Anda di halaman 1dari 8

Proceedings of The Fifteenth (2005) International Offshore and Polar Engineering Conference Seoul, Korea, June 1924, 2005

Copyright 2005 by The International Society of Offshore and Polar Engineers ISBN 1-880653-64-8 (Set); ISSN 1098-6189 (Set)

Surge Motion of Mini TLP in Random Seas- Comparison between Experiment and Theory
Nungsoo Kim and Cheung Hun Kim
Department of Civil Engineering, Ocean Engineering Program, Texas A & M University College Station, TX, USA

ABSTRACT
The objective of the paper is an analysis of the model test data of a mini-TLP, which includes the quality and degree of non-Gaussianity of the measured data; comparison of Gaussian and Non-Gaussian method for the estimation of QTFs. The comparison between the two methods is made with respect to the energy density spectra, coherency function and the reconstructed response time series. A special consideration is given on the slow drift surge in regard to the mean surge value and surge natural period. It was found that the surge response is practically linear in contrast to the other conventional TLPs responses.

fact well known that the mean value in the quadratic response is the most important data for the design of offshore structures. However the mean varies in time. It is necessary to find the reasonably long sample length to reach the statistically stable condition. However, the sample was limited to 3 hours long.

MODEL TEST
Mini TLP concept was designed for deployment off the coast of West Africa with a relatively benign environment. This platform is referred to as a Mini TLP because its maximum dimension in the horizontal plane is less than the outside diameter of a vertical hull column of the Heidrun TLP. Model tests of the Mini TLP were conducted in the OTRC wave basin to qualify its wave loads, air-gaps, motion responses, and tensions in its tendons and risers. The experiments used a 1:40 scale model of the Mini-TLP. The experiments were divided into two parts: fixed model tests and compliant model tests. This analysis takes into account the surge motion and uses the compliant test data in random sea condition. In this compliant model test, the Mini TLP was treated in a conventional way, although the riser system was reduced to 4 risers instead of 12 in the original mini TLP design. A spring was inserted in each riser and tendon model to match the axial stiffness of the prototype risers and tendons. In the tests, both ends of the risers and tendons were hinged to the floor bottom or the TLP. The target wave spectrum was chosen to simulate a 100-year West Africa stormy sea. It is a JONSWAP type spectrum of a peakedness factor of 2.0, peak period of 16 seconds, and significant wave height of 3.9m. Note that the sea severity is very low and sample length is 3 hours. The principal particulars are listed in Table 1, in which we find the measured natural surge period, which will be compared with the result of analysis of the measured surge time history in the current work. Table 1. Principal particulars of the tested mini-TLP model in terms of prototype scale As-built Particulars Full scale value Water depth 673.61 m draft 28.51 m Column diameter 8.64 m

KEY WORDS: cross-bi-spectral estimates, linear transfer function


(LTF), quadratic transfer function (QTF), Volterra quadratic model, Gaussian and non-Gaussian method, mini TLP.

INDRODUCTION
The test data of the mini-TLP was reported by Niedzwecki et al. (2001). The non-Gaussian method for cross-bi-spectral analysis of the mini-TLP was given by Birkelund et al (2003) using non-Gaussian method. The foregoing cross-bi-spectral analysis was retreated by Kim (2004) using both Gaussian and non-Gaussian method, which is similar to the analysis of the surge exciting force on a barge fixed in the head random seas (Kim and Kim, (2004). In the present research we have found that the slow surge drift of such a small mini-TLP in the Gaussian low seas is largely different from the slow drift of conventional large TLP which is usually tested in high non-Gaussian seas. Thus, the test data has attracted our interest in the analysis of the slow drift surge. It is desirable to confirm the quality of the wave data to be acceptable for the analysis. We found that the wave was found to be nearly Gaussian. Similarly the surge response was nearly Gaussian. Since the mini-TLP is a compliant structure, the data has to produce the long surge period. But it was not easy to observe in the original data. Our next interest was to find the natural surge period, mean surge and peak surge amplitude. Birkelund, et al. (2003) conducted higher-order spectral analysis and identification of the mini-TLP system, employing the non-Gaussian method However; they did not consider the mean surge value and surge period. The major objective of this kind of experiment is supposed to be in finding the characteristics of the slow drift motion including the mean offset, natural surge period and distribution of the peak response. It is in

