estimated mean direction of wind generated waves
should not deviate significantly from the actual wind direction,
otherwise the component can not belong to the present wind sea.
Eq. (11) can be useful as well for estimation of wind sea direction.
However, it is observed that using this inequality as a constraint is
problematic for convergence of the optimisation. So, this equation
is merely helpful to adjust the wave direction, fixing the estimated
peak period. For this high frequency part, beside the displacements,
the accelerations are used instead of rates. Similar to the swell part, Fig. 4. Flowchart of the optimisation procedure.
80 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86

6. Results and discussion
6.1. Optimisation algorithms
In this paper, regardless of knowing that whether the wave
spectrum to be estimated is unimodal or multipeaked, the above
mentioned procedure is applied to all wave cases in Table 2. It was
observed that the results from the gradient-based method (Multi-
start) and the genetic algorithm approach are very similar except a
few cases that the genetic algorithm shows more stability in terms
of initialisation and convergence. On the other hand, the computa-
tional time of the Multistart method is approximately 8–10 min for
estimation of the whole spectrum whereas the genetic algorithm
is 2–3 times slower (intel(R) Core(TM) i7, CPU 2.40 GHz, memory
8G). Thus, it is not possible herein to distinguish which algorithm
is more efficient for the current optimisation problem and justifi-
cation of this choice requires further studies. The results from the
genetic algorithm are only shown here.
mathematical models, modelling of boundary value problem, errors
of body geometry modelling and inaccurate mass distribution. So,
it is of interest to consider the uncertainty of transfer functions
in the wave estimation approaches. For this purpose, a different
set of RAOs for the motions (vertical motion, pitch and sway) is
also applied here to estimate the waves. Those transfer functions,
called RAO2 hereafter, are calculated using the in-house linear strip
theory program (ISHIP) that was mentioned before. Fig. 5 shows
the amplitudes of these transfer functions for the same operational
condition as Section 4. Comparing the magnitudes of Fig. 5 with
Fig. 2, the peak values of heave and pitch amplitudes are 10–15%
different. The difference in sway motion, though, is larger partic-
ularly in very low frequencies and also in high frequencies when
the relative directions of waves are 20–40 degrees; i.e. close to fol-
lowing sea condition. As seen in the figures, the calculation error in
this condition can reach 75%. This is consistent with our expecta-
tion about the uncertainty of sway motion calculation as discussed
earlier.

6.2. Uncertainty consideration 6.3. Results

In order to consider the errors in the actual measurements, it
is usual to add a noise to the generated signals. This is imple-
mented here using a white noise with zero mean and standard
deviation equal to 10 percent of the responses amplitudes. Thereby,
the results are not changed more than 2–3%. This is expected,
because the noise does not change the standard deviation or the
variance of the whole time series significantly, and since the calcu-
lations are merely based on the spectral moments or the variances
of responses, it does not affect the results considerably.
Although calculations of transfer functions using e.g. strip the-
ory or 3D panel methods have grown stronger during last decades,
the transfer functions are still subjected to errors due to different
The estimations based on the both sets of RAOs are shown in
Tables A.1–A.4 in Appendix A, which are compared to the real
values. RAO1 represents the original transfer functions as used in
the simulations (HydroStar). The comparisons between the esti-
mations and the true waves are also visualised in Figs. 6–8. The
standard deviations are included as the small-size pin bars. The
average values and the standard deviations correspond to the 15
realisations. As the results show, this number of data sets seems to
be sufficient since the standard deviations are mostly small.
It can be seen in Fig. 6 that for both swell dominated and wind
dominated unimodal cases, the estimations based on RAO1 prop-
erly match the true waves particularly in terms of the significant

RAO Heave RAO Pitch

1.4 0.014
0 deg.
1.2 20 deg. 0.012
40 deg.
60 deg.
1 0.01
80 deg.
RAO [rad/m]
RAO [m/m]

100 deg.
0.8 120 deg. 0.008
140 deg.
0.6 160 deg. 0.006
180 deg.

