27.06.2013 Views

Estimations of on-site directional wave spectra from measured ship ...

Estimations of on-site directional wave spectra from measured ship ...

Estimations of on-site directional wave spectra from measured ship ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Marine Structures 19 (2006) 33–69<br />

<str<strong>on</strong>g>Estimati<strong>on</strong>s</str<strong>on</strong>g> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>on</strong>-<strong>site</strong> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> <strong>from</strong><br />

<strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses<br />

Abstract<br />

ARTICLE IN PRESS<br />

Ulrik Dam Nielsen<br />

www.elsevier.com/locate/marstruc<br />

Department <str<strong>on</strong>g>of</str<strong>on</strong>g> Mechanical Engineering, Technical University <str<strong>on</strong>g>of</str<strong>on</strong>g> Denmark, DK-2800 Lyngby, Denmark<br />

Received 21 October 2005; received in revised form 29 May 2006; accepted 26 June 2006<br />

In general, two main c<strong>on</strong>cepts can be applied to estimate the <strong>on</strong>-<strong>site</strong> directi<strong>on</strong>al <strong>wave</strong> spectrum <strong>on</strong><br />

the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong> resp<strong>on</strong>se measurements: (1) a parametric method which assumes the <strong>wave</strong> spectrum<br />

to be composed by parameterised <strong>wave</strong> <strong>spectra</strong>; or (2) a n<strong>on</strong>-parametric method where the directi<strong>on</strong>al<br />

<strong>wave</strong> spectrum is found directly as the values in a completely discretised frequency-directi<strong>on</strong>al<br />

domain without a priori assumpti<strong>on</strong>s <strong>on</strong> the spectrum. The paper outlines the theory <str<strong>on</strong>g>of</str<strong>on</strong>g> these two<br />

c<strong>on</strong>cepts, and it is shown how to deal with the speed-<str<strong>on</strong>g>of</str<strong>on</strong>g>-advance problem for operating <strong>ship</strong>s. In<br />

additi<strong>on</strong>, the methods include an equivalence <str<strong>on</strong>g>of</str<strong>on</strong>g> energy in the governing equati<strong>on</strong>s and, as regards<br />

the parametric c<strong>on</strong>cept, a frequency-dependent spreading <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong>s is introduced.<br />

The paper includes an extensive analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> full-scale measurements for which the directi<strong>on</strong>al <strong>wave</strong><br />

<strong>spectra</strong> are estimated by the two <strong>ship</strong> resp<strong>on</strong>se-based methods. Hence, comparis<strong>on</strong>s are made<br />

between these estimates and, moreover, the agreement with the corresp<strong>on</strong>ding directi<strong>on</strong>al <strong>wave</strong><br />

<strong>spectra</strong> produced by the <strong>wave</strong> radar system WAVEX is studied. The agreement between the two<br />

methods is reas<strong>on</strong>able, as well is the agreement between the results <str<strong>on</strong>g>of</str<strong>on</strong>g> these methods and those <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

WAVEX. It is difficult to propose <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong> resp<strong>on</strong>se-based methods in favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the other,<br />

since they perform equally well.<br />

r 2006 Elsevier Ltd. All rights reserved.<br />

Keywords: Measured <strong>ship</strong> resp<strong>on</strong>ses; Resp<strong>on</strong>se <strong>spectra</strong>; Directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong>; Bayesian modelling; Parametric<br />

modelling; Full-scale data; Wave radar<br />

Fax: +45 4588 4325.<br />

E-mail address: udn@mek.dtu.dk.<br />

0951-8339/$ - see fr<strong>on</strong>t matter r 2006 Elsevier Ltd. All rights reserved.<br />

doi:10.1016/j.marstruc.2006.06.001


34<br />

1. Introducti<strong>on</strong><br />

1.1. Motivati<strong>on</strong><br />

The operati<strong>on</strong>al safety <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong>s can be increased by use <str<strong>on</strong>g>of</str<strong>on</strong>g> in-service m<strong>on</strong>itoring systems.<br />

For such systems to be used efficiently, the <strong>on</strong>-<strong>site</strong> directi<strong>on</strong>al <strong>wave</strong> spectrum needs to be<br />

estimated and updated c<strong>on</strong>tinuously. Today, means <str<strong>on</strong>g>of</str<strong>on</strong>g> obtaining estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

directi<strong>on</strong>al <strong>wave</strong> spectrum exist. Such means include moored <strong>wave</strong> rider buoys and current<br />

meters, satellite measurements and <strong>wave</strong> radar systems. The latter two <str<strong>on</strong>g>of</str<strong>on</strong>g> these means do<br />

not suffer <strong>from</strong> the problems related to the fixed positi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a moored buoy or current<br />

meters, but do, <strong>on</strong> the other hand, require complex computati<strong>on</strong>al hardware and have a<br />

high initial cost, cf. Tannuri et al. [1], not to menti<strong>on</strong> calibrati<strong>on</strong> and maintenance. For<br />

this reas<strong>on</strong> it is <str<strong>on</strong>g>of</str<strong>on</strong>g> interest to be able to estimate the <strong>wave</strong> spectrum <strong>from</strong> <strong>measured</strong><br />

<strong>ship</strong> resp<strong>on</strong>ses, which are accessible <strong>from</strong> the sensor measurements d<strong>on</strong>e in in-service<br />

m<strong>on</strong>itoring system.<br />

1.2. Existing work<br />

In the literature there exist several papers which deal with the estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> (directi<strong>on</strong>al)<br />

<strong>wave</strong> <strong>spectra</strong> <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses. In general, these works can be split<br />

in two groups or as being part <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> two c<strong>on</strong>cepts: (1) a c<strong>on</strong>cept where the <strong>wave</strong><br />

spectrum is estimated <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> parametric modelling, e.g. Hua and Palmquist [2],<br />

Tannuri et al. [1] and the EC project HullM<strong>on</strong>þ [3]; or (2) a c<strong>on</strong>cept where the estimati<strong>on</strong><br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> spectrum is based <strong>on</strong> n<strong>on</strong>-parametric modelling, e.g. Iseki and Ohtsu [4], Iseki<br />

and Terada [5], Waals et al. [6], Isobe et al. [7] and Nielsen [8]. Fundamentally, the two<br />

c<strong>on</strong>cepts are similar in the sense that both methods in their foundati<strong>on</strong>s use linear <strong>spectra</strong>l<br />

analysis to set up c<strong>on</strong>diti<strong>on</strong>s/equati<strong>on</strong>s which relate the <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses—the <strong>on</strong>e<br />

hand side—with the <strong>wave</strong> energy spectrum through complex-valued transfer functi<strong>on</strong>s—<br />

the other hand side. Thus, <strong>from</strong> this relati<strong>on</strong>, the principle is to minimise, in the least<br />

squares sense, the difference between the two sides, see Fig. 1.<br />

Although Tannuri et al. [1], including Pascoal et al. [9], to some degree compare the two<br />

estimati<strong>on</strong> c<strong>on</strong>cepts with each other, the comparis<strong>on</strong>s apply for numerical simulati<strong>on</strong>s and<br />

for a <strong>ship</strong> not being underway, since these works do not c<strong>on</strong>sider the speed-<str<strong>on</strong>g>of</str<strong>on</strong>g>-advance<br />

problem in the derived theory. Similarly, Benoit and Goasguen [10] compares different<br />

directi<strong>on</strong>al <strong>wave</strong> analysis methods for a so-called ‘‘single-point’’ measuring system, which<br />

record data <strong>from</strong> a three-displacements buoy, that is located at a fixed positi<strong>on</strong>.<br />

The speed-<str<strong>on</strong>g>of</str<strong>on</strong>g>-advance problem, governed by the deep-water relati<strong>on</strong><strong>ship</strong> between the<br />

encounter and the <strong>wave</strong> frequency, leads in certain cases to the so-called triple-valued<br />

functi<strong>on</strong> problem in beam, quartering and following seas, cf. Fig. 3(b). This problem needs<br />

to be strictly incorporated in an estimati<strong>on</strong> procedure based <strong>on</strong> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses.<br />

In Iseki and Ohtsu [4] the elementary problem is dealt with and in that paper it is suggested<br />

how to incorporate it.<br />

1.3. Objectives<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

In the present paper, the theory <str<strong>on</strong>g>of</str<strong>on</strong>g> a parametric modelling procedure and a n<strong>on</strong>parametric<br />

modelling procedure is derived, and it is outlined in detail how to take into


account a vessel being underway, independently if the estimati<strong>on</strong> is performed by<br />

parametric or n<strong>on</strong>-parametric modelling. Furthermore, it is shown how to include<br />

additi<strong>on</strong>al equati<strong>on</strong>s, based <strong>on</strong> Iseki and Terada [5], in both <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimati<strong>on</strong> c<strong>on</strong>cepts,<br />

so that it is sought to secure the c<strong>on</strong>servati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> energy in the resp<strong>on</strong>ses. The main<br />

part <str<strong>on</strong>g>of</str<strong>on</strong>g> the present paper is devoted to the analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> numerical and full-scale data. Thus,<br />

<strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> simulated and <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses, the parametric modelling and<br />

the n<strong>on</strong>-parametric modelling procedure are compared. As regards the analysis <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

full-scale data, comparis<strong>on</strong>s with data <strong>from</strong> the <strong>wave</strong> radar system WAVEX have also<br />

been c<strong>on</strong>ducted.<br />

1.4. Resp<strong>on</strong>ses<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 35<br />

Fig. 1. The fundamental idea in the estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> <strong>spectra</strong> based <strong>on</strong> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses. EC project<br />

HullM<strong>on</strong>+ [3].<br />

In principle, any type <str<strong>on</strong>g>of</str<strong>on</strong>g> resp<strong>on</strong>se signal, obtained as time series, may be utilised in the<br />

<strong>wave</strong> spectrum estimati<strong>on</strong> as l<strong>on</strong>g as a linear complex-valued transfer functi<strong>on</strong> exists, that<br />

is, can be calculated. However, it is believed that the highest rate <str<strong>on</strong>g>of</str<strong>on</strong>g> success is achieved<br />

when global <strong>ship</strong> resp<strong>on</strong>ses, such as pitch, roll, vertically <strong>wave</strong>-induced bending moment,<br />

etc., are used. This means that resp<strong>on</strong>se signals <str<strong>on</strong>g>of</str<strong>on</strong>g> e.g. pressure transducers in<br />

the hull should not be c<strong>on</strong>sidered in the estimati<strong>on</strong>, as local effects may affect such a<br />

resp<strong>on</strong>se. It is important to emphasise that at least <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong> resp<strong>on</strong>ses must have<br />

port/starboard asymmetry, otherwise the modelling procedures cannot differentiate<br />

between port and starboard entering <strong>wave</strong>s. This explains also the need <str<strong>on</strong>g>of</str<strong>on</strong>g> complexvalued<br />

transfer functi<strong>on</strong>s, so that the amplitudes as well as the phase angles <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

resp<strong>on</strong>ses are obtained.<br />

Specifics <strong>on</strong> resp<strong>on</strong>se signals to be used for the estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> <strong>spectra</strong> can be found<br />

in, am<strong>on</strong>g others, HullM<strong>on</strong>+ [3] and Nielsen [8]. Tannuri et al. [1], however, brings<br />

an interesting issue to questi<strong>on</strong> and, thus, argues correctly that the roll moti<strong>on</strong>, in<br />

general, has a n<strong>on</strong>-linear and res<strong>on</strong>ant behaviour and, furthermore, that the roll<br />

moti<strong>on</strong> is extremely sensitive to load variati<strong>on</strong>s. Tannuri et al. [1] therefore proposes to<br />

use the sway resp<strong>on</strong>se instead <str<strong>on</strong>g>of</str<strong>on</strong>g> the roll, as the frequency resp<strong>on</strong>se functi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> sway is less


36<br />

sensitive to loading c<strong>on</strong>diti<strong>on</strong>s. The present work deals primarily with results based <strong>on</strong> the<br />

three resp<strong>on</strong>ses {heave, roll, pitch}. Though, estimati<strong>on</strong>s have also been carried out with<br />

the roll resp<strong>on</strong>se replaced by the sway. Later, in Secti<strong>on</strong> 7, more comments are given <strong>on</strong><br />

this topic.<br />

1.5. Organisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the paper<br />

The paper is organised as follows: In Secti<strong>on</strong>s 2–5 theoretical aspects <str<strong>on</strong>g>of</str<strong>on</strong>g> the modelling<br />

procedures are treated. Secti<strong>on</strong> 6 deals with a numerical example which verifies<br />

and compares the modelling procedures. Then in Secti<strong>on</strong> 7, full-scale data is analysed<br />

and comparis<strong>on</strong>s between the estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the two modelling procedures and<br />

those <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX are carried out. Finally, in Secti<strong>on</strong> 8, the major c<strong>on</strong>clusi<strong>on</strong>s are<br />

formulated.<br />

2. General theory<br />

In this secti<strong>on</strong> the general theory <str<strong>on</strong>g>of</str<strong>on</strong>g> the parametric and the n<strong>on</strong>-parametric modelling<br />

procedure are described. Hence, the secti<strong>on</strong> deals with facts and c<strong>on</strong>diti<strong>on</strong>s which are<br />

shared by the two c<strong>on</strong>cepts and, subsequently, two secti<strong>on</strong>s yield the specifics for the<br />

individual c<strong>on</strong>cepts.<br />

2.1. Fundamentals<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

From time series <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses, the resp<strong>on</strong>se cross <strong>spectra</strong>, SijðoeÞ, can be<br />

obtained by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> fast Fourier transformati<strong>on</strong> (FFT) with appropriate smoothing<br />

or by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> multivariate autoregressive (MAR) modelling, e.g. Akaike and<br />

Nakagawa [11] and Nielsen [8]. Fig. 2 illustrates an example (data <strong>from</strong> [12]) <str<strong>on</strong>g>of</str<strong>on</strong>g>the<br />

outcome <str<strong>on</strong>g>of</str<strong>on</strong>g> a cross <strong>spectra</strong>l analysis c<strong>on</strong>ducted <strong>on</strong> three (simultaneously) <strong>measured</strong> <strong>ship</strong><br />

resp<strong>on</strong>ses. The analysis is carried out by the MAR modelling and it should be noted that<br />

<strong>on</strong>ly the diag<strong>on</strong>al <strong>spectra</strong> are real-valued. The <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> are shown as<br />

functi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the encounter frequency oe. In this study, the <strong>ship</strong> resp<strong>on</strong>ses are taken to be<br />

stati<strong>on</strong>ary. This means that the encounter angle, b, and other operati<strong>on</strong>al parameters are<br />

assumed to be c<strong>on</strong>stant during the time which forms the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>measured</strong> resp<strong>on</strong>se<br />

<strong>spectra</strong>.<br />

The encounter angle b is defined as the heading <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong> relative to the <strong>wave</strong>s and<br />

b ¼ p corresp<strong>on</strong>ds to head <strong>wave</strong>s. From Fig. 3(a) it is seen that when the <strong>ship</strong> course a and<br />

the <strong>wave</strong> directi<strong>on</strong> y are <strong>measured</strong> relative to some fixed directi<strong>on</strong>, the encounter angle is<br />

given by<br />

b ¼ y a. (2.1)<br />

For c<strong>on</strong>venience, the fixed directi<strong>on</strong> in Fig. 3(a) is taken to be coincident to the <strong>ship</strong> course<br />

so that the <strong>wave</strong> directi<strong>on</strong> is given relative to the <strong>ship</strong> course, that is, b ¼ y. This means<br />

that the <strong>wave</strong> directi<strong>on</strong> is coincident to the encounter angle.<br />

On the assumpti<strong>on</strong> that the <strong>ship</strong> resp<strong>on</strong>ses are stati<strong>on</strong>ary and linear with the incident<br />

<strong>wave</strong>s, the complex-valued transfer functi<strong>on</strong>s Fiðoe; bÞ and Fjðoe; bÞ for the ith and jth<br />

resp<strong>on</strong>ses yield the theoretical relati<strong>on</strong><strong>ship</strong> between the ith and the jth comp<strong>on</strong>ents <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

cross <strong>spectra</strong> SijðoeÞ and the directi<strong>on</strong>al <strong>wave</strong> spectrum Eðoe; bÞ through the following


Power<br />

Power<br />

Power<br />

6<br />

4<br />

2<br />

x 10 -3<br />

0<br />

0 0.2 0.4<br />

x 10-3<br />

ARTICLE IN PRESS<br />

pitch/pitch x 10-3<br />

2<br />

pitch/roll<br />

0.02<br />

pitch/ver.acc.<br />

1<br />

real<br />

imag<br />

0<br />

0<br />

0.02<br />

roll/pitch<br />

2<br />

1<br />

0<br />

-1<br />

-2<br />

-3<br />

0 0.2 0.4<br />

ver.acc./pitch<br />

0.02<br />

0<br />

-0.02<br />

-0.04<br />

-0.06<br />

-0.08<br />

0 0.2<br />

Frequency [Hz]<br />

0.4<br />

-1<br />

-2<br />

-3<br />

0 0.2 0.4<br />

5<br />

4<br />

3<br />

2<br />

1<br />

x 10-3<br />

roll/roll<br />

0<br />

0 0.2<br />

ver.acc./roll<br />

0.4<br />

0.02<br />

0.01<br />

0<br />

-0.01<br />

-0.02<br />

0 0.2<br />

Frequency [Hz]<br />

0.4<br />

0.04<br />

0.06<br />

0.08<br />

0<br />

0.02<br />

0.01<br />

0<br />

0.01<br />

roll/ver.acc.<br />

0.02<br />

0 0.2<br />

ver.acc./ver.acc<br />

1<br />

0.5<br />

0.2<br />

0<br />

0 0.2<br />

Frequency [Hz]<br />

Fig. 2. Cross <strong>spectra</strong>l analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> three <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses: {pitch, roll, vertical accelerati<strong>on</strong> at FP}. Data<br />