Paper No. 2005-EJP-01

Kim

244

1 /8

Pontoon height Pontoon width Column center-to-center Total weight of mini-TLP Vessel displacement Tendon & riser pretension Center of Gravity (X) Center of Gravity (Y) Center of Gravity (Z) Pitch radius of gyration Roll radius of gyration Yaw radius of gyration Surge natural period

6.22 m 6.22 m 28.51 m 6445 metric ton 10158 metric ton 3713 metric ton 0 0 29.5 m 21.7 m 21.9 m 17.2 m 139.97 sec

Table 2 Statistics of waves and surge motion No. of oscillations Mean(m) Skewness Wave 968 0.0195 0.0522 Surge 689 0.3698 -0.1627

Kurtosis 3.1405 3.1184

The data in the table indicates that the wave is Gaussian while the response surge is slightly non-Gaussian Probability density distributions of the wave and surge The histograms are constructed and compared with the Gaussian distributions in Fig.2. From the figs, we may conclude that the wave and response are Gaussian. We have plotted the histograms of the positive peaks as shown in Fig. 3. It shows that the distributions of crest heights of waves and surge motion compare less well with the Rayleigh distributions, which might imply that the data are Gaussian but are not quite narrow band or it comes from some error due to the contamination of the waves in the low and high frequency region. Fig. 4 illustrates the Rayleigh distribution compared with the probability of exceedence of crest heights of waves and surge motion. Both crest heights of waves and surge motion appear to be slightly deviated from Gaussian waves and the wave and response are practically regarded Gaussian.
0.5

PRELIMIMINARY INVESTIGATION OF MEASUED DATA It is desirable to test the quality of the measured data before going to the cross-bi-spectral analysis. The quality of the input wave may be tested by comparing the variance of the target and measured energy density spectra. The Gaussianity test may be carried out by comparing the histograms with the Gaussian probability density distribution and Rayleigh density of positive peaks. In addition to these, one may observe the probability of exceedence with the Rayleigh distribution. If these are very close to each other the wave and response data are apparently Gaussian. Energy density spectra of wave and surge

0.5

0.4

0.4

p((x - m)/ )

p((y - m)/ )

0.3

0.3

0.2

0.2

0.1

0.1

The energy density spectra of the wave and surge motion are shown in Fig. 1. The target and measured spectra are in good agreement according to the observation. Since the variances are different from each other within 11 % error, it may be acceptable compared with (Kumar et al., 2002). From the data we presume the surge motion may be in the same range of the error. The surge motion energy density spectrum in Fig.1 has a peak in the low frequency region representing the 2nd-order low-frequency surge motion that is the result of the 2ndorder difference frequency interaction.
7 6 5
2

-4

-3

-2

1 0 -1 (x - m )/
x

-4

-3

-2

1 0 -1 (y - m )/
y

Fig. 2 Probability density of wave elevation (left) and surge motion (right) with Gaussian distribution (solid line)
0.8 0.7 0.6 0.5
p( H )

0.8 0.7 0.6 0.5


p( H )
c

0.4 0.3

0.4 0.3 0.2 0.1 0 0

7 6

0.2

5
U( ) (m s)
0.1

U( ) (m *s)

4 3 2 1 0 0 0.5
(rad/s)

4 3 2 1
0 0 1 2 H (m)
c

2 H (m )
c

Fig. 3 Distribution of crest height of waves (left) and surge motion (right) compared with Rayleigh distribution (solid line)
0.2 0.4 0.6 0.8 1 1.2 1.4
10
0

1.5

0 0

(a) (b) Fig. 1 (a) Measured wave spectrum (solid), target spectrum (dotted): JONSWAP, Hs = 4m, Tp =16 seconds, =2, (b) measured surge motion The principal interest of our study is to investigate the data if they are linear or nonlinear. The state of the sea utilized in the study belongs to the class of rough sea (Hs=2.54.0), which is regarded as Gaussian and the 2nd-order nonlinearity is negligible (Kumar etal., 2002). The surge motion energy density appears to be remarkably weak in the given time series. Thus, we have evaluated the mean, skewness and kurtosis as listed in Table 2.
P