0.4 0.004

0.2 0.002

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency [rad/s] Frequency [rad/s]

RAO Sway
3

2.5

2
RAO [m/m]

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency [rad/s]

Fig. 5. Response Amplitude Operators (RAO2) at V = 20 kn, T = 14.5 m (The legends are identical in all plots).
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86 81

Fig. 6. Estimated parameters for unimodal wave scenarios (The legends are identical in all plots).

Fig. 7. Estimated wind sea parameters for bimodal wave scenarios (The legends are identical in all plots)
82 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86
Fig. 8. Estimated swell parameters for bimodal wave scenarios (The legends are identical in all plots).
wave height, the peak period and the mean wave direction. The
highest wave direction error is 25 deg. in case H. The estimations
based on RAO2 are very similar, which assure the consistency of the
method in these cases. For the pure wind seas, when a secondary
spectrum is fitted to the low frequency part, the algorithm often
converges giving a small magnitude of significant wave height.
However, modification of this secondary peak through the fitting of
the whole spectrum in the final step, always results in a significant
wave height very close to zero. This implies that the secondary peak
can be correctly neglected in this case. For the pure swell events,
the wind speed is quite low, 5 m/s, according to Table 2, so the
separation frequency is a large value and the high frequency range
to be fitted to the secondary spectrum is narrow. Consequently,
the program does not converge to an acceptable point within the
range of wind sea frequencies. As expected from the true wave, it
can be concluded that no wind generated peak is recognised in the
spectrum.
According to Figs. 7 and 8 for bimodal wave spectra, the par-
titioning approach precisely estimates the significant wave height
and the peak period of both swell and wind sea part in all wave
scenarios even in case of closely located peaks, where the two sys-
tems are overlapped. The maximum errors in Hs and Tp are 1 m and
1.12 s, respectively, based on RAO1. The error of mean wave direc-
tions in wind sea, Fig. 7, does not exceed 25 deg., which is fairly
acceptable. However, the standard deviation is quite large in case
P. For the mean wave direction in the swell part, Fig. 8, the errors are
not notable except in case P where the direction has 50 deg. error.
Referring to the results based on RAO2 in bimodal waves, it
can be seen that they match with RAO1 to some extent. However,
particularly in wind sea partitions, the significant wave heights are
up to 1.3 m deviated from the real values (Case M). This could be
due to the uncertainty in the high frequency part of the response
functions. The peak periods have also notable errors in a few cases,
particularly in stern quartering swell partitions; i.e. cases K and O.
It is also observed that the estimated spreading parameter is
critical to the initial guess particularly in the wind sea systems. In
the swell systems, this parameter is usually overestimated, which
conforms with the outcome of other motion based researches, e.g.
[22]. However, it is found that the estimates of the basic three
parameters (Hs , Tp , ) are quite independent of the spreading
parameter and also the peakedness factor. This confirms that the
current method, where only the moments of spectra are fitted, can
estimate the main integral parameters of wave systems, but may
not be proper to be used for evaluating the spectral shape of those
waves. A couple of polar plots of the estimated directional wave
spectra are also illustrated in Appendix A.
6.4. Further discussions
The actual limiting wave characteristics that can be estimated
by ship responses depend highly on the ship length. In general, very
low wavelengths relative to the length of the ship cannot be esti-
mated using the ship responses. The limit is also dependent on the
operational condition, i.e. the ship speed and the relative direction
of waves. Thus, looking at the particular transfer functions is essen-
tial. Focusing on heave and pitch motions in head sea condition, the
transfer functions get close to zero at frequencies corresponding to
wave lengths shorter than the ship length (higher than 0.5 rad/s
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86 83

