Documente Academic
Documente Profesional
Documente Cultură
INTRODUCTION
In the study of the dynamic behavior of floating structures in
irregular waves, two steps are of great importance: the digital simulations of the sea surface elevations, and the spectral analyses of
the generated sea waves time series. And recently great effort has
been spent to successfully represent ocean waves through computational programs, and several methods of spectral analysis and
statistical calculations have been developed to analyze the wave
records. In the present work, some of these methods of wave simulation and spectral analysis are shown and applied in the study of
the motions of floating structures.
Sea waves are of stochastic nature and, mathematically, are
represented as Gaussian stationary and ergodic processes. The
models used to represent these processes are based on the work of
Rice (1944, 1945), who presented two formulations
(Deterministic Spectral Amplitude and Nondeterministic Spectral
Amplitude) to represent electronic signals as Gaussian stationary
processes from previously specified energy spectra. These methods are largely used in the simulation of irregular sea waves and
are known as the Wave Superposition or the Summation models;
they were applied by Morooka and Yokoo (1992) for the simulations of directional waves. Borgman (1969) applied these models
and introduced the simulation through white noise filtering in
time domain. Shinozuka (1971) has shown a method to simulate
a class of multivariate and multidimensional random processes
with specified cross-spectral densities. Later, Wittig and Sinha
(1975) presented a method of synthesis of random processes using
Fast Fourier Transforms (FFTs) of the spectrum in frequency
domain. This method was improved by Hudspeth and Borgman
(1979) with the application of FFT routines for real numbers.
Simulated or measured time domain wave records can be used
for the estimation of wave spectra. There are several methods for
spectral estimation; the most used are the Traditional Methods
based on Fourier transforms of the signals in time domain. The
Parametric Methods are another well-known family of methods
that include the AR, MA and ARMA. Other methods used in signal processing are the Maximum Likelihood Method (MLM) and
Received March 10, 1995: revised manuscript received by the editors
April 12, 1997. The original version (prior to the final revised
manuscript) was presented at the Fifth International Offshore and Polar
Engineering Conference (ISOPE-95), The Hague, The Netherlands,
June 11-16, 1995.
KEY WORDS: Sea waves, simulation of irregular waves, wave statistics,
floating platform dynamics.
(a cosw t + b sinw t)
n
(1)
n =1
where an and bn = independent random variables that are distributed normally around zero with the standard deviation =
s n = S(w n )Dw ; S(w) = specified wave spectrum; and w = 2pf
= frequency (f > 0).
The Fourier coefficients are calculated by filtering the white
noise spectrum with the specified one. As the white noise spectrum is unitary, the filtered spectrum is equal to the specified
spectrum. The coefficients an and bn are random variables, so Eq.
1 is called Nondeterministic Spectral Amplitude (NSA). The second model for Gaussian signals introduced by Rice is represented
by:
h (t ) =
A cos(w t - e )
n
(2)
n =1
where An = 2 S(w n )Dw ; and e = phase angle distributed at random (0 < e < 2p).
190
N n=1
N
N
2p (n -1)( j -1)
Cn = h ( j )exp-i
,
N
j =1
h ( j) =
j = 1,2,, N
(6)
n = 1,2,, N
(7)
n = 2,3,,
N
+1
2
(8)
an = a n [ S(w n )Dw ]
(9)
1/ 2
bn = bn [ S(w n )Dw ]
(10)
Rn = (an2 + bn2 )
(3)
bn
an
(4)
N
Cn = N S(w n-1 )Dw exp(iq n-1 ), n = 2,3,, +1
2
2
(5)
q n = arctan
1/ 2
R cos(w t +q )
n
n =1
(11)
STACKED FFT
The wave simulations are efficiently carried out with the initialization of the complex Fourier coefficients in frequency domain,
followed by the inversion of FFTs to generate time sequences, as
described previously. An important characteristic of the generated
records is that the sea surface elevations are sequences of real
numbers in time domain. Thus, the complex Fourier coefficients
present even real components and odd imaginary components in
frequency domain, and the application of FFTs produces complex
numbers with imaginary components equal to zero in time
domain. Based on this property, Hudspeth and Borgman (1979)
developed a method of initialization of the Fourier coefficients
that uses just half the computer memory for the FFT synthesis
than for the method presented previously.
