Beruflich Dokumente
Kultur Dokumente
Parametric models. In WAFO, we have implemented certain Switching Markov models. In some applications, one wants
models for distributions of wave characteristics found in the to model data, whose properties change according to an under-
literature. For example, one nds lying, often unobserved process, called the regime process. The
state of the regime process controls which parameters to use and
approximations of the density (Tc ; Ac ) in a stationary when to switch the parameter values. The regime process can
Gaussian transformed process (Cavanie, Arhan and Ezraty be modelled by a Markov chain, and this is the fundamental
(1976), Longuet-Higgins (1983)) basis for the set of routines presented. For an application with
a model for the cdf/pdf of breaking limited wave heights switching Markov models for fatigue problems, cf. Johannesson
(Tayfun (1981)) (1998), (1999).
a model for the cdf/pdf of large wave heights (Tayfun Fatigue and Damage. In WAFO, routines for calculation
(1990)) of the ackumulated damage ackording to the Palmgren-Miner
Simulation of random processes. Ecient0 simulation of rule have been implemented. It is possible to compute the total
a Gaussian process X (t) and its derivative X (t) , given the damage from a cycle count as well as from a cycle matrix.
spectral density or the auto-correlation function, can be per-
formed. A routine for simulation of a transformed Gaussian Plot routines, cycles and fatigue. This module contains
process (and its derivative) is also included. For fast and exact routines for visualization of cycle counts and cycle matrices and
simulation, some routines use a technique with circulant em- various histograms.
bedding of the covariance matrix, Dietrich and Newsam (1997).
More traditional spectral simulation methods (FFT) are also Miscellaneous routines. We nd here various plot rou-
used. Simulation of discrete Markov chains, Markov chains of tines, algorithms for numerical integration, and functions for
turning points, switching Markov chains etc. is possible. documentation of WAFO with modules.
Statistical tools and extreme value distributions. Certain
probability distributions are extensively used in ocean engineer-
ing, e.g. Rayleigh, Gumbel, Weibull. The generalized extreme- SOME APPLICATIONS OF WAFO
value distributions (GEV) and generalized Pareto distributions
(GPD) are also important. For the distributions mentioned, it In this section we demonstrate in examples some of the capabil-
is possible to estimate parameters, generate random variables, ities of WAFO. For further examples and knowledge about the
evalute pdf and cumulative distribution function, and plot in algorithms used in the routines, we refer to the tutorial and the
various probability papers. One cathegory of routines handles documentation in the routines. Note that the necessary Matlab
bivariate distributions. Besides having routines for estimation code for generation of the gures in this paper is found in the
of parameters etc. for the two-dimensional Weibull distribution, directory WAFO/paper.
bivariate modelling is possible. We start by dening a frequency spectrum, S (!) , which
will be used in many of the examples; we choose a Torsethaugen
Kernel-density-estimation tools. The routines in this spectrum with the parameters Hm0 = 6 m, Tp = 8 s, describing
cathegory complement the ones found in 'Data analysis' and, signicant wave height and primary peak period, respectively.
obviously, the routines in 'Statistical tools and extreme value The energy is divided between two peaks, corresponding to con-
2
Wave amplitude (m)
0.25
1
0 0.2
1
2 0.15
3
0.1
4
5
0 50 100 150 200 250 300 0.05
Time (s)
Figure 1. Part of a simulation from S (!) , a Torsethaugen
spectrum with Hm0 = 6 m, Tp = 8 s. Total number of points 0
0 2 4 6 8 10
= 1000 , t = 1:05 s. The rst 300 seconds are shown. Wave period, downtoup crossing (s)
Figure 3. Solid line: pdf for wave period, given S (!) . Dash-
A common situation is data given in form of a time series, dotted line: pdf for wave period, given Sest (!) . The histogram
for which one wants to estimate the related spectrum. From the shows the wave periods extracted from simulated data.
realization in Figure 1, we will now nd an estimate Sest (!)
and compare the result to the original Torsethaugen spectrum
S (!) . In Figure 2, the two spectra are displayed. The maximum
lag size of the Parzen window function used can be chosen by Directional spectra
the user or automatically by WAFO.