in this case study). Hence, wave bending moment is added to the it is inevitable that in reality, failure may be faced in estimation
response combination. As it is clear from Fig. 2, this response is procedure because of possible errors in acquiring data during mea-
more sensitive to the higher frequencies up to 1 rad/s in head seas. surements of responses and wind as well as evaluating spectra from
In addition, the filtering effects are compensated to some extent by a finite length of data.
the additional constraints applied to the optimisation problem in
the wind sea parts. As the results show, this strategy is efficient to
estimate waves with minimum peak period of 8 s. However, when Acknowledgements
the actual peak period is lower than that, the current method is not
expected to be reliable since all the responses diminish at frequen- The authors sincerely acknowledge Class NK for supporting the
cies higher than 1 rad/s as seen in Fig. 2. It should be noted that research project in cooperation with the Technical University of
the limiting sea state also depends on the wave height. Because in Denmark.
severe seas, the linear transfer functions are not necessarily accu-
rate enough, so it is fair to claim that this method is more accurate
when the significant wave height is less than, say, 6 m. Appendix A.
All in all, it can be argued from estimations of the considered
waves that the presented method is efficient and fairly robust. The
results based on full-scale data of the same ship have supported
Table A.1
the applicability of the estimation method in unimodal waves [23].
Parameters of unimodal spectrum (Wind Sea).
However, further investigations should be carried out on real data
to verify the applicability of the partitioning approach. It is noted Case Hs (m) Tp (s) (deg.) smax

that the estimations are highly dependent on the wind data and the A real 3 8 45 10
inaccuracy of this information onboard ships is not considered in mean (RAO1) 2.1 7.2 63 15
mean (RAO2) 2.5 7.8 60 17
this study.
std (RAO1) 0.19 0.36 15 4.4
std (RAO2) 0.2 0.16 13 4.2
7. Conclusion B real 3 8 90 10
mean (RAO1) 3.9 7.6 81 18
In this paper, a refined approach for estimation of sea state is mean (RAO2) 3.2 8.7 80 10
std (RAO1) 0.4 0.2 12.5 5
proposed using measured ship responses. The method is based on
std (RAO2) 0.45 0.5 15 8
a parametric model applied on double-peaked directional wave
spectra. The optimisation problem is formulated based on energy C real 3 8 135 10
mean (RAO1) 3.5 6.7 136 16
balance of responses in terms of spectral moments, being measured
mean (RAO2) 3.3 9.1 157 15
and theoretically calculated. This approach applies a sequential par- std (RAO1) 0.43 0.35 10 4
titioning algorithm, which is able to classify swell and wind sea std (RAO2) 0.53 0.7 15 6
events. Real-time wind information is utilised to restrict the fit D real 3 8 180 10
parameters. This can overcome the lack of information in the high mean (RAO1) 4 7.2 −176 12
frequency part of the spectrum, which is due to filtering effects of mean (RAO2) 2.9 8.5 171 18
large ships. The procedure is quite simple compared to traditional std (RAO1) 0.64 0.6 2 10
std (RAO2) 0.5 0.7 6 12
methods and the computation time is short, which is advantageous
for real-time estimations and predictions.
The method is evaluated using simulated data for a large con-
tainer ship as a case study. Although different combination of
responses may influence the estimations, it is believed that a proper Table A.2
weighting of response equations can be accomplished for a par- Parameters of unimodal spectrum (Swell).
ticular ship and operational condition. Selection and weighting of Case Hs (m) Tp (s) (deg.) smax 
the responses should be carried out in such a way that the uncer-
E real 5 15 45 25 4
tainties are decreased in terms of both hydrodynamic calculations mean (RAO1) 4.9 14.6 43 27 5
and measurements. The procedure is examined for different wave mean (RAO2) 4.9 14.7 41 27 5.8
scenarios including unimodal and bimodal spectra. Global optimi- std (RAO1) 0.34 0.52 7 19 2.26
sation is applied and the averaged values of the estimated main std (RAO2) 0.35 0.62 8.5 19 1.9
parameters; i.e. the significant wave heights, the peak periods, and F real 5 15 90 25 4
the mean wave train directions are generally very close to the true mean (RAO1) 4.8 14.8 90.8 23 6
values and the standard deviations are quite satisfactory. mean (RAO2) 4.9 14.5 81 27 6
std(RAO1) 0.5 0.32 1.3 4 1.8
Following previous studies, this paper proves the matureness
std(RAO2) 0.5 1.2 3.2 3 2.1
and reliability of the response-based sea state estimation method
G real 5 15 135 25 4
for ships with advance speed. Beside the original transfer functions,
mean (RAO1) 5.5 15.9 135 52 6
another set of transfer functions are used here in the estimation mean (RAO2) 5.2 16.4 145 57 5.5
stage. This can address the uncertainties in transfer function cal- std (RAO1) 0.5 0.37 2 14.4 0.9
culations. The results based on the two sets of transfer functions std (RAO2) 0.6 0.7 5.5 11.5 1
are consistent to some extent, specially in unimodal cases. The H real 5 15 180 25 4
study on full-scale measurements of the same container ship [23], mean (RAO1) 6 15 155 61 6
also validates the robustness of the proposed method for uni- mean (RAO2) 5.9 15 167 71 6.4
modal waves. However, further studies should be implemented to std (RAO1) 0.4 0.5 2.5 12 0.6
std (RAO2) 0.4 0.4 4.7 12 0.3
assess the overall validity of the partitioning procedure. Because
84 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86

