Download as ppt, pdf, or txt
Download as ppt, pdf, or txt
You are on page 1of 43

INCHOE2014

PERFORMANCE OF A FLOATING PIPE BREAKWATER BASED


ON INTELLIGENT COMPUTING

BY
Dr. Sukomal Mandal
Senior Scientist, Ocean Engineering Division
National Institute of Oceanography, Goa

1
OUTLINE

 INTRODUCTION

 Wave transmission of floating breakwater

 Problem statement

 Research objectives

 HORIZONTAL INTERLACED MULTILAYER MOORED


FLOATING PIPE BREAKWATER (HIMMFPB)
 About experimental horizontally interlaced multilayer moored
floating pipe breakwater (HIMMFPB) model
 Data used to train and test computational intelligence models

2
OUTLINE (Contd.)

 SUMMARY OF THE WORK

 Performance of ANN-1 for wave transmission prediction of


HIMMFPB
 Performance of Genetic algorithm based support vector machine
regression (GA-SVMR-1) model for wave transmission prediction of
HIMMFPB

3
OUTLINE (Contd.)

 CONCLUSIONS

 REFERENCES

4
INTRODUCTION

Typical floating type of breakwater (Fousert, 2006)

5
INTRODUCTION (contd.)
 GAPS

 Vast volume of published work related to floating breakwaters


[(Homma et al, 1964 Harris and Webber, 1968; Adee and martin,
1974; Harms, 1979; Bishop, 1982; Leach et al., 1985; Murali and
Mani, 1997; Sannasiraj et al., 1998; Sundar et al., 2003; Li et al.,
2005; Ruol et al., 2008)]

 Researchers frequently used engineering judgments based on


experience, subjectivity, confidence in models to deal with the
complexity associated with wave structure interaction of floating
breakwater,

 but failed to give simple mathematical model considering all the


coastal boundary conditions
6
INTRODUCTION (contd.)

Pictorial view of Horizontally interlaced multilayer moored floating pipe


breakwater in the flume ( Deepak, 2006; Hegde et al., 2007; Jagadisha, 2007)

7
INTRODUCTION (contd.)

 In floating pipe breakwaters the energy dissipation process depends on


pipe interference effect, the spacing between the pipes (S) and number
of layers (n).

 The effect of all these factors on transmission and forces in the


moorings is not clearly understood. it will be extremely difficult to
quantify them mathematically.

 HIMMFPB is a complex problem.

8
INTRODUCTION (contd.)

 MOTIVATION

 From the literature review it is found that, Intelligent Computing (IC)


techniques are successfully used to solve complex problems associated
with coastal/ocean engineering.

 Patil et al. (2011, 2012) have successfully developed ANN and GA-
SVMR models to predict wave transmission of HIMMFPB. They used
input parameters, such as
 relative spacing of pipes (S/D)
 relative breakwater width (W/L)
 ratio of incident wave height to water depth (Hi/d), and
 incident wave steepness (Hi/L)

9
INTRODUCTION (contd.)

 MOTIVATION

 Performance of CI depends on better selection of input parameters and


one must understand the physics of the wave structure interaction

 Many researchers [Neelamani and Rajendran (2002), Sundar et al.


(2003), Wang et al. (2006), Neelamani and Gayathri (2006), Hegde et al.
(2007), Teh et al. (2010)] based on physical model study and knowing
the physics of the wave structure interaction have revealed that
coefficient of transmission (Kt) decreases with increase in relative depth
of water (d/L).

 They used relative depth of water (d/L) as one of the important


parameter in physical model study of floating break water.

10
INTRODUCTION (contd.)

 MOTIVATION

 However, it is noticed that Patil et al. (2011, 2012) have not used relative
depth of water (d/L) while developing IC models in predicting wave
transmission of HIMMFPB.

 This has motivated us to use relative depth of water (d/L) as one of the
input parameter in developing IC models and study its performance in
terms of wave transmission coefficient (Kt).

11
HORIZONTALLY INTERLACED MULTILAYER
MOORED FLOATING PIPE BREAKWATER
(CONTD.,)
 Range of wave specific and structure specific parameters
used in HIMMFPB study
Wave-specific parameters Experimental range
Incident wave height, Hi (mm) 30, 60, 90, 120, 150 and180
Wave period, T (sec) 1.2, 1.4, 1.6, 1.8, 2.0 and 2.2
Depth of water, d (mm) 400, 450 and 500
Structure-specific parameters Experimental range
Diameter of pipes, D (mm) 32
Ratio of spacing to diameter of pipes, 2, 3, 4 and 5
S/D
Relative breakwater width, W/L 0.4 to 2.65
Number of layers, n 5

12
HORIZONTALLY INTERLACED MULTILAYER
MOORED FLOATING PIPE BREAKWATER
(CONTD.,)
 Data used for training and testing the computational
intelligence models
S/D ratio Data for training Data for testing Total data

2 609 203 812


3 576 233 809
4 366 143 509
5 580 234 814
Combined 2131 813 2944
Total

13
HORIZONTALLY INTERLACED MULTILAYER
MOORED FLOATING PIPE BREAKWATER
(CONTD.,)
 Input parameters and data used to train and test ANN-1
and GA-SVMR-1 models

Input Parameters Data for Data for


training testing
S/D , W/L, Hi/d and d/L 2131 813

14
SUMMARY OF WORK
 Performance of ANN-1 model for wave
transmission (kt) prediction of HIMMFPB

 Introduction

 ANN is an information processing paradigm that is inspired by the way


of biological nervous system, such as brain process information.