<strong>from</strong> [12].<br />

θ<br />

α<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 37<br />

(a) (b)<br />

=30°<br />

following<br />

=0°<br />

=60°<br />

quatering<br />

beam<br />

=120°<br />

bow<br />

=150°<br />

head<br />

=180°<br />

Fig. 3. (a) Ship course, a, and <strong>wave</strong> directi<strong>on</strong>, y, relative to a fixed directi<strong>on</strong>. (b) Definiti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> terms as regards the<br />

encounter angles. Note that b is shown in degrees.<br />

0.4<br />

0.4<br />

0.4


38<br />

integral equati<strong>on</strong>, e.g. Bhattacharyya [13],<br />

SijðoeÞ ¼<br />

Z p<br />

p<br />

Fiðoe; bÞFjðoe; bÞEðoe; bÞ db (2.2)<br />

with denoting the complex c<strong>on</strong>jugate. It should be noted that the transfer functi<strong>on</strong>s are<br />

written as functi<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>on</strong>ly the encounter angle and the encounter frequency, since the<br />

implicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> changing operati<strong>on</strong>al parameters is understood.<br />

The <strong>wave</strong> spectrum is advantageously estimated in the <strong>wave</strong> frequency domain<br />

for which reas<strong>on</strong> the right-hand side <str<strong>on</strong>g>of</str<strong>on</strong>g> (2.2) is transformed into the <strong>wave</strong> frequency,<br />

o, by introducing the deep-water relati<strong>on</strong><strong>ship</strong> between the encounter and the <strong>wave</strong><br />

frequency,<br />

oe ¼ o o 2 A; A ¼ V<br />

cos b (2.3)<br />

g<br />

with V being the <strong>ship</strong> speed and g the accelerati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> gravity. In beam, quartering<br />

and following seas, cf. Fig. 3(b), (2.3) leads to up to three <strong>wave</strong> frequencies for<br />

certain speeds and, hence, the integral in (2.2) must be split apart in order to do the<br />

transformati<strong>on</strong>.<br />

2.2. Discretisati<strong>on</strong><br />

K k=1<br />

m=1<br />

2<br />

3<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

4<br />

By discretising the <strong>wave</strong> field in K M points, with K being the number <str<strong>on</strong>g>of</str<strong>on</strong>g> encounter<br />

angles, i.e. <strong>wave</strong> directi<strong>on</strong>s, and M being the number <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> frequencies, the directi<strong>on</strong>al<br />

<strong>wave</strong> spectrum is thus to be estimated in K M points, see Fig. 4 which illustrates the<br />

discretisati<strong>on</strong> in a polar format. Specifically, the discretisati<strong>on</strong> means that the <strong>wave</strong><br />

spectrum must be estimated in a set <str<strong>on</strong>g>of</str<strong>on</strong>g> prespecified <strong>wave</strong> frequencies. Hence, the <strong>wave</strong><br />

frequency as found by (2.3) is taken to be a weighted distributi<strong>on</strong> between the two closest<br />

specified <strong>wave</strong> frequencies, cf. Fig. 4. C<strong>on</strong>sequently, the transfer functi<strong>on</strong> in (2.2) for the<br />

calculated <strong>wave</strong> frequency o01 is found by linear interpolati<strong>on</strong><br />

-1<br />

Fðo01Þ ¼ e b FðomÞþea Fðomþ1Þ, (2.4)<br />

where the weights are e b ¼ b=ða þ bÞ and ea ¼ a=ða þ bÞ.<br />

2<br />

M<br />

3<br />

5<br />

4<br />

<br />

01<br />

a b<br />

Fig. 4. Discretisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> field, and weighted distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a calculated <strong>wave</strong> frequency between two<br />

specified frequencies.<br />

+1


From the above c<strong>on</strong>siderati<strong>on</strong>s, the fundamental equati<strong>on</strong>, (2.2), can be written in the<br />

following discretised versi<strong>on</strong><br />

SijðoeÞ ¼Db XK<br />

ebFiðom; bkÞFjðom; bkÞEðom; bkÞ k¼1<br />

dom<br />

"<br />

doe b¼bk þeaFiðomþ1; bkÞFjðomþ1; bkÞEðomþ1; bkÞ domþ1<br />

#<br />

ð2:5Þ<br />

doe b¼bk with<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 39<br />

mðoe; bkÞ; m 2½1; MŠ,<br />

k ¼ 1 : K. ð2:6Þ<br />

The calculati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>se cross <strong>spectra</strong> reveals that SijðoeÞ is Hermitian, i.e. Sij ¼ Sji,<br />

see also Fig. 2. Hence, by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> matrix notati<strong>on</strong>, (2.5) is transformed into a<br />

multivariate model expressi<strong>on</strong>, e.g. Iseki and Terada [5] and Nielsen [8],<br />

b ¼ Af. (2.7)<br />

In the event <str<strong>on</strong>g>of</str<strong>on</strong>g> N <strong>ship</strong> resp<strong>on</strong>ses being c<strong>on</strong>sidered, that is, i ¼ j ¼f1; 2; ...; Ng, the<br />

resp<strong>on</strong>se vector, b, is given by<br />

2 3<br />

SiiðoeÞ<br />

6<br />

. 7<br />

6 7<br />

6 . 7<br />

6 7<br />

6 Re½SijðoeÞŠ 7<br />

6 7<br />

b ¼ 6 .<br />

7<br />

7;<br />

sizeðbÞ ¼N<br />

6 . 7<br />

6 7<br />

6 Im½SijðoeÞŠ 7<br />

4 5<br />

.<br />

2<br />

1. (2.8)<br />

The <strong>wave</strong> spectrum is discretised as illustrated in Fig. 4 and the sequence, as regards<br />

the <strong>wave</strong> frequency and the encounter angle, in which the spectrum is evaluated follows<br />

<strong>from</strong><br />

2<br />

3<br />

Eðo1; b1Þ 6<br />

Eðo1; b<br />

7<br />

6<br />

2Þ 7<br />

6<br />

7<br />

6 . 7<br />

6 . 7<br />

6<br />

7<br />

6<br />

Eðo1; b<br />

7<br />

6<br />

KÞ 7<br />

6<br />

7<br />

6<br />

f ¼ Eðo2; b1Þ 7<br />

6<br />

7;<br />

6 . 7<br />

6 . 7<br />

6 .<br />

7<br />

6 . 7<br />

6<br />

7<br />

6<br />

4 EðoM; b 7<br />

K 1Þ 5<br />

EðoM; bKÞ sizeðfÞ ¼ðK MÞ 1. (2.9)


40<br />

The coefficient matrix, A, is thus determined according to<br />

2<br />

6<br />

A ¼ Db 6<br />

4<br />

2 dom<br />

jFiðom; bkÞj doe b¼bk .<br />

.<br />

ReðFiðom; bkÞFjðom; bkÞÞ dom<br />

doe b¼bk .<br />

ImðFiðom; bkÞFjðom; bkÞÞ dom<br />

doe b¼bk .<br />

3<br />

7<br />

7;<br />

sizeðAÞ ¼N<br />

7<br />

5<br />

2<br />

ðK MÞ.<br />

(2.10)<br />

As regards A, the elements must be calculated precisely in the same order as the <strong>wave</strong><br />

spectrum, so that the correct relati<strong>on</strong> between encounter frequencies and <strong>wave</strong> frequencies<br />

are secured. It should be noted that, although not shown in (2.10), the weights ea and e b need<br />

to be included in the set-up <str<strong>on</strong>g>of</str<strong>on</strong>g> the coefficient matrix according to (2.4). Moreover, it should<br />

be realised that A may c<strong>on</strong>tain a large amount <str<strong>on</strong>g>of</str<strong>on</strong>g> zero elements, e.g. Nielsen [8].<br />

So far, <strong>on</strong>ly <strong>on</strong>e encounter frequency has been c<strong>on</strong>sidered. In order to include an interval<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> encounter frequencies discretised as<br />

oe ¼ l Doe; l ¼ 1; 2; 3; ...; L, (2.11)<br />

the fundamental equati<strong>on</strong> can readily be extended by introducing the following resp<strong>on</strong>se<br />

block-vector<br />

eb ¼½b1 b2 ... bLŠ T , (2.12)<br />

where bl ¼ bðl DoeÞ and with bl<br />

sizeðebÞ ¼ðN<br />

determined <strong>from</strong> (2.8). It should be noted that<br />

2 LÞ 1.<br />

The coefficient matrix follows corresp<strong>on</strong>dingly <strong>from</strong> the block-matrix<br />

2 3<br />

A1<br />

6 A2<br />

7<br />

6 7<br />

eA ¼ 6 . 7<br />

4 . 5 ; sizeðe AÞ¼ðN 2 LÞ ðK MÞ (2.13)<br />

AL<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

so that the multivariate model expressi<strong>on</strong> reads<br />

eb ¼ e Af. (2.14)<br />

As menti<strong>on</strong>ed, e A is likely to c<strong>on</strong>tain a large amount <str<strong>on</strong>g>of</str<strong>on</strong>g> zero-elements. This is illustrated in<br />

Fig. 5 which shows an arbitrary example <str<strong>on</strong>g>of</str<strong>on</strong>g> the coefficient matrix e A for a certain speed and<br />

discretisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> encounter angles and frequencies.


2.3. Equivalence <str<strong>on</strong>g>of</str<strong>on</strong>g> energy<br />

0<br />

100<br />

200<br />

300<br />

400<br />

500<br />

600<br />

In additi<strong>on</strong> to the equati<strong>on</strong>s given by (2.14), N equati<strong>on</strong>s can further be specified, where<br />

N, accordingly, is the number <str<strong>on</strong>g>of</str<strong>on</strong>g> c<strong>on</strong>sidered resp<strong>on</strong>se signals. Thus, as menti<strong>on</strong>ed by Iseki<br />

and Terada [5], it can be taken into account that the amount <str<strong>on</strong>g>of</str<strong>on</strong>g> energy in the signals should<br />

be equivalent independently <str<strong>on</strong>g>of</str<strong>on</strong>g> the energy being calculated <strong>from</strong> the <strong>measured</strong> cross<br />

<strong>spectra</strong>, i.e. the left-hand side <str<strong>on</strong>g>of</str<strong>on</strong>g> (2.2), or the theoretically estimated cross <strong>spectra</strong>, i.e. the<br />

right-hand side <str<strong>on</strong>g>of</str<strong>on</strong>g> (2.2). Mathematically, it can be argued that N2 additi<strong>on</strong>al equati<strong>on</strong>s<br />

may actually be specified, since the cross <strong>spectra</strong>l analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> N resp<strong>on</strong>ses, as menti<strong>on</strong>ed,<br />

leads to a total number <str<strong>on</strong>g>of</str<strong>on</strong>g> N2 different real and imaginary <strong>spectra</strong>. However, the complexvalued<br />

cross <strong>spectra</strong>, that is, the real and the imaginary parts <str<strong>on</strong>g>of</str<strong>on</strong>g> the <str<strong>on</strong>g>of</str<strong>on</strong>g>f-diag<strong>on</strong>al cross<br />

<strong>spectra</strong> can have both positive and negative values which means that the area under the<br />

respective curve can be close to zero in some cases. For this reas<strong>on</strong>, the diag<strong>on</strong>al <strong>spectra</strong><br />

are <strong>on</strong>ly c<strong>on</strong>sidered and, hence, N c<strong>on</strong>diti<strong>on</strong>s/equati<strong>on</strong>s are specified to secure the<br />

equivalence <str<strong>on</strong>g>of</str<strong>on</strong>g> energy between the <strong>measured</strong> and the calculated resp<strong>on</strong>ses. C<strong>on</strong>sequently,<br />

the following equati<strong>on</strong>s are also taken into account<br />

Z<br />

Z Z p<br />

SiiðoeÞ doe ¼ jFiðo; bÞj 2 Eðo; bÞ db do; i ¼ 1; ...; N, (2.15)<br />

p<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 41<br />

0 100 200 300 400 500 600 700 800 900<br />

nz = 22509<br />

Fig. 5. Example <str<strong>on</strong>g>of</str<strong>on</strong>g> n<strong>on</strong>-zero elements in the coefficient matrix. ‘nz’ is the total number <str<strong>on</strong>g>of</str<strong>on</strong>g> n<strong>on</strong>-zero elements; the<br />

matrix has a dimensi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> 657 900.<br />

where it should be noted that the right-hand side is based <strong>on</strong> the <strong>wave</strong> frequency.<br />

From a physical point <str<strong>on</strong>g>of</str<strong>on</strong>g> view, the equivalence <str<strong>on</strong>g>of</str<strong>on</strong>g> energy <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>measured</strong> and the<br />

estimated <strong>spectra</strong> is important. In the equati<strong>on</strong> system (2.14), the equivalence, which is<br />

expressed by (2.15), may therefore be weighted relatively heavily.


42<br />

2.4. A least squares problem<br />

In general, the directi<strong>on</strong>al <strong>wave</strong> spectrum, f ¼ Eðom; b kÞ can be sought in the least<br />

squares sense by solving<br />

min w 2 minke Af ebk 2<br />

(2.16)<br />

with w ¼k kbeing the L2 norm. Specifically, the soluti<strong>on</strong> is obtained by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

<strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> two main c<strong>on</strong>cepts: (1) a n<strong>on</strong>-parametric modelling c<strong>on</strong>cept denoted the Bayesian<br />

method or (2) a parametric modelling c<strong>on</strong>cept denoted the Parametric method. In the<br />

following, these c<strong>on</strong>cepts are dealt with. As regards the n<strong>on</strong>-parametric modelling c<strong>on</strong>cept,<br />

it should be noted that in the present work the term Bayesian is used to characterise this<br />

c<strong>on</strong>cept, however, inverse theory, e.g. Dimri [14] and Tarantola [15], could also have been<br />

used as the term, since inverse theory and Bayesian modelling are basically two sides <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

same coin, e.g. Nielsen [8]. In this relati<strong>on</strong> reference may also be given to Herbers and<br />

Guza [16].<br />

3. Bayesian modelling<br />

Without explicit assumpti<strong>on</strong>s <strong>on</strong> the directi<strong>on</strong>al <strong>wave</strong> spectrum, (2.16) expresses ðN 2<br />

L þ NÞ equati<strong>on</strong>s <strong>from</strong> which K M unknowns, Eðom; b kÞ, are to be solved. For a<br />

reas<strong>on</strong>able discretisati<strong>on</strong>, say, K ¼ 18, M ¼ 30 and taking N ¼ 3 <strong>ship</strong> resp<strong>on</strong>ses to form<br />

the basis for the estimati<strong>on</strong>, (2.16) is in general underdetermined. The number, L, <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

encounter frequencies plays a role in the sense that more equati<strong>on</strong>s are established by<br />

increasing L. However, new informati<strong>on</strong> put into the system in this way is <str<strong>on</strong>g>of</str<strong>on</strong>g> limited use<br />

due to degenerati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the equati<strong>on</strong> system. Therefore, additi<strong>on</strong>al equati<strong>on</strong>s must be<br />

established by use <str<strong>on</strong>g>of</str<strong>on</strong>g> some kind <str<strong>on</strong>g>of</str<strong>on</strong>g> prior informati<strong>on</strong>. In the present model two sets <str<strong>on</strong>g>of</str<strong>on</strong>g> prior<br />

informati<strong>on</strong> are introduced. These are: (1) the <strong>wave</strong> spectrum is assumed to be smoothly<br />

changing with both frequency and directi<strong>on</strong>, and (2) the <strong>wave</strong> spectrum is expected to have<br />

negligible values for very low and high frequencies. Moreover, if at hand, informati<strong>on</strong> <strong>on</strong><br />

the <strong>wave</strong> directi<strong>on</strong> can, theoretically, be included.<br />

3.1. Basic assumpti<strong>on</strong>s<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

In additi<strong>on</strong>, and beforehand <str<strong>on</strong>g>of</str<strong>on</strong>g> the introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> prior informati<strong>on</strong>, two basic<br />

assumpti<strong>on</strong>s are introduced to make the Bayesian modelling work. Firstly, as discussed by<br />

Iseki and Ohtsu [4], the Bayesian modelling introduces a stochastic point <str<strong>on</strong>g>of</str<strong>on</strong>g> view as regards<br />

the estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> <strong>spectra</strong>. Thus, in order to facilitate the Bayesian modelling<br />

procedure, cf. Akaike [17], the difference between the left- and the right-hand side <str<strong>on</strong>g>of</str<strong>on</strong>g> (2.14)<br />

is taken as a white noise sequence vector w with zero mean and variance s2 . Sec<strong>on</strong>dly, to<br />

avoid negative <strong>spectra</strong>l estimates, a n<strong>on</strong>-negativity c<strong>on</strong>straint is applied to the <strong>wave</strong><br />

spectrum by use <str<strong>on</strong>g>of</str<strong>on</strong>g> a coordinate transformati<strong>on</strong><br />

Eðom; bkÞ¼expðxk;mÞ, (3.1)<br />

where it should be noted that the order <str<strong>on</strong>g>of</str<strong>on</strong>g> the indices has been changed.<br />