Table A.3
Parameters of bimodal spectrum (1).

Case Wind sea Swell

Hs (m) Tp (s) (deg.) smax Hs (m) Tp (s) (deg.) smax 

I real 3 8 45 10 5 15 −135 25 4
mean (RAO1) 3.1 8.8 66 15 5.2 15 −160 33 1.5
mean (RAO2) 4 8.5 52 15 4.2 15 −101 60 5.8
std (RAO1) 0.7 0.49 10 0 0.65 0.58 12 5.8 0.6
std (RAO2) 0.3 0.55 1.5 5 0.8 0.45 13 20 2

J real 3 8 −90 10 5 15 90 25 4
mean (RAO1) 3.2 8.6 −106 18 4.4 16.6 98 27 4.4
mean (RAO2) 3.6 9.2 −92 20 3.8 15.1 120 47 5.4
std (RAO1) 0.57 0.75 22 2 1.3 1 7.6 14 2.9
std (RAO2) 0.25 1.3 17 0 0.08 0.6 30 20 2.1

K real 3 8 135 10 5 15 45 25 4
mean (RAO1) 2.3 7.3 120 12 5.5 13 49 65 6
mean (RAO2) 3.4 6.7 141 15 5.8 12.8 3 64 7
std (RAO1) 0.5 0.8 13 0 0.4 0.15 14 20 4
std (RAO2) 0.2 0.2 10 3 0.5 0.14 16 23 2

L real 3 8 90 10 5 15 180 25 4
mean (RAO1) 3.6 9.12 89 15 5.3 16 174 49 4.8
mean (RAO2) 3.3 6.8 100 18 5.8 14.15 176 59 5
std (RAO1) 0.5 0.17 2 0 0.93 2.7 4 28 2.1
std (RAO2) 1 0.9 2 4 0.81 0.6 11 12 3

Table A.4
Parameters of bimodal spectrum (2).

Case Wind sea Swell

Hs (m) Tp (s) (deg.) smax Hs (m) Tp (s) (deg.) smax 

M real 3 8 45 10 2 12 −135 25 4
mean (RAO1) 2.7 8.3 56 19 2 13.9 −140 39 7
mean (RAO2) 4.3 9.7 25 17 1.7 14.2 −155 53 3.4
std (RAO1) 0.5 0.8 18.6 3.5 0.3 1.4 31 13 2
std (RAO2) 0.7 1.4 9 6 0.3 0.3 15 11 2

N real 3 8 −90 10 2 12 90 25 4
mean (RAO1) 2.7 9.5 −102 21 2.6 13 110 38 5.2
mean (RAO2) 3.8 7.7 −85 20 2 12.4 85 33 6
std (RAO1) 0.66 1.28 17 3.8 0.19 1.7 25 22 4.3
std (RAO2) 0.6 1.4 13 0 0.3 1.8 29 20 5

O real 3 8 135 10 2 12 45 25 4
mean (RAO1) 4 8.7 132 15 2.3 11.2 50 30 6.8
mean (RAO2) 3.8 7.3 112 15 3.2 9.4 35 75 4.3
std (RAO1) 0.6 0.9 15 0 0.18 0.95 9.3 7 2.4
std (RAO2) 0.7 0.8 23 5 0.6 0.5 4 2 3