(rad/s)

10

10

-1

10

-1

10

-2

10

-2

10

-3

10

-3

10

-4

x / x

10

-4

y / y

Fig. 4 Rayleigh distribution compared with the probability of exceedence of positive peak of wave (left) and surge motion (right)

Paper No. 2005-EJP-01

245 Kim

2 /8

Convergence of mean surge motion The mean of the surge in the random sea is given by
E [surge(t ) ] = G2 ( ,- )U xx ( )d
0

1.2

0.8
Surge(m)

(1)

0.4

which assumes infinitely long data, though the wave data are limited to 3 hours. It is expected that the QTFs for the mean have been affected by the sample length. Thus, it is worth to see the effect of sample length. In the time domain the mean surge motion is examined by the mean cumulative average defined as; ycum (t N ) = 1 N y (t i ) N i =1 (2)

-0.4 6500

7000 Time(s)

7500

8000

Fig. 6 Low pass filtered surge motion Table.3 Surge natural frequency Measured data Surge natural period(sec) 139.97 Present analysis of measured data 165.90

Fig. 5 shows that the mean surge motion has not reached a steady value in 3-hours. It requires more time to arrive at the unchanging mean value. The given data length for this experiment is much shorter than the data length proposed by Kim and Boo (1990).
Mini TLP 0.5

It is concluded that the 2nd-order slow drift is unusually small compared to the conventional TLPs. This is due to the Gaussian wave input that gives negligibly small 2nd-order waves, and forces which cannot give enough effect to produce 2nd-order surge.

0.4
Mean surge (m)

ESTIMATION OF LTFs AND QTFs


We will employ the Volterra quadratic system model (Volterra, 1942) to estimate the QTFs for the 2nd-order surge motion. There are two methods for the estimation; the Gaussian and non-Gaussian method. The former assumes the input is Gaussian, while the latter assumes the input is non-Gaussian. Besides the former approach employs the Blackman-Tukey method for the spectral estimates, whereas the latter the ensemble average method (Bendat, 1991). Volterra quadratic model The Volterra model was proposed by Hasselmann (1966) to be used for the analysis of the nonlinear ship motion in the seas. Application of the model was given by Dalzell (1976) with Gaussian input assumption, Kim and Powers (1987) gave a theory with non-Gaussian input assumption. Bendat (1990) gives a general theory. Kim and Boo (1990) and Kim and Kee (1992) simulated the slow drift force following Dalzell (1976), Stansberg (1997) similarly treated, though slightly different, to the Dalzell (1976). Kim and Kim (2003; 2004) have recently employed the Gaussian and non-Gaussian method. The Volterra quadratic model is originally a time domain model: y ( t ) = g1 ( t1 ) x ( t t1 ) dt1 + g 2 ( t1 , t2 ) x(t t1 ) x (t t2 )dt1dt2 (3)

0.3

0.2

0.1

0 0

1 Time (hour)

Fig. 5 Cumulative mean of surge motion The low frequency surge motion and natural surge period from experimental data We are apt to think the slow drift (2nd-order) motion of any compliant TLP has to be large compared to the linear response However we have found that the linear motion is dominant and 2nd-order motion is negligibly small for the mini-TLP, which contradicts the conventional TLP. This result could be obtained by taking a close look at the slow drift motion. We have filtered the low frequency response between the range of 0.03 and 0.05 rad/s. The band covers the peak energy density spectrum as shown in Fig. 1 (b). The amplitude frequency spectrum over the band has been brought to the time series as shown in Fig.6. The period of the largest surge amplitude has to be the surge natural period, which amounts to 165.9 sec (0.038 rad/s), which is longer than the measured data 139.97 sec as listed in Table 3. The peak amplitude of the surge is 0.5 m, which is unusually small. The small surge slow drift motion implies that the wave forces are not only small but also the size of the structure is much smaller than other conventional compliant TLPs.