Hence, the multivariate model expressi<strong>on</strong> (2.14) is written as<br />

eb ¼ e A expðxÞþw. (3.2)


It is important to realise that the introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the n<strong>on</strong>-negativity c<strong>on</strong>straint leads to a<br />

n<strong>on</strong>-linear problem. This is undesirable for which reas<strong>on</strong> a linearisati<strong>on</strong> should be applied<br />

to the coordinate transformati<strong>on</strong>, e.g. Iseki and Terada [5] and Nielsen [8].<br />

3.2. Prior informati<strong>on</strong><br />

As regards prior informati<strong>on</strong> (1), the c<strong>on</strong>diti<strong>on</strong> is set up by requiring the sec<strong>on</strong>d order<br />

derivative <str<strong>on</strong>g>of</str<strong>on</strong>g> lnðEðo; bÞÞ to be minimum. Hence, by use <str<strong>on</strong>g>of</str<strong>on</strong>g> the L2 norm and introducing<br />

finite differences, the equati<strong>on</strong>s are given <strong>from</strong> the minimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the two functi<strong>on</strong>als<br />

X M<br />

X K<br />

m¼1 k¼1<br />

X K<br />

XM 1<br />

k¼1 m¼2<br />

ðxk 1;m 2xk;m þ xkþ1;mÞ 2 ; ðx0;m ¼ xK;m; xKþ1;m ¼ x1;mÞ, (3.3)<br />

ðxk;m 1 2xk;m þ xk;mþ1Þ 2 . (3.4)<br />

Mathematically, this c<strong>on</strong>diti<strong>on</strong> seeks to secure that the argument <str<strong>on</strong>g>of</str<strong>on</strong>g> the n<strong>on</strong>-negativity<br />

c<strong>on</strong>straint is piecewise c<strong>on</strong>stant, which implies that the <strong>wave</strong> spectrum is smooth due to the<br />

mathematical properties <str<strong>on</strong>g>of</str<strong>on</strong>g> the exp<strong>on</strong>ential functi<strong>on</strong>.<br />

To deal with prior informati<strong>on</strong> (2), the sum <str<strong>on</strong>g>of</str<strong>on</strong>g> the power is minimised for o ! 0 and<br />

o !1. Mathematically, this is d<strong>on</strong>e by minimising the functi<strong>on</strong>al<br />

X K<br />

k¼1<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 43<br />

½ðxk;1 s0Þ 2 þðxk;M s0Þ 2 Š, (3.5)<br />

where s0 denotes a prespecified small value. It should be noted that the indices (here ‘1’ and<br />

‘M’) representing the <strong>wave</strong> frequency may be extended to include a larger interval, if needed,<br />

in practical computati<strong>on</strong>s. The reas<strong>on</strong> or necessity for the introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> this prior is related<br />

to practical computati<strong>on</strong>s as, if not introduced, problems are likely to arise due to the<br />

frequency resp<strong>on</strong>se functi<strong>on</strong>s being zero (or close to) at the frequency boundaries.<br />

With respect to informati<strong>on</strong> <strong>on</strong> the <strong>wave</strong> directi<strong>on</strong>, knowledge can, in principle, be<br />

included in a similar manner as (3.5); however, with both k and m as running indices. Thus,<br />

the limits <strong>on</strong> k and m must be specified in accordance with the <strong>wave</strong> directi<strong>on</strong>(s) and<br />

frequencies wherein the <strong>wave</strong> energy should be minimised. Informati<strong>on</strong> <strong>on</strong> the directi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

sea-<strong>wave</strong>s (i.e. the <strong>on</strong>-<strong>site</strong> wind generated <strong>wave</strong>s) can <str<strong>on</strong>g>of</str<strong>on</strong>g>ten be deduced <strong>from</strong> knowledge <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the wind directi<strong>on</strong>, excluding informati<strong>on</strong> about swell-<strong>wave</strong>s. An evident problem in this<br />

c<strong>on</strong>text is therefore that with a known wind directi<strong>on</strong>, which gives a good indicati<strong>on</strong> about<br />

the directi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the sea-<strong>wave</strong>s, the minimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> energy applied to a certain, so to<br />

speak, oppo<strong>site</strong> domain <str<strong>on</strong>g>of</str<strong>on</strong>g> the frequency-directi<strong>on</strong>al range excludes swells (as well as seas)<br />

<strong>from</strong> this domain. The present suggesti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> including informati<strong>on</strong> <strong>on</strong> the <strong>wave</strong> directi<strong>on</strong><br />

by minimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> energy in a certain domain is, thus, interesting <strong>from</strong> a<br />

theoretical point <str<strong>on</strong>g>of</str<strong>on</strong>g> view, however, in practical applicati<strong>on</strong>s notable problems exist. One<br />

approach to overcome the problem may be to restrict the minimisati<strong>on</strong> to frequencies<br />

above some limit. Still, this approach is not c<strong>on</strong>sidered to be optimal and to end the<br />

discussi<strong>on</strong>, it can be c<strong>on</strong>cluded that instead <str<strong>on</strong>g>of</str<strong>on</strong>g> doing the minimisati<strong>on</strong> for directi<strong>on</strong>s where<br />

the <strong>wave</strong>s, according to the wind directi<strong>on</strong>, do not come <strong>from</strong>, it would be desirable if<br />

directi<strong>on</strong>s, coincident to the wind directi<strong>on</strong>, were given more ‘‘weight’’ compared to other


44<br />

directi<strong>on</strong>s in the <strong>wave</strong> field. Unfortunately, no method for such a procedure has been<br />

developed. In future work it would therefore be interesting to investigate how to include<br />

informati<strong>on</strong> about the wind directi<strong>on</strong>. For now, however, this issue is not dealt with any<br />

further.<br />

3.3. Final equati<strong>on</strong>s and soluti<strong>on</strong><br />

In matrix notati<strong>on</strong> all c<strong>on</strong>diti<strong>on</strong>s, i.e. the prior informati<strong>on</strong>, may be given by kDx ck 2 ,<br />

see Appendix A. Hence, by combinati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the prior informati<strong>on</strong> and the governing<br />

equati<strong>on</strong> system (2.16), a well-c<strong>on</strong>diti<strong>on</strong>ed least squares soluti<strong>on</strong> is found by minimising,<br />

cf. Akaike [17],<br />

ke A expðxÞ ebk 2 þ u 2 kDx ck 2<br />

(3.6)<br />

which is c<strong>on</strong>trolled by a so-called hyperparameter u.<br />

According to the Bayesian modelling procedure, Akaike [17] transforms the minimisati<strong>on</strong><br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> (3.6) into the maximisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the posterior distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> x, ppostðxjs2 ; uÞ, which is<br />

proporti<strong>on</strong>al to<br />

1<br />

2ps 2<br />

ðN 2 LþNÞ=2<br />

u 2<br />

2ps 2<br />

KM=2<br />

exp<br />

1<br />

2s 2 ðke A expðxÞ ebk 2 þ u 2 kDx ck 2 Þ ,<br />

where u is the hyperparameter that expresses the degree <str<strong>on</strong>g>of</str<strong>on</strong>g> smoothness in the soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> x.<br />

For details, see e.g. Nielsen [8], Iseki and Terada [5] and Press et al. [18].<br />

On the assumpti<strong>on</strong> that the hyperparameter, u, and the variance, s2 , are known,<br />

ppostðxjs2 ; uÞ can be maximised by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> various techniques <strong>from</strong> linear algebra,<br />

e.g. QR factorisati<strong>on</strong>. In practical computati<strong>on</strong>s, the maximisati<strong>on</strong> is d<strong>on</strong>e for a range <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

known values <str<strong>on</strong>g>of</str<strong>on</strong>g> the hyperparameter (and the variance). To obtain the optimum value <str<strong>on</strong>g>of</str<strong>on</strong>g> u<br />

and s2 , Akaike [17] proposed a criteri<strong>on</strong> known as Akaike’s Bayesian informati<strong>on</strong><br />

criteri<strong>on</strong>, abbreviated as ABIC. Thus, the determinati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> u and s is based <strong>on</strong> the<br />

minimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> ABIC, which in its general form is given by<br />

Z<br />

ABIC ¼ 2ln ppostðyj ...Þdy. (3.7)<br />

Basically, this finishes the Bayesian modelling, although with most details <strong>on</strong> algebra and<br />

numeric left out. Details can be found in Nielsen [8]. As regards to the treatment <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

prior distributi<strong>on</strong>s it seems relevant, though, to menti<strong>on</strong> that since three functi<strong>on</strong>als are<br />

introduced, three different hyperparameters, as a matter <str<strong>on</strong>g>of</str<strong>on</strong>g> fact, ought to be c<strong>on</strong>sidered.<br />

However, as the computati<strong>on</strong>al cost in c<strong>on</strong>necti<strong>on</strong> with the optimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

hyperparameters is relatively large, compared to the final outcome <str<strong>on</strong>g>of</str<strong>on</strong>g> the directi<strong>on</strong>al<br />

estimati<strong>on</strong> when c<strong>on</strong>sidering three different hyperparameters, <strong>on</strong>ly <strong>on</strong>e hyperparameter is<br />

dealt with. Especially, the computati<strong>on</strong>al cost is important in the study <str<strong>on</strong>g>of</str<strong>on</strong>g> practical<br />

decisi<strong>on</strong> support systems. Therefore, the optimisati<strong>on</strong> is d<strong>on</strong>e for <strong>on</strong>ly <strong>on</strong>e hyperparameter.<br />

4. Parametric modelling<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

Parameterised <strong>wave</strong> <strong>spectra</strong>, e.g. Goda [19], are typically c<strong>on</strong>sidered reliable for<br />

describing the variati<strong>on</strong> with frequency <str<strong>on</strong>g>of</str<strong>on</strong>g> ocean <strong>wave</strong> <strong>spectra</strong>. Moreover, the angular


spread <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> <strong>spectra</strong> can be described by certain parameters, e.g. L<strong>on</strong>guet-Higgins et al.<br />

[20] and Goda [19]. On this assumpti<strong>on</strong>, the <strong>wave</strong> spectrum to be estimated <strong>from</strong> Eq. (2.16)<br />

is based <strong>on</strong> the following 10-parameter bimodal spectrum, e.g. Tannuri et al. [1] and<br />

Hogben and Cobb [21],<br />

Eðo; yÞ ¼ 1 X<br />

4<br />

2<br />

i¼1<br />

ððð4li þ 1Þ=4Þo 4 p;i Þli<br />

GðliÞ<br />

y ymean;i<br />

2si cos<br />

2<br />

exp<br />

H2 s;i<br />

AðsiÞ<br />

o4liþ1 4li þ 1<br />

4<br />

with Hs being the significant <strong>wave</strong> height, l is the shape parameter <str<strong>on</strong>g>of</str<strong>on</strong>g> the spectrum, ymean is<br />

the mean <strong>wave</strong> directi<strong>on</strong>, op is the angular peak frequency, and s represents the spreading<br />

parameter.<br />

AðsÞ ¼ 22s 1 G2ðs þ 1Þ<br />

(4.2)<br />

pGð2s þ 1Þ<br />

is a c<strong>on</strong>stant introduced to normalise the area under the cos2s curve and G denotes the<br />

gamma functi<strong>on</strong>.<br />

The <strong>wave</strong> spectrum expressed by (4.1) c<strong>on</strong>siders basically a sea comp<strong>on</strong>ent ði ¼ 1Þ and a<br />

swell comp<strong>on</strong>ent ði ¼ 2Þ, and <strong>on</strong> this basis it is, in theory, possible to model most ocean<br />

<strong>wave</strong> <strong>spectra</strong>, e.g. Hogben and Cobb [21]. Hence, by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> (4.1), and taking the<br />

<strong>wave</strong> directi<strong>on</strong> to be coincident to the encounter angle, the soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> equati<strong>on</strong> (2.16)<br />

implies a n<strong>on</strong>-linear optimisati<strong>on</strong> problem <strong>from</strong> which the best-fit-values can be<br />

determined.<br />

4.1. Frequency-dependent spreading<br />

From measurements, cf. L<strong>on</strong>guet-Higgins et al. [20] and Mitsuyasu et al. [22], the<br />

spreading parameter s is found to vary with frequency. In the present Parametric method<br />

the frequency dependency <str<strong>on</strong>g>of</str<strong>on</strong>g> the spreading parameter is included in the sense that s takes a<br />

maximum value around the <strong>spectra</strong>l peak frequency, whereas s decreases for frequencies<br />

both lower and higher than the peak frequency. According to Goda [19] the spreading<br />

parameter can basically be modelled as<br />

(<br />

s ¼ ceil½ðo=opÞ 5 smaxŠ; opop;<br />

ceil½ðo=opÞ 2:5 smaxŠ; o4op;<br />

where the peak value <str<strong>on</strong>g>of</str<strong>on</strong>g> s, denoted smax, has been introduced as the principal parameter.<br />

Compared to the original proposal by Goda [19], ceil[ ] has been introduced in (4.3).<br />

Thus, ceil[ ] rounds towards plus infinity and is a numerical technique utilised<br />

here to stabilise the optimisati<strong>on</strong>. As regards smax, the following values for engineering<br />

applicati<strong>on</strong>s have been recommended, cf. Goda [19],<br />

smax ¼<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 45<br />

8<br />

><<br />

10; wind <strong>wave</strong>s;<br />

25;<br />

>:<br />

75;<br />

swell with short decay distance;<br />

swell with l<strong>on</strong>g decay distance:<br />

op;i<br />

o<br />

4<br />

ð4:1Þ<br />

(4.3)<br />

(4.4)


46<br />

4.2. Estimated parameters<br />

In c<strong>on</strong>clusi<strong>on</strong>, the optimal <strong>wave</strong> spectrum estimated by the Parametric method is found<br />

<strong>from</strong> the optimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the parameters<br />

f Hs;1 l1 ymean;1 op;1 smax;1 Hs;2 l2 ymean;2 op;2 smax;2 g. (4.5)<br />

As has been discussed, the Parametric modelling method assumes basically the <strong>wave</strong><br />

spectrum to be composed <str<strong>on</strong>g>of</str<strong>on</strong>g> a sea and a swell part. In case <str<strong>on</strong>g>of</str<strong>on</strong>g> unimodal seas, however, the<br />

method should predict a negligible value for <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the significant <strong>wave</strong> heights (Hs;1 or<br />

Hs;2). Later, this is verified by a numerical example.<br />

5. Soluti<strong>on</strong> procedures<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

From the preceding it appears that both methods, the Bayesian method and the<br />

Parametric method, in their fundamentals are c<strong>on</strong>ceptually similar and deviate <strong>on</strong>ly in the<br />

sense <str<strong>on</strong>g>of</str<strong>on</strong>g> different (prior) assumpti<strong>on</strong>s about the <strong>wave</strong> spectrum to be estimated. In<br />

practical computati<strong>on</strong>s, though, the soluti<strong>on</strong> procedures <str<strong>on</strong>g>of</str<strong>on</strong>g> the two methods are quite<br />

different, since the Bayesian modelling in a number <str<strong>on</strong>g>of</str<strong>on</strong>g> iterati<strong>on</strong>s, based <strong>on</strong> QR<br />

factorisati<strong>on</strong>, solves an overdetermined linear equati<strong>on</strong> system, whereas the Parametric<br />

method yields the soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a n<strong>on</strong>-linear programming problem. This means that<br />

the soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian modelling can be obtained by use <str<strong>on</strong>g>of</str<strong>on</strong>g> ‘standard’<br />

numerical techniques, c<strong>on</strong>trary to the Parametric modelling which requires the use<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> a genetic optimisati<strong>on</strong> algorithm, or a gradient search algorithm with careful input <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

initial values <str<strong>on</strong>g>of</str<strong>on</strong>g> the optimisati<strong>on</strong> parameters, the so-called search basin. In the present<br />

work, the optimisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the Parametric method is based <strong>on</strong> the latter <str<strong>on</strong>g>of</str<strong>on</strong>g> the two<br />

algorithms.<br />

When the basic equati<strong>on</strong> system is set up, cf. (2.2), N different time series, or signals, are<br />

c<strong>on</strong>sidered. The standard deviati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> these N signals will, in some cases, be <str<strong>on</strong>g>of</str<strong>on</strong>g> a very<br />

different magnitude depending <strong>on</strong> the kind <str<strong>on</strong>g>of</str<strong>on</strong>g> signals being recorded. Hence, the soluti<strong>on</strong><br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the equati<strong>on</strong> system will be affected by this fact, which may be reflected by a poor<br />

agreement between the soluti<strong>on</strong> and the data for the signals with the relatively smallest<br />

standard deviati<strong>on</strong>. C<strong>on</strong>versely, the soluti<strong>on</strong>s corresp<strong>on</strong>ding to the signals with the<br />

relatively largest standard deviati<strong>on</strong> may approximate the data in a better way. In order to<br />

meet this problem the equati<strong>on</strong> system can be normalised. Thus, if it is remembered that<br />

the cross <strong>spectra</strong>, in general, are complex with the real part Cij being the co-spectrum and<br />

with the imaginary part Qij being the quadrature-spectrum, e.g. Price and Bishop [23], so<br />

that<br />

SijðoeÞ ¼CijðoeÞþiQijðoeÞ. (5.1)<br />

Bendat and Piersol [24] suggests to normalise the cross <strong>spectra</strong> according to<br />

bCijðoeÞ<br />

CijðoeÞ<br />

¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi<br />

1<br />

2½SiiðoeÞSjjðoeÞþCijðoeÞ 2 QijðoeÞ 2 q ,<br />

Š<br />

Q<br />

bQ<br />

ijðoeÞ<br />

ijðoeÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi<br />

1<br />

2 ½SiiðoeÞSjjðoeÞ CijðoeÞ 2 þ QijðoeÞ 2 q . ð5:2Þ<br />

Š


The normalisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the equati<strong>on</strong> system has also the implicati<strong>on</strong> that the optimal value<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the hyperparameter, which is an important factor in the Bayesian modelling, assumes<br />

almost the same magnitude in all cases for similar data, cf. Iseki and Terada [5].<br />

