Sunteți pe pagina 1din 8

# Safety analyses for bulk carriers using metamodels of still water loads

P. Georgiev
Technical University of Varna, Bulgaria

## Figure 1. Connection between metamodeling tecnique and Monte-Carlo simulation

xi =

dPi , i = 1,2,3 10

(1)

The amount of cargo in each hold is obtained by: Pi = (1 xi / 10) P0i , i = 1,2,3 (2)

where P0 is the planned amount of cargo. The output variables are SWBM at different sections related to the permissible ones and the trim.

Fitting of metamodels is based on RSM (Response Surface Methodology). Polynomial regression is chosen due to its transparency and simplicity and expected low order of non-linearity of considered responses. The response surface y (x) is presented by:
y (x) = b0 + x T b + x T Bx (3)

3 MONTE-CARLO SIMULATION The connection between the Monte-Carlo simulation and fitted metamodels is shown in Figure 1. The Monte-Carlo simulation calculates the fitted metamodels for trim and SWBM each time using different randomly-selected values for the amount of cargo in holds. In reality, the process of cargo loading is monitored by the crew and terminal representatives. To take into account the human factor, the following constraints for the total amount of cargo and obtained trim are accepted:

In this model it is assumed linear effects in x and, two-factor interaction and pure quadratic term in x (the term xTBx) One of the important steps in computer experiment is the selecting a design with and appropriate number of runs and levels for each variable to ensure sufficient design space coverage. The present study uses uniform design proposed by Fang (2006). The designs are marked by Un(qs) where n = number of experiments, q = number of levels and s = maximum number of factors. Designs for different number of factors, levels and experiments could be taken from (www.math.hkbu.edu.hk/). For present study U15(33) design was used. Installed on-board loading instrument ALCOS was used to calculate the responses. It is necessary to evaluate and to record the results for only 15 loading conditions that take less than 15 minutes time. The least squares estimates for trim and SWBM for eight frames are made by JMP software. Table 1 presents the polynomial coefficients for trim and SWBM at the muddle of holds No 2, 3 and 4. For deterministic computer experiment where a random error does not exist, the metamodel accuracy is checked by the modelling error that is the discrepancy between the true output Y from the simulation model and Y from the metamodel. The maximum modelling error for SWBM is less than 0.2% from the permissible values and for trim the maximum error is 17 mm (Georgiev 2010).
Table 1. Polynomial sponses. x1 x1 -0.018583 x2 x3 Trim, m x1 -0.001536 x2 x3 Fr.93, x1 -0.001424 x2 x3 Fr.129, x1 -0.001129 x2 x3 Fr.165, regression coefficients for studied rex2 -0.021596 -0.001311 x3 -0.003580 0.016679 0.016678 b0= 0.003450 -0.001600 -0.003963 b0= 0.002706 -0.001191 -0.002969 b0= 0.001735 -0.000675 -0.001683 b0= b 1.368173 0.251147 -0.979926 0.859706 0.096288 -0.091438 -0.008829 0.758067 0.125465 -0.164884 0.023898 0.601142 0.109279 -0.087475 0.022291 0.963968

## Qmin Pi Qmax ; t min trim t max ,

i =1

(4)

0.000809 0.000946

0.000241 0.000567

where Qmin is the minimum amount of cargo taken as 97% from the planned one, Qmax is the maximum amount limited by summer draught, tmin is the maximum aft trim accepted as 0.005L=0.88 m and tmax is the maximum trim to bow equal to 0.04 m. For the input variables a double truncated normal distribution with lower and upper truncation point ZL=ZU= 2, mean zero and truncated standard deviation T = 5% is accepted. Using the tables (Khasawneh et al, 2005) this corresponds to normal distribution with mean zero and = 5.684%. For samples generation RiskAMP MS Excel Add-in was used (http://www. riskamp.com/). To investigate how accurate are the particular estimates five samples with 10,000, 20,000, 30,000, 40,000 and 50,000 runs are generated. Additionally, the investigated range of trim is subdivided into smaller ranges as follows: T1: [-0.19 0.04); T2: [-0.42 -0.19); T3: [-0.65 -0.42); T4: [-0.88 -0.65). To prepare the calculations by the metamodels and to sift out the feasible points a small computer program, written in Pascal was used. The number of feasible points at different starting number of runs is shown in Table 2. The total number of feasible points is about 25% from starting number of runs. The distribution of cargo is shown in Figure 2. The graph includes the histogram for T1 trim range and fitted normal probability distributions for others. One can see the correct relation between the cargo mass in the three holds in bubble plot. The mass of cargo in middle hold 3 is greatest (bigger circles) when the mass in hold 1 and 5 are smallest (bottom left corner) and vice versa (top right corner).
Table 2. Number of feasible points for trim ranges No of Trim range runs T1 T2 T3 T4 10x103 696 676 567 501 20x103 1430 1347 1158 1020 2197 2034 1797 1465 30x103 40x103 2869 2674 2416 1950 50x103 3565 3415 3012 2397