, where the integral limits of are omitted, the x(t) is a general input, y(t) is a general output to the 2nd -order (e.g. force, motion). g1(t) and g2(t1,t2) are the linear and quadratic impulse response functions respectively. Fourier transform of these gives G1 and G2 making Fourier transform pairs:
g1 (t ) = 1 G 1 ( )e + it d 2

(4) (5) (6)

G1 ( ) = g1 (t )e it dt g 2 (t1 , t2 ) = 1 i t + i t G2 (1 , 2 )e( 1 1 2 2 ) d1d2 (2 ) 2

Paper No. 2005-EJP-01

246 Kim

3 /8

G2 (1 , 2 ) = g 2 (t1 , t2 )e(

i1t1 i2 t2 )

dt1dt2

(7)

formula for the estimate of the cross-bi-spectra or QTFs are given in the form
G2 (1 , 2 ) = C * (1 2 , 1 + 2 ) , 0 1 ,| 2 | + 2 U xx (1 )U xx (2 ) (10)

,where the G1() is LTF which is a function of wave frequency , while G2 (1, 2) is a quadratic transfer function QTF of the bifrequencies 1 and 2. Because the kernel g2 (t1, t2) is assumed to be symmetrical:
g 2 (t1 , t2 ) = g 2 (t2 , t1 )

(8)

G2(1, 2) and G2(1,2) are uniquely defined in the bi-frequency domain of two octants 1>0 and 1|2| ,respectively in Fig.7. The Volterra model Eq. (3) is transformed by Dalzell (1976) in the form:
y (t ) = y1 (t ) + y2 (t )

= Re a jG1 exp(i j t )
j =1

(9)

,where C* is complex conjugate of cross-bi-spectrum, and Uxx is onesided wave energy density spectrum. It should be noted that the algorithm to compute the cross-bi-spectrum was due to Dalzell (1974), which is similar to the Blackman-Tukey method for estimation of energy density spectrum. Eq. (10) is presented schematically in Fig.8. Special care is necessary in the above extraction procedure to avoid the divergence due to the negligibly small amount of the wave energy density distributed in the very low and high frequency regions as discussed in Kim and Kim (2002) Given the input and output time series one obtains QTFs according to Eq (10).

1 Re 2

G2+jk exp i ( j + k )t a j ak j =1 k =1 +G exp i ( j k )t 2 jk


n n

,where a j and ak are the wave amplitudes contained in the random sea.
G1 j and G2 jk represent the complex linear and quadratic transfer

function LTFs and QTFs. The QTFs consist of the mean, double frequency, sum frequency and difference frequency terms. + 1/ 2 G2 jk (j=k) represents the double frequency along the diagonal axis

1 = 2 in the upper octant and 1/ 2 G2jk (j=k) represents mean term


distributed along the diagonal axis 1 = 2, in the lower octant respectively. 1/ 2 G2 jk (jk) are the sum and difference frequency terms
distributed in the upper and lower octants, respectively, excluding the diagonal axes. The LTF is distributed along the 1 axis.
Fig. 8 Schematic diagram for extracting G2 from CBS and energy spectra ( 1 = 1 2 , 2 = 1 + 2 , where 1 & 2 are the difference and sum frequency)

1 = 2

Non-Gaussian method
1 G2 ( j , j ) 2

Kim and Powers (1988) proposed the non-Gaussian method. It is briefly formulated as
H =

G 2 ( j , k )

G1 ( j )
G2 ( j , k )
1 G2 ( j , j ) 2

{ E[ X

* t X ]

E[ X *Y (m)]

(11)

,with

E[ X *Y (m)] =

yx S2

[ S3yx [ S

]
(12)
xx 3

1 = 2 Fig.7 Sum and difference frequency QTFs in the upper and lower octants

E[ X * X t ] =

Gaussian method

Tick (1963) gave formula of cross-bi-spectral estimates assuming the Gaussian input. However the latter is modified by Dalzell by employing a different convention of Fourier transform. The modified