6. Numerical example<br />

As a means to verify the two estimati<strong>on</strong> c<strong>on</strong>cepts, numerical simulati<strong>on</strong>s are c<strong>on</strong>ducted.<br />

Thus, <strong>on</strong> the assumpti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a perfect relati<strong>on</strong><strong>ship</strong> between <strong>ship</strong> resp<strong>on</strong>ses and a <strong>wave</strong><br />

spectrum through (given) complex-valued transfer functi<strong>on</strong>s, generated <strong>ship</strong> resp<strong>on</strong>ses can<br />

be used to estimate the underlying <strong>wave</strong> spectrum.<br />

The numerical simulati<strong>on</strong>s dealt with are based <strong>on</strong> the <strong>ship</strong> resp<strong>on</strong>ses fheave, roll, pitchg<br />

and the complex-valued transfer functi<strong>on</strong>s are calculated by a three-dimensi<strong>on</strong>al time<br />

domain code. The c<strong>on</strong>sidered <strong>ship</strong> is a c<strong>on</strong>tainer <strong>ship</strong> with main dimensi<strong>on</strong>s as given in<br />

Table 1, and in the following a speed <str<strong>on</strong>g>of</str<strong>on</strong>g> V ¼ 23 knots is assumed.<br />

Until now both the <strong>wave</strong> and the encounter frequencies have been expressed in terms <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the angular frequency o ¼ 2pf . However, in practical computati<strong>on</strong>s it is <str<strong>on</strong>g>of</str<strong>on</strong>g>ten more<br />

c<strong>on</strong>venient to use the frequency f, and <strong>from</strong> now <strong>on</strong> the <strong>wave</strong> and the encounter<br />

frequencies will be given as f and f e, respectively.<br />

6.1. Simulati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> resp<strong>on</strong>se <strong>spectra</strong><br />

ARTICLE IN PRESS<br />

As regards resp<strong>on</strong>se <strong>spectra</strong> (<str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong> resp<strong>on</strong>ses), simulati<strong>on</strong>s can be calculated <strong>on</strong> the<br />

basis <str<strong>on</strong>g>of</str<strong>on</strong>g> Eq. (2.14) <strong>on</strong> the assumpti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a theoretically known <strong>wave</strong> spectrum, including<br />

the locati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>spectra</strong>l peak(s). Hence, for given operati<strong>on</strong>al c<strong>on</strong>diti<strong>on</strong>s for a specific<br />

<strong>ship</strong>, the resp<strong>on</strong>se <strong>spectra</strong> are given by<br />

bsimul ¼ A0f þ w0, (6.1)<br />

where w0 represents a certain amount <str<strong>on</strong>g>of</str<strong>on</strong>g> white noise, so that real measurements are better<br />

simulated. In the examples, the standard deviati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the white noise is set to 0.5% <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

standard deviati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the actual resp<strong>on</strong>ses. Although f c<strong>on</strong>tains the <strong>wave</strong> spectrum as<br />

functi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> frequency, it should be realised that the multiplicati<strong>on</strong> by A0<br />

transforms the resp<strong>on</strong>se <strong>spectra</strong> into the encounter frequency domain.<br />

In the example, four c<strong>on</strong>diti<strong>on</strong>s, A, B, C and D, are studied, and in each case resp<strong>on</strong>se<br />

<strong>spectra</strong> are generated <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> a theoretically known <strong>wave</strong> spectrum which is<br />

calculated <strong>from</strong> (4.1). Thus, the <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the four c<strong>on</strong>diti<strong>on</strong>s are characterised by a<br />

set <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> parameters which may be estimated <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> the two estimati<strong>on</strong><br />

methodologies. It should be noted that C<strong>on</strong>diti<strong>on</strong> A corresp<strong>on</strong>ds to unimodal seas, see<br />

Table 2, while C<strong>on</strong>diti<strong>on</strong>s B and C represent bimodal seas. Finally, C<strong>on</strong>diti<strong>on</strong> D is taken<br />

Table 1<br />

Main dimensi<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the c<strong>on</strong>sidered <strong>ship</strong><br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 47<br />

Length, Lpp<br />

275.0 m<br />

Breadth, Bmld 40.0 m<br />

Draught, T 12.0 m<br />

Displacement 50,000 t


48<br />

Table 2<br />

Comparis<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> parameters<br />

as the sum <str<strong>on</strong>g>of</str<strong>on</strong>g> a unimodal and a bimodal spectrum, so that the <strong>wave</strong> spectrum <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong><br />

D exhibits three <strong>spectra</strong>l peaks.<br />

6.2. <str<strong>on</strong>g>Estimati<strong>on</strong>s</str<strong>on</strong>g><br />

ARTICLE IN PRESS<br />

C<strong>on</strong>d. Peak 1 Peak 2 Total<br />

Hs (m) T p (s) l ymean (deg.) smax Hs (m) T p (s) l ymean (deg.) smax Hs (m) T s (s)<br />

A TRUE 2.00 11.0 1.00 225 25 – – – – – 2.00 8.54<br />

BAY 1.95 10.5 – 225 – – – – – – 1.95 8.51<br />

OPT 2.00 11.0 1.00 225 25 0.05 9.77 1.00 55 25 2.00 8.54<br />

B TRUE 2.00 11.0 1.00 225 25 1.50 5.00 1.00 295 25 2.50 6.16<br />

BAY – 10.7 – 225 – – 5.56 – 298 – 2.44 6.67<br />

OPT 2.00 11.0 1.00 225 25 1.50 5.00 1.00 295 25 2.50 6.16<br />

C TRUE 2.00 11.0 1.00 225 25 1.50 5.00 1.00 105 25 2.50 6.16<br />

BAY – 10.3 – 225 – – 5.26 – 102 – 2.37 6.66<br />

OPT 1.07 8.80 3.28 150 75 1.53 5.12 1.00 104 25 1.87 4.98<br />

OPT a<br />

2.00 11.0 1.00 225 25 1.50 5.00 1.00 105 25 2.50 6.16<br />

D TRUE 2.00 11.0 1.00 225 25 1.50 5.00 1.00 105 25 3.20 6.92<br />

2.00 11.0 1.00 5 25<br />

BAY – 10.5 – 225 – – 4.76 – 105 – 3.15 7.32<br />

– 10.5 – 5 –<br />

OPT 1.91 11.4 1.00 334 10 1.97 5.00 1.00 113 25 2.75 5.71<br />

a The initial search basin is changed.<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

Table 2 shows the outcome <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the four c<strong>on</strong>diti<strong>on</strong>s in the form <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

resulting <strong>wave</strong> parameters. For each c<strong>on</strong>diti<strong>on</strong>, the true <strong>wave</strong> parameters (TRUE) are<br />

compared with those based <strong>on</strong> the Bayesian method (BAY) and the Parametric method<br />

(OPT). It should be noted that the peak period T p is specified rather than the peak<br />

frequency f p.<br />

As regards the results <str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian method, the significant <strong>wave</strong> height Hs, the shape<br />

factor l and the spreading parameter smax have not been determined for the individual<br />

peaks in case <str<strong>on</strong>g>of</str<strong>on</strong>g> bimodal seas, since this is not possible. The peak period T p and the mean<br />

<strong>wave</strong> directi<strong>on</strong> ymean, <strong>on</strong> the other hand, are based <strong>on</strong> a detailed visual inspecti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> each <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the peaks <str<strong>on</strong>g>of</str<strong>on</strong>g> the spectrum. C<strong>on</strong>trary, the Parametric modelling yields all the parameters <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the individual peaks, since the parameters are the actual soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> this method. The two<br />

rightmost columns <str<strong>on</strong>g>of</str<strong>on</strong>g> Table 2 show the significant <strong>wave</strong> height and the mean <strong>wave</strong> period<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the ‘total’ spectrum. In general, these numbers are calculated by, e.g. Jensen [25],<br />

Hs ¼ 4 ffiffiffiffiffiffi p<br />

m0<br />

(6.2)<br />

and<br />

T s ¼ m0<br />

, (6.3)<br />

m1


where the <strong>spectra</strong>l moments are given as<br />

Z 1<br />

mn ¼ f<br />

0<br />

n<br />

Z p<br />

p<br />

ARTICLE IN PRESS<br />

Eðf ; bÞ db df . (6.4)<br />

In additi<strong>on</strong> to Table 2, Figs. 6–9 illustrate the comparis<strong>on</strong> between the true <strong>wave</strong><br />

spectrum and the estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the four c<strong>on</strong>diti<strong>on</strong>s. In each figure the left plot<br />

(denoted TRUE) is the true <strong>wave</strong> spectrum, whereas the middle and the right plots are the<br />

estimates by the Bayesian method (denoted BAYESIAN) and the Parametric method<br />

(denoted OPTIMISATION), respectively. With regards to the <strong>ship</strong> course, this is<br />

coincident with the radial line <str<strong>on</strong>g>of</str<strong>on</strong>g> 0 , so that beam, quartering and following <strong>wave</strong>s are<br />

experienced for seas in the interval <str<strong>on</strong>g>of</str<strong>on</strong>g> 90–270 .<br />

From the numbers in Table 2 and <strong>from</strong> Figs. 6–9 and with focus <strong>on</strong> C<strong>on</strong>diti<strong>on</strong>s A and B,<br />

there is seen to be a good agreement between the true and the estimated <strong>wave</strong> parameters.<br />

Indeed, it should be noted that the Parametric method (OPT) yields a unimodal spectrum<br />

in C<strong>on</strong>diti<strong>on</strong> A, as the significant <strong>wave</strong> height <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>on</strong>e peak is negligible. As a matter <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

fact the value <str<strong>on</strong>g>of</str<strong>on</strong>g> Hs ¼ 0:05 m is the lower limit used in the optimisati<strong>on</strong> for the significant<br />

<strong>wave</strong> height. As regards C<strong>on</strong>diti<strong>on</strong> C, the Bayesian method produces a good estimate for<br />

270<br />

300<br />

240<br />

330<br />

210<br />

TRUE<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

Fig. 6. C<strong>on</strong>tour plots <str<strong>on</strong>g>of</str<strong>on</strong>g> the true and the two estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong> A. The <strong>ship</strong> course is 0 and the<br />

<strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

330<br />

210<br />

TRUE<br />

0<br />

180<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 49<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.2<br />

0.1<br />

0.4<br />

OPTIMISATION<br />

Fig. 7. C<strong>on</strong>tour plots <str<strong>on</strong>g>of</str<strong>on</strong>g> the true and the two estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong> B. The <strong>ship</strong> course is 0 and the<br />

<strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

0.4<br />

0.3<br />

30<br />

150<br />

30<br />

150<br />

60<br />

120<br />

60<br />

120<br />

90<br />

90


50<br />

270<br />

270<br />

300<br />

240<br />

300<br />

240<br />

210<br />

330<br />

210<br />

330<br />

TRUE<br />

0<br />

0<br />

180<br />

TRUE<br />

180<br />

0.1<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

0.4<br />

30<br />

150<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

330<br />

210<br />

BAYESIAN<br />

BAYESIAN<br />

0<br />

180<br />

the <strong>wave</strong> spectrum. The Parametric modelling, however, gives a poor result, since the<br />

parameters <str<strong>on</strong>g>of</str<strong>on</strong>g> Peak 1 is incorrect, see also Fig. 8. The reas<strong>on</strong> for the inc<strong>on</strong>sistency is to be<br />

found in the initial search basin used for the optimisati<strong>on</strong>. Hence, by changing the search<br />

basin, the <strong>wave</strong> parameters corresp<strong>on</strong>ding to the label OPT a are obtained and it is seen<br />

that the <strong>wave</strong> parameters are then estimated correctly. Clearly, this indicate an important<br />

issue, since the initial search basin for an optimal parametric modelling procedure needs<br />

not to be changed, independently <strong>on</strong> the underlying c<strong>on</strong>diti<strong>on</strong>s. In practical systems based<br />

<strong>on</strong> parametric modelling it is, therefore, likely that a genetic search algorithm will be the<br />

best choice for the optimisati<strong>on</strong>. In this study, however, the optimisati<strong>on</strong> is based <strong>on</strong> a<br />

gradient-based search algorithm. This means that the initial search basin needs to be rather<br />

comprehensive/detailed for which reas<strong>on</strong> the parametric modelling is more time<br />

c<strong>on</strong>suming than the Bayesian modelling. In this numerical example the Parametric<br />

method takes in the order <str<strong>on</strong>g>of</str<strong>on</strong>g> 8 min, whereas the Bayesian modelling takes in the order <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

5 min (with MatLab 7.0 <strong>on</strong> a Pentium 2.0 GHz processor) for each c<strong>on</strong>diti<strong>on</strong>.<br />

Since the Parametric method is based <strong>on</strong> the summati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> two parameterised <strong>wave</strong><br />

<strong>spectra</strong> it is obvious that the three <strong>spectra</strong>l peaks <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong> D make a problem for the<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

Fig. 9. C<strong>on</strong>tour plots <str<strong>on</strong>g>of</str<strong>on</strong>g> the true and the two estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong> D. The <strong>ship</strong> course is 0 and<br />

the <strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

Fig. 8. C<strong>on</strong>tour plots <str<strong>on</strong>g>of</str<strong>on</strong>g> the true and the two estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> C<strong>on</strong>diti<strong>on</strong> C. The <strong>ship</strong> course is 0 and the<br />

<strong>wave</strong>s are shown as approaching.<br />

300<br />

0<br />

180<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

30<br />

150<br />

60<br />

120<br />

60<br />

90<br />

120<br />

90


Parametric method. It is seen that Peak 1, cf. Table 2, in the estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the Parametric<br />

method represents <strong>wave</strong>s which come <strong>from</strong> a directi<strong>on</strong> somewhere in the middle <str<strong>on</strong>g>of</str<strong>on</strong>g> two <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the true <strong>spectra</strong>l peaks, see also Fig. 9. Moreover, Peak 2 <str<strong>on</strong>g>of</str<strong>on</strong>g> the Parametric method<br />

c<strong>on</strong>tains too much energy, although the directi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> those <strong>wave</strong>s is close to that <str<strong>on</strong>g>of</str<strong>on</strong>g> the third<br />

true <strong>spectra</strong>l peak. The Bayesian method, <strong>on</strong> the other hand, yields a <strong>wave</strong> spectrum which<br />

estimates the true <strong>wave</strong> spectrum with good accuracy.<br />

6.3. C<strong>on</strong>cluding remarks<br />

If the Bayesian and the Parametric method are to be compared based <strong>on</strong> the preceding<br />

results, the Bayesian method seems as the best estimati<strong>on</strong> procedure <str<strong>on</strong>g>of</str<strong>on</strong>g> the two. Thus, this<br />

procedure is the least expensive as regards the computati<strong>on</strong>al cost. Furthermore, and most<br />

important, the Bayesian method did estimate a <strong>wave</strong> spectrum in good agreement with the<br />

true spectrum for all four c<strong>on</strong>diti<strong>on</strong>s. This were also partly the case for the Parametric<br />

method, however, C<strong>on</strong>diti<strong>on</strong> C revealed the importance <str<strong>on</strong>g>of</str<strong>on</strong>g> the initial search basin <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

optimisati<strong>on</strong>, and C<strong>on</strong>diti<strong>on</strong> D showed that the Parametric modelling cannot, by any<br />

means, estimate a <strong>wave</strong> spectrum with more than two <strong>spectra</strong>l peaks, since the method is<br />

based <strong>on</strong> the summati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> two parameterised <strong>wave</strong> <strong>spectra</strong>.<br />

It should be noted that there are different opini<strong>on</strong>s about which c<strong>on</strong>cept is the best.<br />

Tannuri et al. [1] performs also numerical tests, for zero speed though, and, c<strong>on</strong>trary to the<br />

present study, Tannuri et al. [1] c<strong>on</strong>cludes that a parametric approach is superior to a<br />

Bayesian approach. The example seen in the present paper, however, indicates clearly that<br />

both <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimati<strong>on</strong> c<strong>on</strong>cepts, based <strong>on</strong> the derived theory in Secti<strong>on</strong>s 2–4, are able to<br />

produce reas<strong>on</strong>able estimates for directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> when numerical simulati<strong>on</strong>s are<br />

studied.<br />

7. Analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> full-scale data<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 51<br />

In this secti<strong>on</strong> full-scale moti<strong>on</strong> measurements <str<strong>on</strong>g>of</str<strong>on</strong>g> a c<strong>on</strong>tainer <strong>ship</strong> are studied. The main<br />

dimensi<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong> are listed in Table 1, and the <strong>ship</strong> has been in route <strong>on</strong> the Pacific<br />

Ocean where all the data has been recorded. The data c<strong>on</strong>sists <str<strong>on</strong>g>of</str<strong>on</strong>g> nine sets: A; B; ...; I, each<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> a durati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> 15 min. The durati<strong>on</strong> is taken in the middle <str<strong>on</strong>g>of</str<strong>on</strong>g> a 30 min period used for the<br />