-0.000310 0.000131

1000

800

800

600

Frequency

Hold No 1
600

Hold No 3

400

400

200

200

0 11200

11600

12000

12400

12800

13200

## 0 12400 12800 13200 13600 14000 14400 14800 15200

Frequency

Cargo mass, t

Cargo mass, t

a)
1200 1000 800 600 400 200 0 12000 12400 12800 13200 13600 14000 14400

b)
Hold No 5

## -T1 -T2 -T3 -T4

Frequency

Cargo mass, t

c)

d)

Figure 2. Cargo distribution for different trim ranges: -a); -b); -c). Bubble plot for cargo distribution in the holds d).

z1 / 2 N

(5)

0.8

## 1.0 95%CI-Fitted function 99%CI-Fitted function -T1 -T2 -T3 -T4

where is sample mean, is standard deviation, N is number of runs, and z1-/2 is the number that (z)= ( denotes the standard normal c.d.f). It is common practice in simulation to use and report the absolute (wa) and relative (wr) widths of the confidence interval (CI), defined as

0.6

0.4

0.2 0

wa = 2 z1 / 2

1000

and wr =

wa

(6)

## 2000 3000 Number of runs, N

4000

In Figure 3 the relative width in percent (wr) for = 5% and =1% significance levels is shown. The results include all trim ranges and as one can see from the trend lines, the accuracy of estimation of the mean is proportional to N-0.5. The maximum relative width of CI is less than 1% and goes down to less than 0.4% at 50,000 starting runs. The descriptive statistics for mass of cargo distribution in the holds at T1 range for sample size N=3565 are shown in Table 3. The table presents the planned and maximum permissible (MaxLoad) amount of cargo.

Figure 3. Relative width for 95% and 99% confidence for cargo mass in Hold No 3. Table 3. Descriptive statistics for mass of cargo in holds at T1 trim range Item Hold 1 Hold 3 Hold 5 Plan P0, t 12536 13884 13186 MaxLoad, t 12536 13884 13.835 ,t 12446.0 13818.4 13141.5 95% CI, t 9.6 18.3 + 11.2 99% CI, t 12.6 24.0 14.7 ,t 291.9 556.8 341.7 CoV, % 2.3 4.0 2.6 Skewness -0.0001 0.0314 -0.0299 Ex. Kurtosis -0.1283 -0.4916 -0.2440

acters (1 or 0) that represent the overloading in holds No 1, 3 and 5 respectively. The sample with total 12389 points is used and the event tree is shown in Table 4. The last column includes the values of product rules. Having the results from the event tree we can draw different conclusions. The probability of overloading of every separate hold and combination of holds for whole trim range is shown in Figure 5. The probability that there is no overloading is about 31% and probability for overloading of hold No 3 (the middle hold) is about 39%.

## Figure 4. Event tree for two events A and B

P( B T ) =

P( B T ) , P(T) 0. P(T )

(7)

Let the trim ranges Ti i=1,24 are events that partition the whole trim range T. Following the theorem of the total probability we obtain
P ( B) = P(B Ti ).P(Ti ) , i = 1,2...4
4 i =1

(8)

For conditional probability of B, given that Ti has occurred the following is valid: P( B Ti ) = P( B Ti ) , P(Ti ) 0 ; i = 1,2,...,4 P(Ti ) (9)

In case where it is necessary to reverse the order in conditional probabilities the Bayes theorem is used:
P (Ti B ) = P(Ti ) P(B Ti ) , i = 1,2...4 P( B)

(10)

The masses of cargo in the holds are considered as not independent and the event of overloading is labelled B1, B3 and B5. The outcome for every branch is marked with a string that includes the number of trim range and combination of three char-