xx S2 xx S3

S4xx

Paper No. 2005-EJP-01

247 Kim

4 /8

is the estimated solution of the general Volterra model in ,where, H

frequency domain,
xx bispectrum, S 4

xx S2

spectrum,

xx S3

conjugate of bispectrum, S
xx 3

yx cross-spectrum, S 3 cross-bispectrum. The inversion of input spectral momenet matrix, E[ X * X t ] and the multiplication with E[ X *Y (m)] is significantly complicated and affected by the number of ensemble average over the data segment used xx estimate has higher because the variance of the trispectrum, S 4

yx trispectrum, S 2

frequency range and the experimental LTFs are bigger in the high frequency region. Fig. 11 shows QTFs estimated by two methods. As it may be noticed, irrespective of the two methods, there is a single distinctive bump along 1-axis, and there is no bump along 2 -axis in bi-frequency plan. This is aligned with the fact that the 2nd-order components of a compliant Mini TLP in a surge motion are mainly created by the low frequency 2nd -order motion.
1.5 Gaussian Non-Gaussian Theory

xx variance than the power spectrum, S 2 . And the small variance of trispectrum gives better and stable solution. Ensemble average over data segment provides reasonable inversion results for {E[ X * X t ]} with numerical stability. This implies the estimated solution of the Vollterra model has close realtionhsip with the variance of input spectral moments. Thus, we need to invert this matrix with small variance to compute a better solution. This non-Gaussian method is modified with Principal Component Analysis (PCA) algorithm, which yields considerably lower variance in the estimation of Volterra model (Birkelund et al., 2003).

LTF (m/m)

0.5

0 0

0.2

0.4

0.6

Fig. 10 Estimated LTFs

(rad/s)

0.8

1.2

1.4

In this study for the Gaussian method, each 1-hour data was computed, and an ensemble average was made over 3 hours of data to obtain the LTF and QTF. In the foregoing computation, the maximum lag 70 is used. In the non-Gaussian method, both input and output data were decimated by a factor R = 4. After the down sampling, the time step becomes 0.632 seconds. The maximum frequency, max = 9.9 rad/s, is large enough to cover the 2nd-order frequency range associated with input frequency components. The segment length was chosen to be 256 and the number of segments was 66. The non-Gaussian method was modified with the principal component analysis (PCA) algorithm The QTF by the non-Gaussian method is smaller in volume than by the Gaussian method. Given a long enough data, a QTF would increase in volume.
Mini TLP (Gaussian)

fN: Nyquist frequency Fig. 9 HQ in bi-frequency domain It should be noted that this method dose not include frequencies along fi= fj in difference interaction region (Fig. ) when H2 is computed. H1 corresponds to G1 and H2 to G2 of Gaussian input method, respectively. The non-Gaussian method necessarily involves division in the multiplication of the inverse input spectral momenet matrix, E[ X * X t ] and E[ X *Y (m)] like the Gaussian method. Consequently, the large diverging magnitude of QTF occurs in some parts of H2 plane due to the large numerator compared to the small denominator (tails of input energy spectra) in the division process. In this method, the unrealistic QTF should be accurately eliminated as done in the Gaussian method. This process could lose some true QTF. Another disadvantage in the use of non-Gaussian method is that the algorithm cannot determine the mean QTFs. Thus in our analysis we have approximated the mean value by taking the QTFs close to the terms along the diagonal axis.
Estimation of LTFs and QTFs from experimental data
G2(1,2) (m/m2)

1.5 1 0.5 0 0 0.5


2(rad/s)

0 0.5 1 1.5 1.5 1


1(rad/s)

Mini TLP (Non-Gaussian)

G2(1,2) (m/m2)

1.5 1 0.5 0 0 0.5


2(rad/s)

0 0.5 1 1.5 1.5 1


1(rad/s)

Estimated LTFs by both methods and potential theory are compared in Fig. 10. Theoretically estimated LTF by WAMIT code appears to be smooth over entire frequency range, while experimentally estimated LTFs by the Gaussian and non-Gaussian method give less smooth curves, which is attributed to the test data. A smoother LTF curve would be obtained with much longer data. All estimated LTFs appear to be in the same trend, but the theoretical LTF is bigger in the low

Fig.11 Estimated QTFs

Paper No. 2005-EJP-01

248 Kim

5 /8

ANALYSIS OF RECONSTRUCTED DATA


One may reconstruct the output using the formula Eq. (9), in the time domain, so that one may compare the reconstructed with the original measured data.
Reconstruction of response data