WAVEX estimati<strong>on</strong> and during this time, the operati<strong>on</strong>al c<strong>on</strong>diti<strong>on</strong>s were (nearly)<br />

c<strong>on</strong>stant. The estimati<strong>on</strong>s are based <strong>on</strong> the three resp<strong>on</strong>ses {heave, roll, pitch}, with<br />

complex-valued transfer functi<strong>on</strong>s calculated by a three-dimensi<strong>on</strong>al time domain code.<br />

Unfortunately, no visual observati<strong>on</strong>s were carried out to hold against the estimati<strong>on</strong>s.<br />

It is important to emphasise that the complex-valued transfer functi<strong>on</strong>s are believed to<br />

give a perfect relati<strong>on</strong><strong>ship</strong> between the <strong>wave</strong> excitati<strong>on</strong>s and the <strong>ship</strong> resp<strong>on</strong>ses, similar to<br />

the numerical example studied in the previous secti<strong>on</strong> where the ‘<strong>measured</strong>’ resp<strong>on</strong>ses were<br />

generated by utilisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the transfer functi<strong>on</strong>s. In reality, however, there is a certain<br />

amount <str<strong>on</strong>g>of</str<strong>on</strong>g> uncertainty related to the transfer functi<strong>on</strong>s due to phenomena such as: not<br />

perfectly c<strong>on</strong>trolled/known operati<strong>on</strong>al c<strong>on</strong>diti<strong>on</strong>s (in particular, as regards the loading<br />

c<strong>on</strong>diti<strong>on</strong>), unknown noise phenomena in the resp<strong>on</strong>se signals, the degree <str<strong>on</strong>g>of</str<strong>on</strong>g> n<strong>on</strong>-linearity<br />

between excitati<strong>on</strong>s and resp<strong>on</strong>ses is difficult to predict about, relative to the size<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong> the high-frequency <strong>wave</strong> comp<strong>on</strong>ents induce no significant moti<strong>on</strong>s,<br />

etc. C<strong>on</strong>sequently, the estimated directi<strong>on</strong>al <strong>wave</strong> spectrum, including its underlying<br />

<strong>wave</strong> parameters, ought to be given with some uncertainty, although the estimati<strong>on</strong>


52<br />

methodology in itself, <strong>on</strong> the assumpti<strong>on</strong>, works perfectly in theory and yields the true<br />

directi<strong>on</strong>al <strong>wave</strong> spectrum. So far, though, nothing can be said about the related<br />

uncertainty <str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian and the Parametric method. As regards WAVEX, the technical<br />

manual [26] c<strong>on</strong>siders the statistical uncertainty <str<strong>on</strong>g>of</str<strong>on</strong>g> the measurements. Thus, based <strong>on</strong><br />

theory assuming that the sea surface elevati<strong>on</strong> is a Gaussian process, it is specified that the<br />

statistical measurement error <str<strong>on</strong>g>of</str<strong>on</strong>g> the standard deviati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the significant <strong>wave</strong> height is<br />

Hs ¼ 0:66 g T p<br />

, (7.1)<br />

2p Lx<br />

where g is the accelerati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> gravity, T p is the peak period <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> spectrum, and Lx is<br />

the image size. The uncertainty <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimated results by WAVEX will not be dealt with<br />

in detail. In the following, reference will, however, be given to the number in (7.1) when<br />

comparis<strong>on</strong>s are c<strong>on</strong>ducted. (The image size is assumed to be Lx ¼ 520 m [26]).<br />

As a general comment to the recorded data it should be noted that n<strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the data sets<br />

have been recorded during particularly light sea states. The reas<strong>on</strong> for this is that the<br />

recording system saves <strong>on</strong>ly entire time histories <str<strong>on</strong>g>of</str<strong>on</strong>g> the hundred so-called worst cases (with<br />

a durati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> 30 min) experienced within a m<strong>on</strong>th. Otherwise, <strong>on</strong>ly statistical values are<br />

stored. The lowest possible sea states are chosen in order to minimise the influence <str<strong>on</strong>g>of</str<strong>on</strong>g> n<strong>on</strong>linearity<br />

between <strong>wave</strong> excitati<strong>on</strong> and resp<strong>on</strong>se.<br />

Furthermore, it should be noted that the cross <strong>spectra</strong>l analysis <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses is<br />

carried out by use <str<strong>on</strong>g>of</str<strong>on</strong>g> MAR modelling, e.g. Akaike and Nakagawa [11], by applicati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the so-called stepwise least squares algorithm, see Neumaier and Schneider [27], for the<br />

determinati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the MAR coefficients. Finally, it should be menti<strong>on</strong>ed that the author<br />

himself has not carried out the calculati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the transfer functi<strong>on</strong>s, for which reas<strong>on</strong><br />

sensitivity analyses <str<strong>on</strong>g>of</str<strong>on</strong>g> the transfer functi<strong>on</strong>s have not been c<strong>on</strong>ducted. The transfer<br />

functi<strong>on</strong>s have been calculated by use <str<strong>on</strong>g>of</str<strong>on</strong>g> a three-dimensi<strong>on</strong>al time domain code.<br />

7.1. Estimated <strong>wave</strong> parameters<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

The nine data sets are analysed by the Bayesian modelling and the Parametric modelling<br />

procedure and the analyses lead to the <strong>wave</strong> parameters presented in Table 3, see also<br />

Table 3<br />

Comparis<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> parameters: the significant <strong>wave</strong> height Hs, the mean <strong>wave</strong> period T s, and the primary<br />

(mean) <strong>wave</strong> directi<strong>on</strong> ym<br />

Data Hs (m) T s (s) ym (deg.)<br />

BAY OPT WAV BAY OPT WAV BAY OPT WAV<br />

A 2.7 2.5 3.0 ( 0.4) 7.5 7.2 6.9 200 220 280<br />

B 2.5 2.3 2.5 ( 0.3) 7.1 7.3 6.8 240 240 225<br />

C 2.3 1.7 2.6 ( 0.3) 7.1 7.4 6.5 240 240 240<br />

D 2.4 1.4 2.5 ( 0.3) 7.0 8.3 6.6 160 130 115<br />

E 3.6 3.4 4.9 ( 1.5) 8.2 11.3 9.2 0 5 330<br />

F 4.0 3.4 4.8 ( 1.0) 7.8 8.6 8.6 180 145 180<br />

G 3.3 3.5 4.5 ( 0.6) 10.2 10.7 8.7 280 265 60<br />

H 3.6 3.7 4.4 ( 1.4) 9.0 9.5 8.8 280 295 35<br />

I 5.7 7.4 7.2 ( 2.2) 13.1 13.0 10.8 185 180 190


270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

ARTICLE IN PRESS<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

Figs. 10–18. The significant <strong>wave</strong> height Hs is calculated according to (6.2), and the mean<br />

<strong>wave</strong> period T s is found <strong>from</strong> (6.3). In the table, numbers are also given for the estimati<strong>on</strong>s<br />

produced by WAVEX. Thus, results <str<strong>on</strong>g>of</str<strong>on</strong>g> the former two methods are denoted by BAY and<br />

OPT, respectively, whereas parameters calculated <strong>from</strong> WAVEX are denoted by WAV. It<br />

is seen that the uncertainty related to the significant <strong>wave</strong> height estimated by WAVEX is<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

WAVEX<br />

Fig. 10. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data A. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

0<br />

180<br />

WAVEX<br />

Fig. 11. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data B. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

BAYESIAN<br />

330<br />

0<br />

0.4<br />

0.3<br />

210<br />

180<br />

0.2<br />

0.1<br />

30<br />

150<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 53<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

OPTIMISATION<br />

330<br />

0<br />

0.4<br />

0.3<br />

210<br />

180<br />

0.2<br />

0.1<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

0.4<br />

WAVEX<br />

0<br />

0.4<br />

Fig. 12. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data C. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

300<br />

240<br />

330<br />

210<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

30<br />

150<br />

30<br />

150<br />

30<br />

150<br />

60<br />

120<br />

60<br />

120<br />

60<br />

90<br />

120<br />

90<br />

90


54<br />

270<br />

270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

ARTICLE IN PRESS<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

included in the parenthesis (calculated by use <str<strong>on</strong>g>of</str<strong>on</strong>g> (7.1)). As regards the columns <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

primary <strong>wave</strong> directi<strong>on</strong> ym, these numbers are relative in the sense that they are given<br />

relative to the <strong>ship</strong>’s l<strong>on</strong>gitudinal axis. The definiti<strong>on</strong> follows <strong>from</strong> WAVEX, so that beam,<br />

quartering and following <strong>wave</strong>s are observed for directi<strong>on</strong>s in the interval <str<strong>on</strong>g>of</str<strong>on</strong>g> 90–270<br />

(similarly to the numerical example). It should be noted that ym represents the mean<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

WAVEX<br />

Fig. 14. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data E. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

BAYESIAN<br />

0<br />

0.4<br />

0.3<br />

30<br />

180<br />

0.1<br />

0.2<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

150<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

60<br />

60<br />

90 270<br />

120<br />

90 270<br />

120<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

OPTIMISATION<br />

0<br />

0.4<br />

0.3<br />

30<br />

180<br />

0.2<br />

0.1<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.2<br />

0.1<br />

0.4<br />

WAVEX<br />

0<br />

0.4<br />

Fig. 15. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data F. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

WAVEX<br />

Fig. 13. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data D. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

0.3<br />

30<br />

150<br />

30<br />

150<br />

30<br />

150<br />

60<br />

120<br />

60<br />

120<br />

60<br />

120<br />

90<br />

90<br />

90


270<br />

270<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

BAYESIAN<br />

0<br />

180<br />

0.1<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

0.4<br />

30<br />

150<br />

30<br />

150<br />

60<br />

60<br />

90 270<br />

120<br />

90 270<br />

120<br />

ARTICLE IN PRESS<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

OPTIMISATION<br />

0<br />

180<br />

directi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the primary <strong>spectra</strong>l peak, which is the peak that c<strong>on</strong>tains the most energy in<br />

the case <str<strong>on</strong>g>of</str<strong>on</strong>g> more peaks being present. Moreover, it should be menti<strong>on</strong>ed that Table 4<br />

c<strong>on</strong>tains all the specific parameters obtained <strong>from</strong> the Parametric method. In the table, S1<br />

ði ¼ 1Þ and S2 ði ¼ 2Þ represent the two comp<strong>on</strong>ents <str<strong>on</strong>g>of</str<strong>on</strong>g> the individual <strong>wave</strong> parameters,<br />

which c<strong>on</strong>trol the parameterised <strong>wave</strong> spectrum, in accordance with expressi<strong>on</strong> (4.1).<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

WAVEX<br />

Fig. 17. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data H. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

0.3<br />

0.2<br />

0.1<br />

0.4<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

WAVEX<br />

Fig. 16. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data G. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 55<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

0.1<br />

0.4<br />

0.3<br />

0.2<br />

30<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

0<br />

180<br />

0<br />

180<br />

0.3<br />

0.2<br />

0.1<br />

WAVEX<br />

Fig. 18. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data I. The <strong>ship</strong> course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

0.4<br />

0.3<br />

0.2<br />

0.1<br />

0.2<br />

0.1<br />

0.4<br />

0.4<br />

0.3<br />

30<br />

150<br />

30<br />

150<br />

30<br />

150<br />

60<br />

120<br />

60<br />

120<br />

60<br />

120<br />

90<br />

90<br />

90


56<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

Table 4<br />

Wave parameters <strong>from</strong> the Parametric method: the peak <strong>wave</strong> period T p, the significant <strong>wave</strong> height Hs, the<br />

(mean) <strong>wave</strong> directi<strong>on</strong> ym, the shape factor l, and the spreading parameter smax<br />

Data T p (s) Hs (m) ym (deg.) l smax<br />

S1 S2 S1 S2 S1 S2 S1 S2 S1 S2<br />

A 7.7 9.1 1.7 1.7 220 295 4.0 1.0 10 75<br />

B 9.1 7.7 1.2 1.9 295 240 1.0 4.0 25 75<br />

C 16.5 7.6 0.4 1.7 295 240 1.0 4.0 25 75<br />

D 16.7 7.3 0.7 1.2 55 130 1.0 4.0 10 25<br />

E 14.3 12.5 0.4 3.4 135 5 1.0 4.0 25 75<br />

F 14.2 7.6 1.9 2.8 130 145 4.0 3.7 10 10<br />

G 10.0 16.7 2.8 2.2 265 0 1.6 4.0 75 75<br />

H 14.4 8.3 2.7 2.6 295 250 2.5 4.0 75 75<br />

I 14.4 20.0 7.0 2.4 180 305 4.0 2.8 25 75<br />

For the individual <strong>wave</strong> parameters, S1 and S2 represent the two <strong>spectra</strong>l peaks <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> spectrum.<br />

From a study <str<strong>on</strong>g>of</str<strong>on</strong>g> the numbers in Table 3 it is seen that, in general, there is a fair<br />

agreement between the Bayesian and the Parametric method. However, notable differences<br />

(in the relative sense) are observed for the significant <strong>wave</strong> height <str<strong>on</strong>g>of</str<strong>on</strong>g> Data C, D and I. This<br />

goes also for the mean <strong>wave</strong> period <str<strong>on</strong>g>of</str<strong>on</strong>g> Data E and F. With regard to the primary <strong>wave</strong><br />

directi<strong>on</strong> and the overall distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> energy <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the two modelling<br />

procedures, see Figs. 10–18, the c<strong>on</strong>sistency is reas<strong>on</strong>able. Especially, it seems that the<br />

introduced frequency dependency, cf. (4.3), <str<strong>on</strong>g>of</str<strong>on</strong>g> the spreading parameter <str<strong>on</strong>g>of</str<strong>on</strong>g> the Parametric<br />

modelling yields directi<strong>on</strong>al estimati<strong>on</strong>s which exhibit somewhat similar properties as seen<br />

<strong>from</strong> the Bayesian estimati<strong>on</strong>s. At this stage it is worth to menti<strong>on</strong> that, although not<br />

shown in the present paper, the introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a frequency-dependent spreading leads<br />

actually to results which are in better agreement with the Bayesian as well as the WAVEX<br />

estimates compared to what is seen without the introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a frequency-dependent<br />

spreading parameter; both with respect to the <strong>wave</strong> parameters and the directi<strong>on</strong>al<br />

distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> energy.<br />

Compared to the <strong>wave</strong> parameters estimated by WAVEX, the parameters <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

Bayesian and the Parametric method deviate slightly with a trend towards better<br />

agreement for Data A–D as opposed to Data E–I. However, with respect to the significant<br />

<strong>wave</strong> height, the best agreement is, in general, found between the Bayesian method and<br />

WAVEX. As regards the mean <strong>wave</strong> period, the agreement between WAVEX and the<br />

Bayesian and the Parametric method is equally good, with reservati<strong>on</strong> to some data (as<br />

menti<strong>on</strong>ed above). The general tendency <str<strong>on</strong>g>of</str<strong>on</strong>g> less good agreement between the significant<br />

<strong>wave</strong> height <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX and the two modelling procedures for Data E–I could be the<br />

existence <str<strong>on</strong>g>of</str<strong>on</strong>g> severer sea states compared to the sea states <str<strong>on</strong>g>of</str<strong>on</strong>g> Data A–D. Thus, it is seen that<br />

WAVEX and the Bayesian method as well as the Parametric method estimate substantially<br />

higher values for Hs, when the last five sets in Table 3 are compared with the first four data<br />

sets. Hence, the assumed linearity between the <strong>ship</strong> resp<strong>on</strong>ses and the <strong>wave</strong>s might be less<br />

good in the last five cases, a fact which would influence the Bayesian and the parametric<br />

results since linearity is fundamental for these methods to be reliable. On the other hand, it<br />

is believed that due to the high inertia <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong>, the errors in the Bayesian and the


ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 57<br />

Parametric estimati<strong>on</strong>s should decrease as the <strong>wave</strong> period increases. The reas<strong>on</strong> is that the<br />

<strong>ship</strong> acts as a filter and, therefore, the <strong>ship</strong> will filter the high-frequency <strong>wave</strong> comp<strong>on</strong>ents<br />

which induce no significant moti<strong>on</strong>s. To judge the increasing deviati<strong>on</strong>s between the<br />

Bayesian/parametric estimates and the WAVEX results observed for Data E–I compared<br />

to Data A–D, it is thus important to keep the filtering aspect in mind as well.<br />

Returning to the deviati<strong>on</strong>s in each individual set <str<strong>on</strong>g>of</str<strong>on</strong>g> data, cf. Table 3, it should be noted<br />

that with respect to the significant <strong>wave</strong> height, WAVEX yields always results which are<br />

larger than those obtained by the Bayesian and the Parametric method, in particular this<br />

applies to Data E–I. The general trend <str<strong>on</strong>g>of</str<strong>on</strong>g> larger significant <strong>wave</strong> heights estimated by<br />

WAVEX is difficult to explain since, <strong>on</strong> the face <str<strong>on</strong>g>of</str<strong>on</strong>g> it, it may be the Bayesian/Parametric<br />

estimati<strong>on</strong>s which exhibit too little energy or it may be the WAVEX results which exhibit<br />

too much energy. As will be seen in a later secti<strong>on</strong>, studies suggest that energy is lacking in<br />

the resp<strong>on</strong>se <strong>spectra</strong> when measurements and estimati<strong>on</strong>s based <strong>on</strong> the Bayesian/<br />