Table 4. Event tree for overloading of double bottom Ti B1 B3 B5 Outcome 228/ 0/228 Y 1111 0.0 Y 1369 1369/ 228/228 N 1110 0.01840 Y 3565 67/1141 Y 1101 0.00541 1141/ N 1369 3565/ 1074/1141 N 1011 0.08669 12389 1403/ 0/1403 Y 1011 0.0 Y 2196 2196/ 1403/1403 N 1010 0.11325 N 3565 0/793 Y 1001 0.0 793/ N 2196 793/793 N 1000 0.06401 37/ 0/37 Y 2111 0.0 Y 811 811/ 37/37 N 2110 0.00299 Y 3415 154/774 Y 2101 0.01243 774/ N 811 620/774 N 2011 0.05004 3415/ 12389 1488/ 0/1488 Y 2011 0.0 Y 2604 2604/ 1488/1488 N 2010 0.12010 N 3415 1116/ 0/1116 Y 2001 0.0 N 2604 1116/1116 N 2000 0.09008 1/ 0/1 Y 3111 0.0 Y 406 406/ 1/1 N 3110 0.00008 Y 3012 273/405 Y 3101 0.02204 405/ N 406 132/405 N 3011 0.01065 3012/ 12389 1306/ 13/1306 Y 3011 0.00105 Y 2606 2606/ 1293/1306 N 3010 0.10437 N 3012 1300/ 36/1300 Y 3001 0.00291 N 2606 1264/1300 N 3000 0.10203 4111 0.0 0/ Y 156 156/ 4110 0.0 Y 2397 155/156 Y 4101 0.07199 156/ N 156 2397/ 1/156 N 4011 0.00046 12389 1036/ 56/1036 Y 4011 0.00302 Y 2241 980/1036 N 4010 0.05292 2241/ N 2397 234/1205 Y 4001 0.01264 1205/ N 2241 971/1205 N 4000 0.05244

100

No

H1 only

H3 only 1.2

H5 only

80

Probability, %

39.4

43.6 42.9

6.5

60

40

30.1

18.2

4.4

27.4 0.2

20 22.2 0 T1

32.7

42.0 27.1 T3 T4

T2

Figure 5. Probability (%) of overloading of every cargo hold and combination of holds Table 5. Descriptive statistics for SWBM at trim range T1 Statistic

Figure 6. Probability for overloading of separate holds at given trim range T1-T4

## Frames Fr.129 Fr.139 Fr.147 Fr.165 Fr.183

,95% CI .102, 99% CI .102, ,CoV, % Skewness Ex. Kurtosis Plan, X/L

Fr. 75

Fr. 86

Fr. 93

Fr.111

## 0.6449 0.110 0.145 0.0335 5.2% -0.0538 -0.3238 0.6515 0.7913

2.0

1.5

1.0

4.2 Still water bending moments During the Monte-Carlo simulation for feasible points the SWBM is calculated by fitted metamodels. Descriptive statistics and confidence interval for the mean at all controlled sections for trim range T1 are shown in Table 5. The last two rows include values for SWBM for the planned loading condition

0.5

## Relative width, wr, %

and location of the section. The relative width of 95% confidence interval for three sections and corresponding trend line are shown in Figure 7. The least accuracy of estimation of the mean is obtained for frame 129. The frame is located at the middle of the ship and the standard deviation of SWBM for this frame is greatest.
3.0 wr = 42.475N-0.5013 ; R2=0.9956 2.5 wr = 64.005N-0.5038 ; R2=0.9946 wr = 24.328N-0.4971 ; R2=0.9966

Fr. 111
0.0 0 1000

Fr. 129

Fr. 165
4000

## 2000 3000 Number of runs, N

Figure 7. Relative width (wr) for 95% confidence for SWBM for three frames

## 16 14 12 10 8 6 4 2 0.3 0.4 0.5 0.6 0.7

780/3565 = 0.22; T2 - 741/3415 = 0.22; T3 576/3012=0.19 and T4 - 362/2397 = 0.15. The curve for mean values is very close to the curve of the plan i.e. the values obtained by loading instrument. For all frames the samples for SFBM have negative excess kurtosis that shows distribution with lower, wider peak around the mean and thinner tails. A negative skew for all frames indicates longer left tail in the distribution, but the values are very small and one can conclude that the data are not skewed. 5 CONCLUSIONS
x/L
0.8

1.0

0.8

SWBM, -

0.6

0.4

## Figure 9. Region around the mean value of SFBM

The standard deviation and CoV are biggest at the middle sections. From Figure 8 one can see that the maximum standard deviation is about 9% from the permissible SFBM and CoV is about 16%. The most important remark based on the descriptive statistics of SWBM is illustrated in Figure 9. The plot includes the SWBM related to the permissible values for initial loading condition (the plan), and obtained for controlled sections mean and mean 2 . The shaded area corresponds to 95.45% of all possible values for SWBM that could exist for this trim range. Due to the deviation from cargo plan and uncertainties in cargo distribution we will have a range of possible SWBM values. For frame 165 (X/L=0.71) there are values greater than the permissible ones (rounded rectangle) and overloading may occur. The relative frequency for such overloading for considered trim ranges is as follows: T1 -

%
- CoV -

max permissible

2 +2