1.2 1 0.8
Surge (m)

Non-Gaussian Gaussian

0.6 0.4 0.2 0 5000 5100 5200 t (s) 5300 5400 5500

The reconstructed data may provide a way for comparison of the two methods. It is important to observe the behavior of the response due to the two different methods in the time domain. The comparison of the reconstructed with the original measured response will provide valuable information that is not be available in the frequency domain. The reconstruction procedure is explained referring to Fig. 12; at the first step we estimate LTFs and QTFs; at second step multiply these with the amplitude spectrum of the measured nonlinear random wave according to Eqs.(9) and (11) to reconstruct the nonlinear surge.

Fig. 15 Comparisons of time series of quadratic component Fig. 13 illustrates a comparison between the measured and reconstructed response time series by the two methods. The difference of both methods is not clearly seen in the time series. The reconstructed data are used to produce the surge energy density spectra as shown in Fig. 14. It is uneasy to distinguish the results from each other. In Fig. 15 we see that the 2nd-order surge motion by the Gaussian method is larger than the non-Gaussian method. This is attributed to the fact that the distribution of QTFs by Gaussian method is larger than the nonGaussian method as shown in Fig.11. Fig. 16 shows the reconstructed low-frequency surge motion as it was shown in Fig. 6. Zero-upcrossing analysis was applied to the 2nd-order low-frequency surge motion to compare each low-frequency surge period between the reconstructed time series. It was found that the periods from the reconstructed data are larger than the measured periods as shown in Table 4.
Gaussian

Fig. 12 Schematic diagram of Volterra quadratic model with measured input and reconstructed output, using estimated LTF and QTF
5 4 3 2
Surge(m)

0.8

Experimet Non-Gaussian Gaussian

Surge(m)

0.6

0.4

1 0 -1 -2 -3 -4 100 200 300 Time(s) 400 500 600


Surge(m)

0.2 6500 7000 Time(s) Non-Gaussian 7500 8000

0.4

Fig.13 The reconstructed time series compared with the original test data.
Mini TLP (Gaussian) 7 6 5
U( ) (m s)

0.3

0.2

Mini TLP (Non-Gaussian) 7 6 5


U( ) (m s)
2

0.1 6500 7000 Time(s) 7500 8000

4 3 2 1 0 0 0.2 0.4 0.6 (rad/s) 0.8 1

4 3 2 1 0 0 0.2 0.4 0.6 (rad/s) 0.8 1

Fig.16 Reconstructed low frequency surge motion due to Gaussian and non-Gaussian method Table.4 Reconstructed low frequency surge motion (0.03rad/s< < 0.05rad/s) Methods Average of surge period (s) Gaussian method 162.8 Non-Gaussian method 163.3 Present analysis of measured data 165.9 Measured value (Liagre 2000) 140.0 The natural periods estimated from the reconstructions are less than the data directly determined from the measured data. This is due to the loss of QTFs in the process of cross-bi-spectral estimates.

Fig.14 Spectrum of surge motion: measured (solid), reconstructed data (dotted)

Paper No. 2005-EJP-01

249 Kim

6 /8

Mini TLP (Gaussian)

Normalized mean square error

1.2 1

Linear Quadratic Total

To compare the reconstructed and measured time series quantitatively, the normalized mean square error (NMSE) was computed. The experimental data was filtered by eliminating the frequency components higher than 4 rad/s, which may be regarded as noise. The NMSE dose not include mean of surge.
1 N

Coherency

0.8 0.6 0.4 0.2

NMSE =

{[ x
N t =1

rec (t ) xrec (t )] [ xexp (t ) xexp (t )]}

1 N

[ x
t =1

(12)

0 0

0.2

0.4

exp

(t ) xexp (t )]

1 0.8 0.6 (rad/sec)

1.2

1.4

Mini TLP (Non-Gaussian) 1.2 1 0.8 0.6 0.4 0.2 0 0 Linear Quadratic Total

Table 5 Normalized mean square error Gaussian 0.43 Mini TLP Non-Gaussian 0.37

3 hours

Coherency

, where, N is the number of data, xrec(t) reconstructed time series and xexp(t) experimental time series. The NMSE are compared in Table 5. It appears that non-Gaussian method is better than Gaussian method.