Parametric results are compared. The same trend is, however, seen for the estimated<br />

resp<strong>on</strong>se <strong>spectra</strong> based <strong>on</strong> the WAVEX results, although the tendency is not as unique as<br />

observed for the Bayesian/Parametric results. Therefore, more detailed investigati<strong>on</strong>s and<br />

analyses would be required to settle <strong>on</strong> the discrepancy in the significant <strong>wave</strong> height<br />

estimated by the Bayesian and the Parametric method and by WAVEX. In the end, this<br />

would require additi<strong>on</strong>al knowledge <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>on</strong>-<strong>site</strong> sea state in the given situati<strong>on</strong> and,<br />

unfortunately, this informati<strong>on</strong> cannot be attained. However, it should also be realised<br />

that—taking the uncertainty <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX into account—almost all <str<strong>on</strong>g>of</str<strong>on</strong>g> the cases, the<br />

Bayesian as well as the Parametric method, are within the c<strong>on</strong>fidence range <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX,<br />

c<strong>on</strong>sidering the statistical uncertainty.<br />

As regards the mean <strong>wave</strong> period T s, Table 3 shows no c<strong>on</strong>sistent tendency in the<br />

deviati<strong>on</strong>s between the Bayesian/Parametric estimates and the WAVEX estimates<br />

and, in general, the magnitude <str<strong>on</strong>g>of</str<strong>on</strong>g> the differences must be c<strong>on</strong>sidered acceptable <strong>from</strong> an<br />

engineering point <str<strong>on</strong>g>of</str<strong>on</strong>g> view.<br />

The primary (mean) <strong>wave</strong> directi<strong>on</strong> ym, cf. Table 3, is found with reas<strong>on</strong>able c<strong>on</strong>sistency<br />

when the results <str<strong>on</strong>g>of</str<strong>on</strong>g> the two modelling procedures are compared with those <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX, see<br />

also Figs. 10–18. However, both for Data G and H there seems to be a c<strong>on</strong>flict whether<br />

<strong>wave</strong>s are approaching <strong>from</strong> <strong>on</strong>e side or the other side <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>ship</strong>. Thus, the Bayesian as<br />

well as the Parametric estimati<strong>on</strong>s have <strong>wave</strong>s approaching <strong>on</strong> the port side, whereas<br />

WAVEX yields <strong>wave</strong>s entering <strong>from</strong> the starboard side. It is by no means easy to give a<br />

reas<strong>on</strong> for this c<strong>on</strong>troversy since nothing is changed in the computati<strong>on</strong>al set-up <str<strong>on</strong>g>of</str<strong>on</strong>g> the two<br />

data sets. An apparent and straightforward explanati<strong>on</strong> for the discrepancy, though, could<br />

be related to port/starboard symmetric frequency resp<strong>on</strong>se functi<strong>on</strong>s used in the Bayesian/<br />

Parametric estimati<strong>on</strong>s. However, as the roll resp<strong>on</strong>se is used as <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses in the<br />

present data analysis, the port/starboard asymmetry is reflected in the estimati<strong>on</strong>s. In this<br />

c<strong>on</strong>text it should be remembered that by use <str<strong>on</strong>g>of</str<strong>on</strong>g> complex-valued transfer functi<strong>on</strong>s, the<br />

amplitude as well as the phase <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses are taken into account. For this reas<strong>on</strong> the<br />

c<strong>on</strong>troversy in the <strong>wave</strong> directi<strong>on</strong> for Data G and Data H has also not to do with the phase<br />

shift <str<strong>on</strong>g>of</str<strong>on</strong>g> the roll transfer functi<strong>on</strong> experienced at frequencies close to the natural roll<br />

frequency in beam seas. An estimate, e.g. Lloyd [28], <str<strong>on</strong>g>of</str<strong>on</strong>g> the natural roll frequency <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

<strong>ship</strong> is found to be f r ’ 0:06 Hz ð T r ’ 15 s) and Fig. 19, which depicts the roll phase<br />

angle, c<strong>on</strong>firms, more or less, this value and illustrates the menti<strong>on</strong>ed shift in phase.<br />

It can therefore be excluded that the roll resp<strong>on</strong>se as such has to do with the c<strong>on</strong>troversy<br />

in the <strong>wave</strong> directi<strong>on</strong> observed for the two data sets. For this reas<strong>on</strong> it would be interesting


58<br />

[deg.]<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

Roll Phase<br />

15<br />

10<br />

5<br />

0<br />

-5<br />

-10<br />

-15<br />

0 0.05 0.1 0.15 0.2<br />

Wave Freq. [Hz]<br />

Fig. 19. The phase angle <str<strong>on</strong>g>of</str<strong>on</strong>g> the roll transfer functi<strong>on</strong> for an encounter angle equal to 90 . The operati<strong>on</strong>al<br />

c<strong>on</strong>diti<strong>on</strong>s corresp<strong>on</strong>d to Data G and H.<br />

Table 5<br />

Estimated <strong>wave</strong> parameters based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}<br />

Data Hs (m) T s (s) ym (deg.)<br />

BAY OPT BAY OPT BAY OPT<br />

A 2.4 ( 11%) 2.2 ( 12%) 8.1 (8%) 7.6 (6%) 200 (0) 300 (80)<br />

B 2.3 ( 8%) 1.9 ( 17%) 8.1 (14%) 6.7 ( 8%) 240 (0) 275 (35)<br />

C 2.0 ( 13%) 1.1 ( 35%) 7.6 (9%) 6.5 ( 12%) 240 (0) 270 (30)<br />

D 2.2 ( 8%) 1.2 ( 14%) 8.0 (14%) 8.4 (1%) 125 ( 35) 90 ( 40)<br />

E 3.1 ( 14%) 3.5 (3%) 9.5 (16%) 10.0 ( 12%) 0 (0) 0 ( 5)<br />

F 3.9 ( 3%) 3.5 (3%) 8.9 (14%) 11.5 (34%) 225 (45) 200 (55)<br />

G 3.4 (3%) 3.7 (6%) 9.5 ( 7%) 11.4 (7%) 270 ( 10) 300 (35)<br />

H 3.7 (3%) 4.1 (11%) 9.5 (6%) 9.0 ( 5%) 270 ( 10) 250 ( 45)<br />

I 7.2 (26%) 8.5 (15%) 10.1 ( 23%) 12.5 ( 4%) 185 (0) 180 (0)<br />

Deviati<strong>on</strong> relative to the estimated parameters based <strong>on</strong> {heave, roll, pitch} is shown in parenthesis.<br />

to do the Bayesian and the Parametric estimati<strong>on</strong>s with the roll resp<strong>on</strong>se replaced by<br />

another resp<strong>on</strong>se, e.g. the sway resp<strong>on</strong>se which is also an odd functi<strong>on</strong> and therefore it<br />

exhibits port/starboard asymmetry as well. Such estimati<strong>on</strong>s, with the roll replaced by the<br />

sway, have been c<strong>on</strong>ducted and, as seen <strong>from</strong> Table 5, the sway-based <strong>wave</strong> parameters<br />

deviate <strong>on</strong>ly slightly compared to the estimati<strong>on</strong>s where the roll resp<strong>on</strong>se is used. Similarly,<br />

the directi<strong>on</strong>al distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> energy is more or less identical, independently if the roll or<br />

the sway resp<strong>on</strong>se is used in the Bayesian/Parametric estimati<strong>on</strong>s, cf. Figs. B1–B5 in<br />

Appendix B which show the directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong>, when the roll resp<strong>on</strong>se is replaced by<br />

the sway. As regards Data G and H, the sway-based estimati<strong>on</strong>s do not specifically c<strong>on</strong>firm<br />

the roll-based Bayesian/Parametric estimati<strong>on</strong>s and nor do they specifically indicate results<br />

in favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the WAVEX estimati<strong>on</strong>s. Thus, it is seen that the sway-based Bayesian<br />

estimati<strong>on</strong>s have <strong>wave</strong>s approaching <strong>from</strong> the starboard side as well as the port side,<br />

however, with the <strong>wave</strong>s <strong>on</strong> the port side c<strong>on</strong>taining the most energy. Therefore, in<br />

c<strong>on</strong>clusi<strong>on</strong> <strong>on</strong> the c<strong>on</strong>troversy in the <strong>wave</strong> directi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data G and H, it cannot be<br />

decisively stated whether <strong>wave</strong>s in reality were entering <strong>on</strong> the port or the starboard side.<br />

As menti<strong>on</strong>ed in Secti<strong>on</strong> 1, Tannuri et al. [1] suggests to use the sway resp<strong>on</strong>se instead <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

the roll. The present paper will not argue against this proposal, however, to follow <strong>on</strong> the<br />

0.25


topic <str<strong>on</strong>g>of</str<strong>on</strong>g> sway vs. roll, it should be noted that, based <strong>on</strong> the estimati<strong>on</strong>s carried out in the<br />

present work, it seems hard to justify whether the <strong>on</strong>e resp<strong>on</strong>se should be proposed in<br />

favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the other.<br />

The Parametric method is based <strong>on</strong> the summati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> two parameterised <strong>wave</strong> <strong>spectra</strong><br />

and <strong>from</strong> the data sets c<strong>on</strong>sidered here, it appears to be reas<strong>on</strong>able to use a number <str<strong>on</strong>g>of</str<strong>on</strong>g> two<br />

<strong>spectra</strong>. Thus, it is seen, see Figs. 10–18, that in neither <str<strong>on</strong>g>of</str<strong>on</strong>g> the data sets no more than two<br />

(fully separated) <strong>spectra</strong>l peaks are estimated by the Bayesian method and by WAVEX.<br />

Moreover, the Parametric modelling predicts (nearly) unimodal seas, cf. Table 4, for Data B,<br />

C and I, which is c<strong>on</strong>sistent to the Bayesian and the WAVEX estimati<strong>on</strong>s. However, with<br />

c<strong>on</strong>siderati<strong>on</strong> to the directi<strong>on</strong>al <strong>wave</strong> spectrum <str<strong>on</strong>g>of</str<strong>on</strong>g> Data E, cf. Fig. 14, it should be noted that<br />

both the Bayesian and the WAVEX estimati<strong>on</strong>s exhibit two adjoined <strong>spectra</strong>l peaks, that is,<br />

the peaks are evidently separated but very close to each other. This property is not captured<br />

by the Parametric method <str<strong>on</strong>g>of</str<strong>on</strong>g> Data E, cf. Table 4. In general, the property <str<strong>on</strong>g>of</str<strong>on</strong>g> adjoining peaks<br />

is hard to be handled by the Parametric method because <str<strong>on</strong>g>of</str<strong>on</strong>g> the prespecified shape <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

<strong>wave</strong> spectrum, frequency- as well as directi<strong>on</strong>al-wise. A fact which is also touched partly by<br />

Pascoal et al. [9], which in relati<strong>on</strong> to a parametric formulati<strong>on</strong>/modelling state: ‘‘The<br />

formulati<strong>on</strong> has shown to lack enough informati<strong>on</strong> to allow multiple peaked <strong>spectra</strong> with<br />

str<strong>on</strong>g frequency overlap to be correctly estimated in a reas<strong>on</strong>able time span, thus multiple<br />

sea or swell comp<strong>on</strong>ents are, for now, out <str<strong>on</strong>g>of</str<strong>on</strong>g> reach in this formulati<strong>on</strong>.’’ Although Data E to<br />

some extent justifies this statement, at least directi<strong>on</strong>al-wise, the present paper is not able to<br />

give a c<strong>on</strong>clusive accedence to the statement.<br />

7.2. Resp<strong>on</strong>se measurements<br />

Based <strong>on</strong> the <strong>measured</strong> resp<strong>on</strong>se spectrum, cf. Appendix C, for each <str<strong>on</strong>g>of</str<strong>on</strong>g> the c<strong>on</strong>sidered<br />

resp<strong>on</strong>ses, heave, roll, and pitch, the variance <str<strong>on</strong>g>of</str<strong>on</strong>g> the individual resp<strong>on</strong>ses has been<br />

calculated for all the data sets. The results are seen in Table 6.<br />

Corresp<strong>on</strong>dingly, the variances <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses can be calculated by use <str<strong>on</strong>g>of</str<strong>on</strong>g> the<br />

estimated <strong>wave</strong> spectrum. Hence, by calculati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>se spectrum in the <strong>wave</strong><br />

frequency domain (cf. (2.2))<br />

SRðf Þ¼<br />

Z p<br />

p<br />

ARTICLE IN PRESS<br />

jFRðf ; bÞj 2 Eðf ; bÞ db (7.2)<br />

Table 6<br />

Measured variances, calculated <strong>from</strong> the resp<strong>on</strong>se <strong>spectra</strong><br />

Data Heave ðm 2 Þ Roll ðrad 2 Þ Pitch ðrad 2 Þ<br />

A 5:7 10 2<br />

B 4:0 10 2<br />

C 1:5 10 2<br />

D 3:6 10 2<br />

E 22 10 2<br />

F 14 10 2<br />

G 72 10 2<br />

H 61 10 2<br />

I 65 10 2<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 59<br />

3:1 10 4<br />

13 10 4<br />

8:3 10 4<br />

2:1 10 4<br />

0:60 10 4<br />

19 10 4<br />

0:88 10 4<br />

3:0 10 4<br />

30 10 4<br />

1:3 10 5<br />

1:1 10 5<br />

0:46 10 5<br />

0:69 10 5<br />

20 10 5<br />

4:7 10 5<br />

18 10 5<br />

13 10 5<br />

26 10 5


60<br />

Table 7<br />

Deviati<strong>on</strong> <strong>on</strong> variances for the c<strong>on</strong>sidered resp<strong>on</strong>ses<br />

the variance is given, or estimated, by<br />

VarR ¼<br />

Z 1<br />

0<br />

ARTICLE IN PRESS<br />

Data Heave Roll Pitch<br />

BAY (%) OPT (%) WAV (%) BAY (%) OPT (%) WAV (%) BAY (%) OPT (%) WAV (%)<br />

A 26 5.6 1.6 5.2 20 38 22 13 40<br />

B 15 3.8 37 13 40 63 15 9.3 20<br />

C 4.9 3.1 10 14 37 36 14 3.5 33<br />

D 16 6.6 25 4.2 11 49 30 10 75<br />

E 21 6.4 86 22 40 9.6 73 49 14<br />

F 6.8 6.1 47 3.8 0.5 83 66 53 0.5<br />

G 26 17 34 2.0 9.6 19 79 24 40<br />

H 19 5.2 8.5 7.6 0.2 27 62 20 7<br />

I 21 4.8 16 2.7 5.9 362 70 37 39<br />

Figures relative to the <strong>measured</strong> variance.<br />

[m 2 ]<br />

0.8<br />

0.7<br />

0.6<br />

0.5<br />

0.4<br />

0.3<br />

0.2<br />

0.1<br />

0<br />

Heave variance<br />

Measured<br />

BAY<br />

OPT<br />

WAV<br />

A B C D E F G H I<br />

Data set<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

[rad 2 ]<br />

0.014<br />

0.012<br />

0.01<br />

0.008<br />

0.006<br />

0.004<br />

0.002<br />

0<br />

Measured<br />

BAY<br />

OPT<br />

WAV<br />

Roll variance<br />

A B C D E F G H I<br />

Data set<br />

Fig. 20. Visualisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Tables 6 and 7.<br />

A B C D E F G H I<br />

Data set<br />

SRðf Þ df , (7.3)<br />

where index R represents the specific resp<strong>on</strong>se.<br />

Table 7 shows the relative deviati<strong>on</strong> between the estimated variance and the <strong>measured</strong><br />

variance (Table 6) <str<strong>on</strong>g>of</str<strong>on</strong>g> the individual resp<strong>on</strong>ses for each data set, when the estimati<strong>on</strong> is<br />

based <strong>on</strong> the Bayesian method (BAY), the Parametric method (OPT) and WAVEX<br />

(WAV), respectively. The results <str<strong>on</strong>g>of</str<strong>on</strong>g> Tables 6 and 7 are visualised graphically in Fig. 20<br />

which shows comparis<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the variances by use <str<strong>on</strong>g>of</str<strong>on</strong>g> bar-diagrams.<br />

As seen <strong>from</strong> Fig. 20 and the numbers in Table 7 it appears, first and foremost, that the<br />

estimated variances based <strong>on</strong> the Bayesian and the Parametric method (almost)<br />

c<strong>on</strong>sistently are less than the <strong>measured</strong> variances. This tendency is somewhat peculiar<br />

and indicates that energy is lacking in those estimati<strong>on</strong>s. As regards to the estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

WAVEX, a similar trend is not seen. However, the estimati<strong>on</strong>s by WAVEX do not, in any<br />

way, reflect a better agreement with the measurements. On the c<strong>on</strong>trary, it is the variances<br />

based <strong>on</strong> the Parametric estimati<strong>on</strong>s which generally deviate the least <strong>from</strong> the <strong>measured</strong><br />

variances. Likewise, it is seen that the Bayesian estimates in many cases are closer to the<br />

<strong>measured</strong> variances compared to the estimati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> WAVEX. An excepti<strong>on</strong> is the pitch<br />