Spectral density In WAFO one nds means for evaluation and visualization of
7 fp = 0.79 (rad/s)
directional spectra, that is
6
S (!; ) = S (!)D(; !)
5
S(w) (m s/rad)
3
ing function. A number of common spreading functions can be
chosen by the user.
2 One way of visualizing S (!; ) is a polar plot. In Figure 4
we show the resulting directional spectrum (solid line) for the
1 Torsethaugen spectrum used above. The spreading function is
of the cos-2s type, that is (in the frequency dependent case)
0
0 0.5 1 1.5 2 2.5 3
Frequency (rad/s)
Figure 2. Solid: original spectrum. Dashed: spectrum esti- D() = 2p (s(s++1)1=2) cos2s 2
mated from data (1000 observations) and 95% condence inter-
val (dotted). Maximum lag size of the Parzen wind = 80 .
Probability distributions of wave characteristics with s=15 . Note that the two peaks can be distinguished. The
dash-dotted line is the corresponding result when the spread-
WAFO gives the possibility to compute exact probability dis- ing function is frequency dependent, cf. Krogstad and Barstow
tributions for a number of wave characteristics, given a spectral (1999).
150 30
Tc crest period
0.5
180 0
10
30
50
70
240 300 90
95
270 99
10 99.9
amplitude [m]
Torsethaugen spectrum and the spreading function is of cos-2s
type with s = 15 . Solid line: directional spectrum. Dash-dotted
line: directional spectrum, using frequency dependent spreading
function. 5
density (solid line) has been calculated given an estimated spec- Figure 6. Solid line: theoretical joint density of Vcf and 2Ac .
trum with two peaks. The data are marked as dots. Further, Dashed line: kernel density estimate of Vcf and Hd . Dots:
the kernel density estimates (dashed line) are estimated from the data from the North Sea.
square root of the data using an Epanechnikov kernel and the
one-dimensional
p optimal
p value for the smoothing parameter for
both Vcf and Hd , see also Silverman (1986). In these cases
it also ensures that the kernel density estimate is zero for nega-
tive values due to the compact support of the kernel used. The
nal estimate is then obtained by a simple change of variables.
For the kernel density estimate in Figure 6, the contour levels of
the kernel density estimates enclose the given percentage of the
data and have been calculated by splitting the bi-variate data,
(Vcf ; Hd ) , into groups formed by ranking the data by density
estimate.
Time
Tu
estimated from ( Ac ; Tcf ) using an Epanechnikov kernel and
Figure 5. Wave parameters, denitions. In the list below, the one-dimensional optimal value for the smoothing parameter.
0
0 1 2 3 4 5 mk = mrfc
k
period [s]
m+k
Figure 7. Solid line: theoretical joint density of Tcf and Ac Figure 9. Denition of the rain
ow cycle, as given by Rychlik
given that Tc = 5 . Dashed line: kernel density estimate. Dots: (1987). From each local maximum Mk one shall try to reach
data from the Japan Sea. above the same level, in the backward(left) and forward(right)
directions, with an as small downward excursion as possible. The
minimum, of mk and m+k , that represents the smallest devi-
Spatio-temporal waves. Distribution of extremal crest ation from the maximum Mk is dened as the corresponding
heights. rain
ow minimum mrfc k . The k :th rain
ow cycle is dened as
Consider spatial waves evolving in time, spatio-temporal waves (mrfc
k ; Mk ) .