0.2

0.4

Comparison of cohebrcy fucntions

Fig. 17 Coherency of reconstructed surge motion


Comparison of surge drift of conventional TLP and mini-TLP

1 0.8 0.6 (rad/sec)

1.2

1.4

The reconstructed data is used for the coherency test. The linear and quadratic coherence functions are:
2 ( ) = xy 1

Uy

U yy ( )
Uy
y 2 2

y ( ) 1 1

(13) (14)

2 q xy ( ) = 2

( )

U yy ( )

We compare the characteristic behaviors of surge drifts of the conventional TLPs (Fig 18a) and that of the mini-TLP (Fig. 18b). The surge drift of the conventional TLPs is dominated by the 2nd-order response compared to the small linear surge overridden on the 2nd-order drift. The mini TLP is dominated by the linear surge with negligibly small 2nd-order drift. The above comparison may used for visualization of the coherency test (Fig.16).

,where 2 and q2 represent the linear and quadratic coherence functions respectively. U indicates one-sided energy density spectra. The suffixes y1 and y2 stand for the reconstructed linear and quadratic forces, respectively, and y denotes the measured force. The sum of these coherence functions must lie in between 0 and 1. A goodness-of-fit measure for the validity of the quadratic nonlinear model can be defined by seeing how close is the sum of these linear and quadratic coherence function to unity (Bendat, 1990). The reconstructed data and measured data are used for the coherency test. The linear and quadratic coherency function (Eqs. (13) and (14)) were calculated. The results are shown in Fig. 17. The linear coherency is close to 1 over wide frequency region, 0.20.9 rad/s for all methods, which means that the linear surge motion is dominant and the quadratic surge motion is negligibly small in this region; the quadratic coherency is about 0.4 for the Gaussian method, 0.2 for the non-Gaussian method, while the non-Gaussian method is much higher than the Gaussian method beyond 1 rad/s. It is to be noted that the coherency at zero frequency is unit for the Gaussian method whilst 0.2 in the non-Gaussian method.

Fig. 18 (a) Conventional TLP surge motion

Fig. 18 (b) Mimi TLP surge motion

CONCLUDING REMARKS
We have conducted the analysis of slow 2nd-order surge of a mini-TLP in two ways; one is to utilize the raw data and low pass filtering of the response and the other is to conduct the cross-spectral estimates and reconstruction of the surge response. We have used Gaussian and nonGaussian method for the estimation of QTF. Overall the two methods give almost the same result. From the filtered 2nd-order surge responses we have found the mean surge (offset) and natural slow drift surge period. The period is longer than the measured data. It was found that

Paper No. 2005-EJP-01

250 Kim

7 /8

the input and output were practically linear and thus, the system is also linear. The behavior of the 2nd-order surge of the mini-TLP is totally different from the conventional TLPs. The linear surge was dominant compared to the 2nd-order surge for the mini-TLP, whereas the 2ndorder surge is dominant compared to the linear surge for the conventional TLPs. Assuming that the system is quadratically nonlinear and analyzing the system, we found the system is practically linear. The present quadratic system analysis has been used as a useful method in identifying that the system is practically linear.

ACKNOWLEGEMENTS
The authors are grateful for access of the test data from the Offshore Technology Research Center (OTRC).The authors also would like to thank Drs. John Niedzwecki and Pierre Liagre for their assistance in providing the data. The authors are thankful to Professor Powers and Dr. Birkelund for their precious comments on the work.