P real 3 8 90 10 2 12 180 25 4
mean (RAO1) 3.8 8.1 96 15 2.2 12.8 132 53 5.6
mean (RAO2) 4.1 8.2 88 19 2.4 12.6 127 45 6
std (RAO1) 0.37 1 34 4 0.3 0.4 6 11 2.1
std (RAO2) 0.8 0.8 42 7 0.3 0.5 5 13 2.8
 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86 85

Generated Spectrum, Case I Estimated Spectrum

90 0.2 [Hz] 90 0.2 [Hz]
120 60 120 60

150 0.1 30 150 0.1 30

180 0 [deg.] 180 0 [deg.]

210 330 210 330

240 300 240 300

270 270

Generated Spectrum, Case K Estimated Spectrum

90 0.2 [Hz] 90 0.2 [Hz]
120 60 120 60

150 0.1 30 150 0.1 30

180 0 [deg.] 180 0 [deg.]

210 330 210 330

240 300 240 300

270 270

Generated Spectrum, Case N Estimated Spectrum

90 0.2 [Hz] 90 0.2 [Hz]
120 60 120 60

150 0.1 30 150 0.1 30

180 0 [deg.] 180 0 [deg.]

210 330 210 330

240 300 240 300

270 270

Generated Spectrum, Case P Estimated Spectrum

90 0.2 [Hz] 90 0.2 [Hz]
120 60 120 60

150 0.1 30 150 0.1 30

180 0 [deg.] 180 0 [deg.]

210 330 210 330

240 300 240 300

270 270

Fig. A.1. Examples of contour plots of the generated and the estimated bimodal spectra.