In frequency domain, the Fourier coefficients (Cn) are modified
to be represented by a new sequence of complex coefficients (Tk)
producing a stacked Fourier pair:
4p (k -1)(l -1)
2 N /2
Tk exp i
,
N k =1
N
N /2
4p (k -1)(l -1)
Tk = h (l )exp-i
,
N
l =1
h (l ) =
l = 1,2,,
N
2
(12)
N
2
(13)
k = 1,2,,
2pk
+ cos
( IC + ICN / 2-k )
N k
2pk
1
ITk = ICk + ICN / 2-k - sin
( IC - ICN / 2-k )
N k
2
2pk
- cos
( RCk + RCN / 2-k )
N
N
k = 2,3,,
2
RTk =
1
( RC1 + RCN / 2 )
2
1
IT1 = - ( RC1 - RCN / 2 )
2
(14)
191
WALSH FUNCTIONS
The development of digital computers leads to the application
of orthogonal functions in the solution of numerical problems.
Sine and cosine orthogonal systems are of fundamental importance in these theories, and the Fourier analyses are the most
important mathematical tools in signal processing and analysis.
Another well-developed family of orthogonal functions with the
evolution of computers (including hardware) is that of the binary
functions, which represent series of rectangular pulses or square
waves. The Walsh functions form an ordered group of rectangular
waves with only two possible amplitude values: 1 or +1. The
Walsh functions can be divided in odd and even functions:
WAL(2 k,t ) = CAL(k,t ),
WAL(2 k -1,t ) = SAL(k,t ),
k = 1,2,, N /2
These functions define two new Walsh series similar to sines
and cosines. The CAL and SAL functions are of simple representation (specially for hardware) and can easily represent stepped
functions. However, the trigonometric functions present better
accuracy when representing smooth functions. More details about
Walsh functions and their applications are presented by
Beauchamp (1984).
The simulation models represented by Eqs. 1 and 2 can be
applied through substitution of the sines and cosines by SAL and
CAL functions, as suggested by Totsuka (1993). However, the direct
summation of rectangular functions produces stepped waves. This
problem can be attenuated through the representation of sines and
cosines with summations of Walsh functions with precalculated
amplitudes, as proposed by Kitai (1975). This technique produces
wave records close to those produced by sines and cosines.
(15)
PPER ( f ) =
(16)
(17)
Through inversion to time domain, the discrete sequence contained in both real and imaginary parts represents a random process with the same length as that generated from the nonstacked
method. The advantage is that the stacked FFT requires just half
the computational memory of the method presented before.
(18)
N
1
Dt h j exp( -i 2pfjDt )
NDt j =1
(19)
192
ent to the spectral estimation via FFT. The effects of leakage can
be reduced with the application of windows of nonuniform
weights on the wave records. In this work the Hanning window
was used:
h (t ) =
1
2pt
1- cos
2
T
(20)
where apk are the predictor parameters and sp2 is the prediction
error power. The determination of the order of the predictor, the
coefficients apk and the constant sp2 are the crucial points of the
maximum entropy spectral estimation.
Andersen (1974) presented an algorithm to determine the coefficients apk based on the method proposed by Burg, through the
recursive equation:
apk = ap1k appap1pk
where ap0 = 1and apk = 0 for k p, and:
N-p
a pp =
1+ a pk exp( -i 2pfkDt )
k =1
t =1
N-p
2
pt
2
pt
+ b
(23)
t =1
2 bpt bpt
(b
PMEM ( f ) =
(22)
(24)
(25)
(26)
1 N 2
ht
N t =1
(27)
N - ( p +1)
(28)
where the selected order p is that for which the FPE is minimum.