REFERENCES
Bendat, JS (1990). "Nonlinear System Analysis and Identification from Random Data," John Wiley & Son, New York, N.Y. Birkelund, Y, and Powers, EJ (2003). "Higher-Order Spectral Estimators and Nonlinear System Identification," Proc 11th Int\Offshore and Polar Eng Conf, ISOPE, Stavanger, Norway, Vol III, pp 78-84. Dalzell, JF (1972). "Application of Cross-Bi-Spectral Analysis to Ship Resistance in Waves, "SIT-DL-72-1606, Stevens Institute of Technology, Hoboken, NJ, May. Dalzell, JF (1974). "Cross-Bispectral Analysis: Application to Ship Resistance in Waves, " J Ship Res, Vol 18, No 1, pp 62-72. Dalzell, JF (1976). "Application of the Functional Polynomial Model to the Ship Added Resistance Problem, 11th Symposium on Naval Hydrodynamics, London, March 1976. Dalzell, JF and Kim, CH. (1979). "An Analysis of Quadratic Frequency Response for Added Resistance, J Ship Res, Vol 23, No 23, pp 198 -208. Hasselmann, K. (1966). "On Non-Linear ship Motions in Irregular Waves," J of Ship Research, Vol 10, No. 1. pp 64-68. Kim, CH and Boo, SY (1990) "Statistical analysis of slow drift forces in random seas," Proc. 1st Pacific /Asia Offshore Mechanics Symposium, Seoul, Korea, pp 169-177. Kim, KI, and Powers, EJ (1988). "A Digital Method of Modeling Quadratically Nonlinear Systems with a General Random Input, IEEE Trans Acoustics, Speech, and Signal Processing, Vol 36, pp 1758-1769. Kim, CH and Kee ST (1992). "Non Gaussian Effect of Input and Output on Cross-Bi-Spectral Estimates," Proc 2nd Int Offshore and Polar Eng Conf, ISOPE, San Francisco, USA, Vol III, pp 496-503. Kim, NS (2004). Extraction of the Second-Order Nonlinear Response from Model Test Data in Random Seas and Comparison of the Gaussian and non-Gaussian Models. Dissertation, Texas A&M University. Kim, NS and Kim, CH (2002). "Cross-Bi-Spectral Estimation of Nonlinear Force on Fixed Structure in Nonlinear Waves," Proc 12th Int Offshore and Polar Eng Conf, ISOPE, Kitakyshu, Japan, Vol III, pp 188-195. Kim, NS and Kim, CH (2003). "The Effect of Sea Severity on the CrossBi-Spectral Estimate of Quadratic Response Function for Surge Exciting Forces," Proc 13th Int Offshore and Polar Eng Conf, ISOPE, Honolulu, Hawaii, Vol III, pp 413-420. Kim, NS and Kim, CH (2004). "The Effect of Sea Severity on the CrossBi-Spectral Estimate of Quadratic Response Function for Surge Exciting Forces," Proc 13th Int Offshore and Polar Eng Conf, ISOPE, Toulon,

France, Vol III, pp 413-420. Kim, SB and Powers, EJ (1995). "Estimation of Volterra Kernels via higher-order statistical signal processing," in: Boashash, B, Powers, EJ, and Zoubir, AM, eds, Higher-Order Statistical Signal Processing, Wiley, New York, NY, pp 213-239. Kumar, A, Kim, CH, and J, Zou (2002). "Limitation of the 2nd-order Theories for Laboratory High Sea Waves and Forces on Structures," Int J of Offshore and Polar Engineering, Vol 12, No 4, pp 243-248. Niedzwecki, JM, Liagre, PF, Roesset JM, and Kim MH (2001). " An Experimental Research Study of a Mini-TLP," Proc 11th Int Offshore and Polar Eng Conf, ISOPE, Stavanger, Norway, Vol IV, pp 631-633. Stansberg, CT (1992) "On the Estimation of Extreme Mooring Line Forces," Proc the OMAE 1992 Conf, Calgary, Canada, I-A, pp 291300. Stansberg, CT (2001). "Data Interpretation and System Identification in Hydrodynamic Model Testing," Proc 11th Int Offshore and Polar Eng Conf, ISOPE, Stavanger, Norway, Vol III, pp 1-9. Tick, LJ (1961). "The Estimation of the Transfer Functions of Quadratic Systems," Technometrics, Vol 3, No 4. pp 563-567. Volterra, V (1943). "Theory of Functionals and of Integral and IntegroDifferential Equations," Dover Publications Inc., New York, 1959.

Paper No. 2005-EJP-01

251 Kim

8 /8