References
[1] Tannuri EA, Sparano JV, Simos AN, Da Cruz JJ. Estimating directional wave
spectrum based on stationary ship motion measurements. Appl Ocean
Res 2003;25(5):243–61, http://dx.doi.org/10.1016/j.apor.2004.01.003 http://
linkinghub.elsevier.com/retrieve/pii/S0141118704000070.
[2] Nielsen UD. Response-based estimation of sea state parameters - influence of
filtering. Ocean Eng 2007;34(13):1797–810, http://dx.doi.org/10.1016/
j.oceaneng.2007.03.002 http://linkinghub.elsevier.com/retrieve/pii/
S0029801807000625.
[3] Nielsen UD, Stredulinsky DC. Sea state estimation from an advancing ship
A comparative study using sea trial data. Appl Ocean Res 2012;34:33–44,
86 N. Montazeri et al. / Applied Ocean Research 54 (2016) 73–86
http://dx.doi.org/10.1016/j.apor.2011.11.001 http://linkinghub.elsevier.com/
retrieve/pii/S0141118711000873.
[4] Pascoal R, Guedes Soares C, Sorensen AJ. Ocean wave spectral estimation using
vessel wave frequency motions. J Offshore Mech Arct Eng 2007;129(2):90,
http://dx.doi.org/10.1115/1.2426986 http://offshoremechanics.
asmedigitalcollection.asme.org/article.aspx?articleid=1456353.
[5] Pascoal R, Guedes Soares C. Non-parametric wave spectral estimation
using vessel motions. Appl Ocean Res 2008;30(1):46–53, http://dx.doi.
org/10.1016/j.apor.2008.03.003 http://linkinghub.elsevier.com/retrieve/pii/
S0141118708000163.
[6] Tannuri EA, da Silva Bispo IB, Simos AN, Filho A, da Cruz J, Carvalho R. Devel-
opment of an on-board wave estimation system based on the motions of a
moored FPSO: preliminary assessments from a field campaign. In: Proceedings
of the ASME 2013 32nd International Conference on Ocean, Offshore and Arctic
Engineering, ASME. 2013. p. 1–8.
[7] Nielsen UD. Estimation of directional wave spectra from measured ship. Tech-
nical University of Denmark; 2005 [Ph. D. thesis].
[8] Ewans KC, Vanderschuren L, Tromans PS. FPSO conference, estimating wind-
sea and swell for FPSO operability. J Offshore Mech Arct Eng 2006;128(4):314,
http://dx.doi.org/10.1115/1.2166653 http://offshoremechanics.
asmedigitalcollection.asme.org/article.aspx?articleid=1456271.
[9] Portilla J, Ocampo-Torres FJ, Monbaliu J. Spectral partitioning and identifi-
cation of wind sea and swell. J Atmos Ocean Technol 2009;26(1):107–22,
http://dx.doi.org/10.1175/2008JTECHO609.1.
[10] Guedes Soares C. Representation of double-peaked sea wave spectra. Ocean
Eng 1984;11(2):185–207, http://dx.doi.org/10.1016/0029-8018(84)90019-2
http://linkinghub.elsevier.com/retrieve/pii/0029801884900192.
[11] Ewans KC, Bitner-Gregersen E, Guedes Soares C. Estimation of
wind-sea and swell components in a bimodal sea state. J Offshore
Mech Arct Eng 2006;128(4):265, http://dx.doi.org/10.1115/1.2166655
http://offshoremechanics.asmedigitalcollection.asme.org/article.
aspx?articleid=1456243.
[12] Boukhanovsky AV, Guedes Soares C. Modelling of multipeaked
directional wave spectra. Appl Ocean Res 2009;31(2):132–41,
http://dx.doi.org/10.1016/j.apor.2009.06.001 http://linkinghub.elsevier.
com/retrieve/pii/S0141118709000455.
[13] Hanson JL, Phillips OM. Automated analysis of ocean surface direc-
tional wave spectra. J Atmos Ocean Technol 2001;18(2):277–93,
http://dx.doi.org/10.1175/1520-0426(2001)018<0277:AAOOSD>2.0.CO;2.
[14] Nielsen UD, Jensen JJ, Pedersen P, Ito Y. Onboard monitoring of fatigue damage
rates in the hull girder. Mar Struct 2011;24:182–206.
[15] Guedes Soares C. Effect of transfer function uncertainty on short-term ship
responses. Ocean Eng 1991;18(4):329–62.
[16] Pascoal R, Guedes Soares C. Kalman filtering of vessel motions for ocean
wave directional spectrum estimation. Ocean Eng 2009;36(6–7):477–88,
http://dx.doi.org/10.1016/j.oceaneng.2009.01.013 http://linkinghub.elsevier.
com/retrieve/pii/S0029801809000183.
[17] Jensen JJ, Capul J. Extreme response predictions for jack-up units in sec-
ond order stochastic waves by FORM. Prob Eng Mech 2006;21(4):330–7,
http://dx.doi.org/10.1016/j.probengmech.2005.11.007 http://linkinghub.
elsevier.com/retrieve/pii/S0266892005000834.
[18] Jensen JJ. Load and global response of ships. Elsevier Ocean Engineering Book
Series; 2001.
[19] Rodriguez G, Guedes Soares C, Pacheco M, Perez-Martell E. Wave height dis-
tribution in mixed sea states. J Offshore Mech Arct Eng 2002;124(1):34,
http://dx.doi.org/10.1115/1.1445794 http://offshoremechanics.
asmedigitalcollection.asme.org/article.aspx?articleid=1475388.
[20] Tutorial for WAFO version 2.5, no. March. Lund University; 2011.
[21] Simos AN, Sparano JV, Tannuri EA, Matos VLF. Directional wave spectrum esti-
mation based on a vessel 1 st order motions: field results. In: Proceedings of the
seventeenth international offshore and polar engineering conference, ISOPE.
2007. p. 1938–44.
[22] Tannuri EA, Simos AN, Sparano JV, Matos VLF. Motion-based wave
estimation: small-scale tests with a crane-barge model. Mar Struct
2012;28(1):67–85, http://dx.doi.org/10.1016/j.marstruc.2012.05.002 http://
linkinghub.elsevier.com/retrieve/pii/S0951833912000366.
[23] Montazeri N, Nielsen UD, Jensen JJ. Trend modelling of wave parameters and
application in onboard prediction of ship responses. In: Proc. of MTS/IEEE
OCEANS 15. 2015.