or space-time waves, which, from a probabilistic point of view, The Markov model is dened by the min-max pdf, which is
are modelled as Gaussian random elds W (x; t) . A wave of obtained from the power spectral-density by using approxima-
this type is called an extremal wave if its crest height attains tions in Slepian model processes, see e.g. Lindgren and Rych-
a local maximum in time. Podgorski et al. (1999) discussed lik (1991) and references therein. For the Markov model
dierent aspects of extremal waves, and computed the joint den- there is an explicit solution for the intensity of rain
ow cy-
sity function of extremal crest height and half-wave length. A cles, see Frendahl and Rychlik (1993). By using the rou-
comparison of the result for one-dimensional (spatial) waves and tines in WAFO the intensity of rain
ow cycles can be found
spatio-temporal waves was made; cf. Figure 8. The spectrum using Markov approximation; see Figure 10, where also the
used as input was a JONSWAP spectrum with frequencies cut rain
ow cycles found in the simulated load signal are shown.
o outside (0:5!p ; 1:5!p ) and signicant crest height 3.8 m. 6
8 8
7 7
6 6 4
5 5
H1 [m]
H1 [m]
4 4 2
3 3
max
2 2
0
1 1 Level curves at:
0 4.2994e05
0 50 100 150 200 250 0 50 100 150 200 250 8.5987e05
L1 [m] L [m]
1 2 0.00017197
0.00042994
and half-wave length for spatial waves (left) and extremal waves
0.0034395
4 0.0051592
In fatigue applications the exact sample path is not important, Figure 10. Intensity of rain
ow cycles computed from the psd
but only the tops and bottoms of the load, called the sequence of through Markov approximation, compared with the cycles found
turning points (TP). From the TP cycles can be extracted, from in the simulation.
0
butions in Ocean Wave Spectra. Proceedings of the 9th ISOPE
Conference, Vol III, pp. 79-86.
Level curves at:
0.0010513
0.0021027
2
Lindgren, G. and Rychlik, I. (1991). Slepian models and re-
0.0042054
0.010513
0.021027
0.042054 gression approximations in crossing and extreme value theory.
4
0.084108
0.12616
0.16822
International Statistical Review 59, 195-225.
Longuet-Higgins, M.S. (1983). On the joint distribution wave
6 periods and amplitudes in a random wave eld. Proc. R. Soc.
6 4 2 0
min
2 4 6 A389, pp 24{258.
Figure 12. Intensity of trough-crest cycles computed from the Podgorski, K., Rychlik I. and Machado, U. E. B. (1999). Exact
psd through Markov approximation, compared with the cycles Distributions for Apparent Waves in Irregular Seas, 1-54. Ocean.
found in the simulation. Engng. to appear.
Podgorski, K., Rychlik, I., Ryden, J. and Sjo, E. (1999). How
ACKNOWLEDGEMENTS big are the big waves? Proceedings of the 9th ISOPE Conference,
Vol III, pp. 53-60.
Many persons have put a great deal of eort into WAFO and its
predecessors FAT and WAT. Among others, we are grateful to Rychlik, I. (1987). A new denition of the rain
ow cycle count-
Mats Frendahl, Finn Lindgren, and Ulla Machado, all at Math- ing method. Int. J: Fatigue 9, 119-121.
ematical Statistics at Lund University, and Sylvie van Iseghem,
IFREMER, Brest. Rychlik I. and Lindgren G. (1993). CROSSREG { a tech-
The wave data from Gullfaks C were prepared and nique for rst passage and wave density analysis. Probability in
made available by Dr. S. Haver, Statoil. The wave data the Engineering and Informational Sciences 7, 125-148.
from Yura, Japan Sea were prepared and made available by
Dr. Sc. H. Tomita, Ship Research Institute, Ministry of Trans- Rychlik I. and Lindgren G. (1995). WAVE Analysis Toolbox
port. { a Tutorial. Dept. of Math. Stat., Lund University.
The rst author was nancially supported by the Norwe-
gian Research Council (NFR), Det Norske Veritas and Lund Uni- Rychlik I., Johannesson P., Leadbetter M. R. (1997). Modelling
versity. and Statistical Analysis of Ocean-Wave Data using transformed