In Marple (1987) some other methods can be found, but no consensus has been reached about their efficiency. Holm and Hovem
(1979) suggest a test with the partial autocorrelation function
coefficients:
Rxx (t ) =
R (t - k )a
xx
tk
(29)
k =1
where Rxx(0). Houmb and Overvik (1981) call the partial autocorrelation coefficients of PAKF and suggest that their values can be
compared with the variance of the partial auto-correlation, given
by:
(21)
s2 =
1
N
(30)
193
records were simulated using the Wave Superposition, FFT inversions and Walsh functions. The specified wave spectrum is the
ITTC with Hs = 7.5 m and Ts = 10 s. Fig. 1 shows part of a simulation with the DSA model with Wave Superposition and with
Walsh functions. It can be noted that the simulation via Walsh
functions produces a record very close to that produced via Wave
Superposition.
Fig. 2 shows the results of the simulations via Wave
Superposition and via Walsh functions applying the NSA model.
The differences between the records are due to the substitutions of
the sine and cosine functions by the summations of Walsh functions.
Fig. 3 presents the simulated record using the DSA model via
FFT inversion. The application of the FFTs has shown it to be the
fastest method. The Wave Superposition simulations have spent
twice the computer time of the FFTs, while the Walsh functions
were up to 20 times slower. The application of the stacked FFT
has produced wave records exactly equal to the standard FFT and
without any advantage in computation time. However, the stacked
FFT routine requires about half the computer memory.
Wave Statistics
RESULTS
Wave Simulation
Following the formulations shown above, irregular wave
X
XMAX
XMIN
s
HMEAN
H1/3
H1/10
T0
TP
SK
K
(m)
(m)
(m)
(m)
(m)
(m)
(m)
(s)
(s)
SUPER
0.318E-5
6.212
6.305
1.872
4.623
7.179
8.881
9.648
6.943
0.0430
2.887
WALSH
0.369E-3
6.403
6.918
1.934
3.750
6.820
8.973
6.011
1.964
0.0173
2.879
FFT
0.589E-6
6.604
6.935
1.875
4.604
7.164
8.878
9.422
6.030
0.0011
2.862
194
SPECTRAL ANALYSIS
Complementing the statistical analyses, the simulated wave
records were used to estimate the wave spectra, which can be
compared with the specified one (ITTC with Hs = 7.5m and Ts =
10 s). By applying the Direct Method with Hanning spectral windowing, averaging 15 blocks of 1024 points with overlapping of
half a block, the following spectra were estimated. From the estimated spectra, the moments are calculated and can be compared
with those from the specified spectrum, shown in Table 2.
M0
M1
M2
M3
M4
(m2)
(m2/s )
(m2 /s2)
(m2/s3 )
(m2/s4 )
3.513
2.156
1.536
1.351
1.549
195
X
XMAX
XMIN
s
HMEAN
H1/3
H1/10
T0
TP
(m)
(m)
(m)
(m)
(m)
(m)
(m)
(s)
(s)
SK
K
DSA
0.656E-2
2.546
2.420
0.786
1.889
2.837
3.502
13.63
10.97
0.0555
2.823
NSA
0.905E2
2.865
2.968
0.966
2.195
3.185
3.827
16.75
10.62
0.0106
2.642
REFERENCES
DISPLACEMENT
LENGTH
BEAM
DRAFT
40000 ton
115 m
61.5 m
23.5 m
CONCLUSIONS
The present work discusses the main methods of irregular
waves simulation and wave records analysis, for DSA and NSA
models. The method of the inversion of FFT has shown itself to
be computationally more efficient, being two times faster than the
Wave Superposition. The Walsh functions have not been competitive with the other methods; however, the study of these functions is suggested by references to application in hardware for
196
Mail or fax to: ISOPE Membership Dept., P.O. Box 1107, Golden, Colorado 80402-1107, USA
(Fax +1-303-420-3760)