[rad 2 ]<br />

2<br />

1<br />

0<br />

x 10 -4 Pitch variance<br />

Measured<br />

BAY<br />

OPT<br />

WAV


ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 61<br />

resp<strong>on</strong>se, where the Bayesian estimati<strong>on</strong>s are particularly poor for Data E–I. In this<br />

relati<strong>on</strong>, it should be noted, though, that the deviati<strong>on</strong>s, in general, are significantly larger<br />

for the last five data sets compared to Data A–D, when the heave and the pitch estimati<strong>on</strong>s<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian and the Parametric method are studied. This fact could be related to the<br />

severer sea states observed for Data E–I, cf. Table 3, which may mean a less good linear<br />

relati<strong>on</strong><strong>ship</strong> between the <strong>ship</strong> resp<strong>on</strong>ses and the <strong>wave</strong>s. Besides, it is interesting to note how<br />

the heave and the pitch resp<strong>on</strong>ses are markedly increased for all <str<strong>on</strong>g>of</str<strong>on</strong>g> the last five data sets,<br />

whereas the roll resp<strong>on</strong>se experiences <strong>on</strong>ly an increase for Data F and I. Both <str<strong>on</strong>g>of</str<strong>on</strong>g> these data<br />

sets, F and I, are characterised by following (and quartering) seas according to the <strong>wave</strong><br />

<strong>spectra</strong> produced by all three estimati<strong>on</strong> procedures, cf. Figs. 15 and 18. This indicates that<br />

the roll moti<strong>on</strong> is a resp<strong>on</strong>se which is quite sensitive to the encounter angle between the<br />

<strong>ship</strong> and the <strong>wave</strong>s. Hence, in the respect <str<strong>on</strong>g>of</str<strong>on</strong>g> decisi<strong>on</strong> support for the specific <strong>ship</strong>, it is<br />

evident that course changes are likely to be effective to decrease the roll resp<strong>on</strong>se, in spite<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> a severer sea state. This is to some extent seen by comparis<strong>on</strong> with Data G and H, cf.<br />

Figs. 16 and 17, which are characterised by beam and quartering seas (although with a<br />

port/starboard c<strong>on</strong>troversy as discussed previously) and characterised by an energy<br />

c<strong>on</strong>tent <str<strong>on</strong>g>of</str<strong>on</strong>g> notable size. Still, the roll variances <str<strong>on</strong>g>of</str<strong>on</strong>g> Data G and H are am<strong>on</strong>gst the lowest<br />

recorded (and estimated) <str<strong>on</strong>g>of</str<strong>on</strong>g> the nine data sets.<br />

Previously, it was argued that it is the variances estimated by the Parametric method<br />

which, generally, deviate the least <strong>from</strong> the <strong>measured</strong> variances. Without further<br />

explanati<strong>on</strong> this statement may seem rather subjective. However, by assigning points to<br />

the individual estimati<strong>on</strong> procedures according to the value <str<strong>on</strong>g>of</str<strong>on</strong>g> the deviati<strong>on</strong> between the<br />

estimated and the <strong>measured</strong> variance <str<strong>on</strong>g>of</str<strong>on</strong>g> the respective resp<strong>on</strong>ses, a kind <str<strong>on</strong>g>of</str<strong>on</strong>g> objective<br />

argument can be stated. Thus, by c<strong>on</strong>siderati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the individual resp<strong>on</strong>ses <str<strong>on</strong>g>of</str<strong>on</strong>g> each data set<br />

in Table 7, 1 point is given to the method which has the smallest relative deviati<strong>on</strong> <strong>on</strong> the<br />

variance, and 2 and 3 points are given to the methods with the sec<strong>on</strong>d smallest and the<br />

largest relative deviati<strong>on</strong>s <strong>on</strong> the variances, respectively. Based <strong>on</strong> this allotment, Table 8 is<br />

obtained and it is seen that, indeed, the Parametric modelling (OPT) is the method which<br />

scores the least total points. Specifically, it is observed that the Bayesian method, <strong>on</strong><br />

average, yields the best agreement, when the roll resp<strong>on</strong>se is c<strong>on</strong>sidered. From the<br />

viewpoint <str<strong>on</strong>g>of</str<strong>on</strong>g> the soluti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the estimati<strong>on</strong> procedures, it is not surprising that the smallest<br />

relative deviati<strong>on</strong>s are found for the Parametric method and the Bayesian method. This is<br />

so, because the estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> these methods are based <strong>on</strong> the factually<br />

<strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong>. For this reas<strong>on</strong>, it could be anticipated that the agreement<br />

between the <strong>measured</strong> and the estimated variances <str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian (as well as the<br />

Parametric) method would be better than what is seen <strong>from</strong> Table 7. As regards the<br />

Bayesian method, the explanati<strong>on</strong>, for this not being so, is believed to has to do with<br />

the hyperparameter, u, because no matter whether the frequency resp<strong>on</strong>se functi<strong>on</strong>s are<br />

Table 8<br />

Points assigned to the estimati<strong>on</strong> methods for the nine cases A–I according to the agreement with the <strong>measured</strong><br />

variances <str<strong>on</strong>g>of</str<strong>on</strong>g> the respective resp<strong>on</strong>ses (The less points, the better the agreement.)<br />

Heave Roll Pitch Total<br />

BAY 21 12 23 56<br />

OPT 10 18 12 40<br />

WAV 23 24 19 66


62<br />

(completely) wr<strong>on</strong>g, whether n<strong>on</strong>-linearities in reality dominate the relati<strong>on</strong> between<br />

excitati<strong>on</strong>s and resp<strong>on</strong>ses, whether the cross <strong>spectra</strong>l analysis c<strong>on</strong>tains errors, etc., the<br />

soluti<strong>on</strong> x, i.e. the estimated <strong>wave</strong> spectrum, is basically found by minimising<br />

ke A expðxÞ ebk 2 þ u 2 kDx ck 2 . (7.4)<br />

Hence, the soluti<strong>on</strong>, although maybe unstable and completely unrealistic, will in the best<br />

possible way approximate the <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> c<strong>on</strong>tained in the vector eb,by proper multiplicati<strong>on</strong> and with reservati<strong>on</strong> about smoothing. Thus, the more smoothing,<br />

i.e. the larger the hyperparameter, which is needed in accordance with ABIC, cf. (3.7), the<br />

less good agreement can be expected between soluti<strong>on</strong> and data. And, apparently, the<br />

studied data sets require some smoothing to be applied. From a more general point <str<strong>on</strong>g>of</str<strong>on</strong>g> view<br />

it can be said that the fundamental equati<strong>on</strong> system (2.14), which applies both to the<br />

Bayesian and the Parametric method, without introducti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> any assumpti<strong>on</strong>s <strong>on</strong> the<br />

<strong>wave</strong> spectrum to be estimated, is underdetermined (or degenerate) in such a way that it<br />

permits <strong>on</strong>ly fitting with soluti<strong>on</strong>s which lacks energy compared to the data.<br />

Although the problem <str<strong>on</strong>g>of</str<strong>on</strong>g> lacking energy in the resp<strong>on</strong>se <strong>spectra</strong> has been looked into<br />

intensively, the latter explanati<strong>on</strong> is, unfortunately, the <strong>on</strong>ly reas<strong>on</strong> that can be given to<br />

explain the lack <str<strong>on</strong>g>of</str<strong>on</strong>g> energy in the resp<strong>on</strong>se <strong>spectra</strong>.<br />

7.3. Predicti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> sway resp<strong>on</strong>se<br />

ARTICLE IN PRESS<br />

Since the sway resp<strong>on</strong>se has also been <strong>measured</strong>, it should be interesting to see how well<br />

the variance <str<strong>on</strong>g>of</str<strong>on</strong>g> this resp<strong>on</strong>se can be predicted by use <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> spectrum—estimated <strong>on</strong><br />

the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> the heave, the roll and the pitch resp<strong>on</strong>ses—and the transfer functi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> sway.<br />

Table 9 shows the <strong>measured</strong> variance <str<strong>on</strong>g>of</str<strong>on</strong>g> sway for each <str<strong>on</strong>g>of</str<strong>on</strong>g> the data sets and, moreover, the<br />

table lists the predicted values <str<strong>on</strong>g>of</str<strong>on</strong>g> the variance in case the <strong>wave</strong> spectrum is estimated by the<br />

Bayesian method, the Parametric method and by WAVEX, respectively. As it appears<br />

<strong>from</strong> the table the agreement between the <strong>measured</strong> and the predicted values is far <strong>from</strong><br />

impressive. However, the latter five data sets reveal an agreement which is markedly<br />

improved compared to the agreement observed for Data A–D; this applies (almost)<br />

Table 9<br />

Comparis<strong>on</strong> between the <strong>measured</strong> sway variance and the estimated sway variances <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong><br />

spectrum estimated <strong>from</strong> the resp<strong>on</strong>ses {heave, roll, pitch}<br />

Data Sway ð10 2 m 2 Þ<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

Measured BAY OPT WAV<br />

A 2:44 11:5 (376%) 4.89 (101%) 8.22 (241%)<br />

B 3:17 11.5 (263%) 3.95 (27%) 6.39 (105%)<br />

C 1:25 11.2 (872%) 3.42 (198%) 11.2 (873%)<br />

D 2:35 13.2 (470%) 2.59 (12%) 3.04 (31%)<br />

E 1:53 0.82 ( 46%) 0.28 ( 81%) 1.81 (20%)<br />

F 20:9 23.4 (17%) 20.4 (2.1%) 20.8 (3.7%)<br />

G 19:2 17.1 ( 10%) 9.07 ( 52%) 5.35 ( 72%)<br />

H 24:1 13.5 ( 45%) 8.84 ( 64%) 7.04 ( 71%)<br />

I 124 31.2 ( 75%) 20.7 ( 83%) 40.9 ( 67%)<br />

Deviati<strong>on</strong> relative to <strong>measured</strong> variance shown in parenthesis.


c<strong>on</strong>sistently to all the estimati<strong>on</strong> procedures. Keeping in mind the severity <str<strong>on</strong>g>of</str<strong>on</strong>g> the sea states<br />

for the individual sets <str<strong>on</strong>g>of</str<strong>on</strong>g> data, it is somewhat peculiar that the best agreement is found for<br />

Data E–I.<br />

The predicti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> resp<strong>on</strong>ses is <str<strong>on</strong>g>of</str<strong>on</strong>g> vital importance in the specific area <str<strong>on</strong>g>of</str<strong>on</strong>g> decisi<strong>on</strong> support<br />

for safe navigati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong>s which means that the above c<strong>on</strong>siderati<strong>on</strong>s <strong>on</strong> the predicti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

sway does not as such bel<strong>on</strong>g to the objective, or the topic, <str<strong>on</strong>g>of</str<strong>on</strong>g> the present work; the<br />

estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>wave</strong> <strong>spectra</strong>. However, it is interesting to see that the combined use <str<strong>on</strong>g>of</str<strong>on</strong>g> a<br />

<strong>wave</strong> spectrum and a transfer functi<strong>on</strong> leads to relatively different results compared to<br />

what is <strong>measured</strong>, although the estimated directi<strong>on</strong>al <strong>wave</strong> spectrum in the overall sense,<br />

including the underlying <strong>wave</strong> parameters, is more or less similar in the cases, when the<br />

spectrum is estimated <strong>from</strong> {heave, roll, pitch} and <strong>from</strong> {heave, sway, pitch}, cf. Table 5.<br />

This fact illustrates the importance <str<strong>on</strong>g>of</str<strong>on</strong>g> reliable (complex-valued) transfer functi<strong>on</strong>s.<br />

7.4. Bayesian vs. parametric modelling<br />

Based <strong>on</strong> the discussi<strong>on</strong>s in the preceding secti<strong>on</strong>s <strong>on</strong> <strong>wave</strong> parameters and resp<strong>on</strong>se<br />

measurements, it is not straight forward to c<strong>on</strong>clude, which estimati<strong>on</strong> procedure is the<br />

(most) correct. To draw such a c<strong>on</strong>clusi<strong>on</strong> knowledge <str<strong>on</strong>g>of</str<strong>on</strong>g> the absolute true <strong>wave</strong> spectrum<br />

would be needed. Though, it should be noted that, in general, the Parametric modelling<br />

and the Bayesian modelling method yield results which are in good agreement with each<br />

other. Moreover, the results are comparable with the estimati<strong>on</strong>s by WAVEX, so that<br />

the directi<strong>on</strong>al distributi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> energy is nearly similar as well are the underlying <strong>wave</strong><br />

parameters. In the view <str<strong>on</strong>g>of</str<strong>on</strong>g> the produced results it is, therefore, difficult to propose the<br />

Bayesian method in favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the Parametric method, or vice versa. There are, however,<br />

some general points which should also be addressed to get a more subtle picture when<br />

judging the two modelling procedures.<br />

For the analysed data sets, the computati<strong>on</strong>al time were always <str<strong>on</strong>g>of</str<strong>on</strong>g> the l<strong>on</strong>gest durati<strong>on</strong><br />

for the Parametric method; in the order <str<strong>on</strong>g>of</str<strong>on</strong>g> 10 min compared to the order <str<strong>on</strong>g>of</str<strong>on</strong>g> 5 min for the<br />

Bayesian method (with MatLab 7.0 <strong>on</strong> a Pentium 2.0 GHz processor).<br />

The initial search basin is paramount and must be detailed for the Parametric method,<br />

when a genetic search algorithm is not applied.<br />

The number <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>spectra</strong>l peaks is limited (to a certain number) in the Parametric method,<br />

and in the event <str<strong>on</strong>g>of</str<strong>on</strong>g> str<strong>on</strong>g frequency/directi<strong>on</strong>al overlap the individual <strong>spectra</strong>l peaks<br />

may not be ‘‘caught’’.<br />

As regards the ABIC criteri<strong>on</strong> in the Bayesian method, problems have been experienced<br />

in the sense that it appears that the best soluti<strong>on</strong> was not obtained for the minimum<br />

value <str<strong>on</strong>g>of</str<strong>on</strong>g> ABIC in certain cases; ABIC kept decreasing for decreasing hyperparameter<br />

until the soluti<strong>on</strong> stopped c<strong>on</strong>verging. In such cases the ‘optimal’ soluti<strong>on</strong> must be<br />

chosen manually.<br />

8. C<strong>on</strong>clusi<strong>on</strong>s<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 63<br />

The present work has described two methods to estimate the directi<strong>on</strong>al <strong>wave</strong><br />

energy spectrum <strong>on</strong> the basis <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses. The two c<strong>on</strong>cepts dealt with<br />

are a parametric and a n<strong>on</strong>-parametric method. The Parametric method is based <strong>on</strong> a<br />

summati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> a number <str<strong>on</strong>g>of</str<strong>on</strong>g> parameterised <strong>wave</strong> <strong>spectra</strong>, whereas the n<strong>on</strong>-parametric


64<br />

method, denoted the Bayesian method, estimates the directi<strong>on</strong>al <strong>wave</strong> spectrum in a<br />

number <str<strong>on</strong>g>of</str<strong>on</strong>g> discretised points <str<strong>on</strong>g>of</str<strong>on</strong>g> the <strong>wave</strong> field, which is divided into a set <str<strong>on</strong>g>of</str<strong>on</strong>g> frequencies<br />

and, as well, a set <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>s. For both methods, linear <strong>spectra</strong>l analysis is applied to the<br />

moti<strong>on</strong> measurements to set up equati<strong>on</strong>s <strong>from</strong> which the <strong>wave</strong> energy spectrum may be<br />

determined.<br />

The overall c<strong>on</strong>clusi<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the present work are<br />

The Bayesian and the Parametric estimati<strong>on</strong> method are capable <str<strong>on</strong>g>of</str<strong>on</strong>g> estimating<br />

directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> <strong>from</strong> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses.<br />

It is paramount that at least <strong>on</strong>e <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses has an asymmetric frequency resp<strong>on</strong>se<br />

functi<strong>on</strong> with respect to port/starboard entering <strong>wave</strong>s. It is understood here that<br />

complex-valued transfer functi<strong>on</strong>s are utilised, so that the amplitudes as well as the<br />

phases <str<strong>on</strong>g>of</str<strong>on</strong>g> the resp<strong>on</strong>ses are reflected by the frequency resp<strong>on</strong>se functi<strong>on</strong>.<br />

The uncertainty related to the transfer functi<strong>on</strong>s due to different physical phenomena is<br />

not included in the methodologies.<br />

The present analysis did not suggest to use the roll resp<strong>on</strong>se in favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the sway<br />

resp<strong>on</strong>se, or vice versa. However, in general, it may be noted that the sway resp<strong>on</strong>se has<br />

properties which may be preferred compared to the properties <str<strong>on</strong>g>of</str<strong>on</strong>g> the roll resp<strong>on</strong>se, e.g.<br />

Tannuri et al. [1].<br />

Reas<strong>on</strong>able agreements have been observed for the Bayesian and the Parametric<br />

estimati<strong>on</strong> method when estimated <strong>wave</strong> parameters and directi<strong>on</strong>al distributi<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

energy have been compared with results <strong>from</strong> a <strong>wave</strong> radar system.<br />

It seems difficult to propose the <strong>on</strong>e <strong>ship</strong> resp<strong>on</strong>se-based method in favour <str<strong>on</strong>g>of</str<strong>on</strong>g> the other,<br />

since they perform almost equally well as regards results (for the specific full-scale data<br />

studied here). However, the Parametric method is the most time c<strong>on</strong>suming (due to the<br />

n<strong>on</strong>-linear programming problem) and problems with frequency/directi<strong>on</strong>al overlap <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

<strong>spectra</strong>l peaks may influence the method badly. On the other hand, the Parametric<br />

method yields a soluti<strong>on</strong> which, by definiti<strong>on</strong>, is smooth. In c<strong>on</strong>trast, the smoothness<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> the Bayesian method depends <strong>on</strong> the hyperparameter, <str<strong>on</strong>g>of</str<strong>on</strong>g> which the optimal<br />

value is determined by the ABIC criteri<strong>on</strong> and, in this relati<strong>on</strong>, problems have been<br />

encountered.<br />

Acknowledgements<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

The author heartily thanks Associate Pr<str<strong>on</strong>g>of</str<strong>on</strong>g>essor Toshio Iseki, Tokyo University <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

Marine Science and Technology, for his great willingness to share knowledge <str<strong>on</strong>g>of</str<strong>on</strong>g> related<br />

research subjects. Moreover, the assistance and inspirati<strong>on</strong> provided by Pr<str<strong>on</strong>g>of</str<strong>on</strong>g>essor Jørgen<br />

Juncher Jensen, Technical University <str<strong>on</strong>g>of</str<strong>on</strong>g> Denmark, is highly appreciated.<br />

The analysed full-scale data has been provided by Det Norske Veritas (DNV) and the<br />

author would like to thank Dr. Bo Cerup Sim<strong>on</strong>sen and Mr. Øyvind Lund-Johansen for<br />

making the data available and for valuable discussi<strong>on</strong>s. Posthumous thanks go to<br />

Mr. Andreas Aschim, DNV, for his work <strong>on</strong> the complex-valued transfer functi<strong>on</strong>s.<br />

During the work, a stay at DNV was organised and for this, the European Network <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

Excellence MARSTRUCT should be acknowledged for the financial support.<br />

Finally, it should be menti<strong>on</strong>ed that the overall work, which has lead to the paper, has<br />

been funded by the Eureka project E!2097 MONITUS.


Appendix A. Matrix formulati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> prior distributi<strong>on</strong><br />

The matrix D is composed <str<strong>on</strong>g>of</str<strong>on</strong>g> elements corresp<strong>on</strong>ding to (3.3)–(3.5). From the<br />

organisati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> x, D has the form<br />

2 3<br />

D1<br />

6 7<br />

D ¼ 4 D2 5; sizeðDÞ ¼ð2 K MÞ ðK MÞ. (A.1)<br />

D3<br />

Here D1 is composed as, cf. (3.3)<br />

2<br />

D1 6 0<br />

6<br />

D1 ¼ 6<br />

4<br />

0<br />

D1 . . .<br />

0<br />

3<br />

7<br />

7;<br />

7<br />

5<br />

sizeðD1Þ ¼ðK MÞ ðK MÞ (A.2)<br />

0 D1 with<br />

2 1 0 0 0 1<br />

D1 ¼<br />

1<br />

0<br />

. 2<br />

1<br />

1<br />

2<br />

0<br />

1<br />

. . .<br />

0<br />

0<br />

0<br />

0<br />

. 2<br />

3<br />

6<br />

4<br />

7<br />

7;<br />

7<br />

5<br />

sizeðD1Þ¼K K. (A.3)<br />

1 0 0 0 1 2<br />

Regarding D2 this is determined <strong>from</strong> the following, cf. (3.4)<br />

2<br />

1 2 1 0<br />

3<br />

0<br />

6 0<br />

D2 ¼ 6<br />

4<br />

1 2 1<br />

.<br />

. .<br />

7<br />

5<br />

0 1 2 1<br />

; sizeðD2Þ ¼ðK M 2KÞ ðK MÞ, (A.4)<br />

where<br />

2<br />

1 0<br />

3<br />

0<br />

2<br />

1 0<br />

3<br />

0<br />

6 0<br />

6<br />

1 ¼ 6<br />

4<br />

1<br />

.<br />

. .<br />

7<br />

7;<br />

7<br />

5<br />

2 ¼<br />

6 0<br />

6<br />

2 6<br />

4<br />

1<br />

.<br />

. .<br />

7<br />

7,<br />

7<br />

5<br />

ðA:5Þ<br />

0 1<br />

0 1<br />

sizeðÞ¼size 1 ð 2Þ<br />

¼ K K.<br />

D3 is given by (3.5)<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 65<br />

D3 ¼<br />

1<br />

0<br />

0<br />

0<br />

0<br />

0<br />

0<br />

1 ; sizeðD3Þ ¼2K ðK MÞ (A.6)<br />

with 1 being the K K identity matrix as above.


66<br />

C<strong>on</strong>cerning the vector c in kDx ck2 2 3<br />

0<br />

, this is given as, cf. (3.5),<br />

6 . 7<br />

6 . 7<br />

6 7<br />

6<br />

. 7<br />

6 . 7<br />

6 7<br />

6 7<br />

c ¼ 6 0 7;<br />

6 7<br />

6 s0<br />

7<br />

6 . 7<br />

4 . 5<br />

sizeðcÞ ¼ð2 K MÞ 1, (A.7)<br />

s0<br />

ARTICLE IN PRESS<br />

where s0 fills the same number <str<strong>on</strong>g>of</str<strong>on</strong>g> rows which follows <strong>from</strong> D3.<br />

Appendix B. Directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong>—roll replaced by Sway<br />

270<br />

The estimated <strong>wave</strong> <strong>spectra</strong> for Data A–I are given in Figs. B1–B5.<br />

330<br />

BAYESIAN<br />

0 0.4<br />

30<br />

0.3<br />

300<br />

0.2<br />

0.1<br />

60<br />

270<br />

240<br />

300<br />

240<br />

210<br />

180<br />

BAYESIAN<br />

0 0.4<br />

330<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

150<br />

90 270<br />

120<br />

60<br />

90 270<br />

120<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

300<br />

240<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0.4<br />

30<br />

0.3<br />

(A) (B)<br />

OPTIMISATION<br />

0 0.4<br />

330<br />

30<br />

0.3<br />

210<br />

150<br />

210<br />

150<br />

210<br />

(C) 180<br />

180<br />

(D)<br />

0<br />

180<br />

0.2<br />

0.1<br />

0.2<br />

0.1<br />

150<br />

60<br />

60<br />

90<br />

120<br />

90<br />

120<br />

270<br />

300<br />

240<br />

BAYESIAN<br />

0 0.4<br />

330<br />

30<br />

0.3<br />

180<br />

0.2<br />

0.1<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

OPTIMISATION<br />

0 0.4<br />

330<br />

30<br />

0.3<br />

Fig. B2. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data C and D. Based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}. The <strong>ship</strong><br />

course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.4<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

210<br />

0.2<br />

0.1<br />

180<br />

0.4<br />

30<br />

0.3<br />

Fig. B1. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data A and B. Based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}. The <strong>ship</strong><br />

course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

0.2<br />

0.1<br />

150<br />

150<br />

60<br />

90<br />

120<br />

60<br />

90<br />

120


270<br />

300<br />

240<br />

330<br />

BAYESIAN<br />

0<br />

0.4<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

Appendix C. Measured resp<strong>on</strong>se <strong>spectra</strong><br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

ARTICLE IN PRESS<br />

OPTIMISATION<br />

0.4<br />

30<br />

0.3<br />

210<br />

150<br />

210<br />

150<br />

210<br />

(E) 180<br />

180<br />

(F)<br />

0<br />

0.2<br />

0.1<br />

The <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> for Data A–I are given in Figs. C1–C3.<br />

60<br />

90<br />

120<br />

270<br />

300<br />

240<br />

330<br />

BAYESIAN<br />

0<br />

180<br />

0.4<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0<br />

180<br />

0.4<br />

30<br />

0.3<br />

Fig. B3. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data E and F. Based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}. The <strong>ship</strong><br />

course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

330<br />

BAYESIAN<br />

0 0.4<br />

30<br />

0.3<br />

300<br />

0.2<br />

0.1<br />

60<br />

270<br />

240<br />

210<br />

180<br />

150<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0.4<br />

30<br />

0.3<br />

(G) (H)<br />

0<br />

180<br />

0.2<br />

0.1<br />

150<br />

60<br />

90<br />

120<br />

270<br />

300<br />

240<br />

330<br />

210<br />

BAYESIAN<br />

0<br />

180<br />

0.4<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

150<br />

60<br />

90 270<br />

120<br />

300<br />

240<br />

330<br />

210<br />

0.2<br />

0.1<br />

0.2<br />

0.1<br />

150<br />

OPTIMISATION<br />

0.4<br />

30<br />

0.3<br />

Fig. B4. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data G and H. Based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}. The <strong>ship</strong><br />

course is 0 and the <strong>wave</strong>s are shown as approaching.<br />

270<br />

300<br />

240<br />

330<br />

210<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 67<br />

BAYESIAN<br />

0<br />

180<br />

0.4<br />

30<br />

0.3<br />

0.2<br />

0.1<br />

150<br />

60<br />

120<br />

90<br />

270<br />

300<br />

240<br />

330<br />

210<br />

OPTIMISATION<br />

0.1<br />

0.4 30<br />

Fig. B5. The estimated <strong>wave</strong> <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> Data I. Based <strong>on</strong> the resp<strong>on</strong>ses {heave, sway, pitch}. The <strong>ship</strong> course is 0<br />

and the <strong>wave</strong>s are shown as approaching.<br />

0<br />

180<br />

0.3<br />

0.2<br />

150<br />

60<br />

120<br />

0<br />

180<br />

90<br />

150<br />

60<br />

90<br />

120<br />

60<br />

90<br />

120


68<br />

References<br />

ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69<br />

(A) (B) (C)<br />

Fig. C1. The <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> {heave, roll, pitch} for Data A, B and C.<br />

(D) (E) (F)<br />

Fig. C2. The <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> {heave, roll, pitch} for Data D, E and F.<br />

(G) (H) (I)<br />

Fig. C3. The <strong>measured</strong> resp<strong>on</strong>se <strong>spectra</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> {heave, roll, pitch} for Data G, H and I.<br />

[1] Tannuri EA, Sparano JV, Simos AN, Da Cruz JJ. Estimating directi<strong>on</strong>al <strong>wave</strong> spectrum based <strong>on</strong> stati<strong>on</strong>ary<br />

<strong>ship</strong> moti<strong>on</strong> measurements. Appl Ocean Res 2003;25:243–61.<br />

[2] Hua J, Palmquist M. Wave estimati<strong>on</strong> through <strong>ship</strong> moti<strong>on</strong> measurement. Technical Report, Naval<br />

Architecture, Department <str<strong>on</strong>g>of</str<strong>on</strong>g> Vehicle Engineering, Royal Institute <str<strong>on</strong>g>of</str<strong>on</strong>g> Technology; 1994.<br />

[3] Aschehoug M. Scientific paper <strong>on</strong> the sea state estimati<strong>on</strong> methodology. Technical Report, SIREHNA,<br />

France; 2003 [Paper prepared in the HullM<strong>on</strong>þ project].


ARTICLE IN PRESS<br />

U.D. Nielsen / Marine Structures 19 (2006) 33–69 69<br />

[4] Iseki T, Ohtsu K. Bayesian estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> based <strong>on</strong> <strong>ship</strong> moti<strong>on</strong>s. C<strong>on</strong>trol Eng Pract<br />

2000;8:215–9.<br />

[5] Iseki T, Terada D. Bayesian estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> for <strong>ship</strong> guidance systems. Int J Offshore<br />

Polar Eng 2002;12:25–30.<br />

[6] Waals OJ, Aalbers AB, Pinkster JA. Maximum likelihood method as a means to estimate the directi<strong>on</strong>al<br />

<strong>wave</strong> spectrum and the mean <strong>wave</strong> drift force <strong>on</strong> a dynamically positi<strong>on</strong>ed vessel. Proceedings <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

OMAE2002, Oslo, Norway; 2002.<br />

[7] Isobe M, K<strong>on</strong>do K, Horikawa K. Extensi<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> MLM for estimating directi<strong>on</strong>al <strong>wave</strong> spectrum. Proceedings<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> symposium <strong>on</strong> descripti<strong>on</strong> and modeling <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al seas, vol. A-6; 1984.<br />

[8] Nielsen UD. Estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> <strong>from</strong> <strong>measured</strong> <strong>ship</strong> resp<strong>on</strong>ses. PhD thesis, Secti<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

Coastal, Maritime and Structural Engineering, Department <str<strong>on</strong>g>of</str<strong>on</strong>g> Mechanical Engineering, Technical University<br />

<str<strong>on</strong>g>of</str<strong>on</strong>g> Denmark; May 2005.<br />

[9] Pascoal R, Soares CG, Sørensen AJ. Ocean <strong>wave</strong> <strong>spectra</strong>l estimati<strong>on</strong> using vessel <strong>wave</strong> frequency moti<strong>on</strong>s.<br />

Proceedings <str<strong>on</strong>g>of</str<strong>on</strong>g> OMAE2005, Halkidiki, Greece; 2005.<br />

[10] Benoit M, Goasguen G. Comparative evaluati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> analysis techniques applied to field<br />

measurements. Proceedings <str<strong>on</strong>g>of</str<strong>on</strong>g> ninth internati<strong>on</strong>al <str<strong>on</strong>g>of</str<strong>on</strong>g>fshore and polar engineering c<strong>on</strong>ference, Brest, France;<br />

1999.<br />

[11] Akaike H, Nakagawa T. Statistical analysis and c<strong>on</strong>trol <str<strong>on</strong>g>of</str<strong>on</strong>g> dynamic systems. Tokyo: KTK Scientific<br />

Publishers; 1988.<br />

[12] Data recorded in Tateyama Bay, Japan. The data has kindly been <str<strong>on</strong>g>of</str<strong>on</strong>g>fered by the Tokyo University <str<strong>on</strong>g>of</str<strong>on</strong>g> Marine<br />

Science and Technology. The data originates <strong>from</strong> the research and training <strong>ship</strong> Shioji-Maru.<br />

[13] Bhattacharyya R. Dynamics <str<strong>on</strong>g>of</str<strong>on</strong>g> marine vehicles. New York: Wiley; 1978.<br />

[14] Dimri V. Dec<strong>on</strong>voluti<strong>on</strong> and inverse theory: applicati<strong>on</strong> to geophysical problems. Amsterdam: Elsevier;<br />

1992.<br />

[15] Tarantola A. Inverse problem theory: methods for data fitting and model parameter estimati<strong>on</strong>. Amsterdam:<br />

Elsevier; 1987.<br />

[16] Herbers THC, Guza RT. Estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong> <strong>from</strong> multicomp<strong>on</strong>ent observati<strong>on</strong>s. J Phys<br />

Oceanogr 1990;20:1703–24.<br />

[17] Akaike H. Likelihood and Bayes procedure. In: Bernado JM, De Groot MH, Lindley DU, Smith AFM,<br />

editors. Bayesian Statistics. Valencia: University Press; 1980. p. 143–66.<br />

[18] Press WH, Flannery BP, Teukolsky SA, Vetterling WT. Numerical recipes in FORTRAN77: the art <str<strong>on</strong>g>of</str<strong>on</strong>g><br />

scientific computing, 2nd ed. Cambridge: Cambridge University Press; 1992.<br />

[19] Goda Y. Random seas and design <str<strong>on</strong>g>of</str<strong>on</strong>g> maritime structures. Advanced series <strong>on</strong> ocean engineering, vol. 15.<br />

Singapore: World Scientific; 2000.<br />

[20] L<strong>on</strong>guet-Higgins MS, Cartwright DE, Smith ND. Observati<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> the directi<strong>on</strong>al spectrum <str<strong>on</strong>g>of</str<strong>on</strong>g> sea <strong>wave</strong>s<br />

using the moti<strong>on</strong>s <str<strong>on</strong>g>of</str<strong>on</strong>g> a floating buoy. Ocean Wave Spectra 1961; 111–36.<br />

[21] Hogben N, Cobb FC. Parametric modelling <str<strong>on</strong>g>of</str<strong>on</strong>g> directi<strong>on</strong>al <strong>wave</strong> <strong>spectra</strong>. Proceedings <str<strong>on</strong>g>of</str<strong>on</strong>g> 18th <str<strong>on</strong>g>of</str<strong>on</strong>g>fshore<br />

technology c<strong>on</strong>ference, Houst<strong>on</strong>, Texas 1986.<br />

[22] Mitsuyasu H, Tasai F, Suhara T, Mizuno S, Onkusu M, H<strong>on</strong>da T, et al. Observati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> the directi<strong>on</strong>al<br />

spectrum <str<strong>on</strong>g>of</str<strong>on</strong>g> ocean <strong>wave</strong>s using a cloverleaf buoy. J Phys Oceanogr 1975;5:750–60.<br />

[23] Price WG, Bishop RED. Probalistic theory <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong> dynamics. L<strong>on</strong>d<strong>on</strong>: Chapman & Hall; 1974.<br />

[24] Bendat JS, Piersol AG. Random data—analysis and measurement procedures, 3rd ed. New York: Wiley;<br />

2000.<br />

[25] Jensen JJ. Load and global resp<strong>on</strong>se <str<strong>on</strong>g>of</str<strong>on</strong>g> <strong>ship</strong>s. Elsevier ocean engineering book series, vol. 4. Amsterdam:<br />

Elsevier; 2001.<br />

[26] WAVEX—principles <str<strong>on</strong>g>of</str<strong>on</strong>g> operati<strong>on</strong>. Technical documentati<strong>on</strong> by MIROS; 1998.<br />

[27] Neumaier A, Schneider T. Estimati<strong>on</strong> <str<strong>on</strong>g>of</str<strong>on</strong>g> parameters and eigenmodes <str<strong>on</strong>g>of</str<strong>on</strong>g> multivariate autoregressive models.<br />

ACM Trans Math S<str<strong>on</strong>g>of</str<strong>on</strong>g>tware 2001;27(1):27–57.<br />

[28] Lloyd ARJM. Seakeeping, 2nd ed. Chichester: Ellis Horwood; 1998.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!