 In the present work three layer feed forward error back-propagation


neural network was used.

 First layer represents the input nodes, hidden nodes as second layer
and output node as third layer.

 The main objective of neural network is to train the model such that the
result output are nearer to the desired values.

15
SUMMARY OF WORK (Contd.)
 Feed forward back-propagation network
x1 SINGLE NEURON NODE

y1 B WEIGHTS

x2
BIAS

z1
x3
y2
zo

x4 y3

ym
xn

INPUT LAYER HIDDEN LAYER OUTPUT LAYER

i = 1, 2…n j = 1, 2…m k = 1, …o

Typical Feed Forward Back-Propagation Network

16
SUMMARY OF WORK (Contd.)
 Performance of ANN-1 model for wave
transmission (Kt) prediction of HIMMFPB (Contd.)

 Kt calculation by ANN-1

 The input parameters that influence the wave transmission


(Kt) of HIMMFPB such as S/D, W/L, Hi/d, and d/L are
considered

 Based on the input parameters, ANN-1 models are


constructed to predict the transmission coefficient of
HIMMFPB

17
SUMMARY OF WORK (Contd.)
 Feed forward back-propagation network
bias at hidden layer
bias at output layer

S/D

W/L
Kt

Hi /d

d /L

Input Layer Hidden Layer Output Layer

18
SUMMARY OF WORK (Contd.)
 Performance of ANN-1 model for wave
transmission (Kt) prediction of HIMMFPB (Contd.)

Correlation coefficient of Kt for two network models

Test Data
Network
Model
(Input nodes - Hidden nodes - Epochs)
CCtest RMSE SI

ANN-1 4-5-200 0.9504 0.05007 0.08932

ANN
(Patil et 4-5-200 0.9488 0.05395 0.09625
al 2012)

19
SUMMARY OF WORK (Contd.)
 Test correlation between predicted and measured Kt for ANN-1
model with S/D=5.
1
CC Test = 0.9504
0.9

0.8

0.7

0.6
Predicted Kt

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Measured Kt

20
SUMMARY OF WORK (Contd.)
 Performance of GA-SVMR-1 model for wave
transmission (Kt) prediction of HIMMFPB
 Introduction to SVMR
 Vapnik (1998) proposed the support vector machines (SVMRs),
which is based on statistical learning theory.

 The basic idea of support vector machine regression is to map


the original data into a feature space with high dimensionality
through a non-linear mapping function (kernel) and construct
an optimal hyper-plane in new space.

 The role of the kernel function simplifies the learning process by


changing the representation of the data in the input space to a
linear representation in a higher-dimensional space called a
feature space.
21
SUMMARY OF WORK (Contd.)
 Introduction to SVMR (Contd.)

The loss margin setting corresponds to one dimensional linear SV machine


22
SUMMARY OF WORK (Contd.)
 Proposed GA-SVMR-1 Model
Training Testing
Data Data

Training Training Testing Testing


Input Target Input Target

Testing
GAs SVMRs
Training using Compare to get
Testing best model
Input
Parameter
Selection
Kernel BEST
Functions MODEL

Figure 3. Proposed GA-SVMR-1


model

23
SUMMARY OF WORK (Contd.)
 Kernel Functions

Kernels Functions

1 1
spline K ( xi , x j )  1  ( xi , x j )  ( xi , x j ) min( xi , x j )  min( xi , x j ) 3
2 6

b-spline K ( xi , x j )  B2 d 1 ( xi  x j )

24
SUMMARY OF WORK (Contd.)
 Genetic Algorithm for selecting parameters in the GA-SVMR-1 Model

 Genetics algorithms (GAs) are the search methods based on the


principles of natural selection and genetics (Holland, 1975)

 GAs are used for better combination of SVM parameters and


kernel parameters for better generalization of SVMR model

 The steps involved in GA for selecting SVMs and kernel parameters


are as follows:

25
SUMMARY OF WORK (Contd.)
 Genetic Algorithm for selecting parameters in the SVMR Model (Contd.,)

 Step 1.( Initialization ): Population size was set to 50

 Step 2. (Evaluating Fitness): fitness of each chromosome is


evaluated by
1 N

 Fitness Function =
 2
N
 (d
i 1
i  yi ) 2

1 N
where  2

N
 (d
i 1
i  di )2

 Step 3. (New Population) : New population is created using the


following steps

26
SUMMARY OF WORK (Contd.)
 Genetic Algorithm for selecting parameters in the SVMR Model (Contd.,)

 Step 3. (New Population)


 (i) [Selection]: two parent chromosomes are selected according to
fitness function
 (ii) [Crossover]: cross over the parents to form new offspring .
the intermediate crossover principle was used
 (iii) [Mutation]: the variable is selected randomly, and
incremental operator is used
 The rate of crossover and mutation was set to 0.8 and 0.05
respectively
 (iv) [Accepting]: place new offspring in the new population

 Step 4. (Replace): New population is used for further run of the


algorithm
 Step 5. (Stop Condition): satisfy end condition otherwise go to Step 2
27
SUMMARY OF WORK (Contd.)
 Values of optimal parameters for GA-SVMR (Patil et al. 2012) and GA-
SVMR-1 models with different kernels

Model Kernel nsv


C  d
GA-SVMR (Patil et al. 2012) 2131 40000 0.05 -
spline
GA-SVMR -1 2131 40000 0.05 -

GA-SVMR (Patil et al. 2012) 2131 15 0.05 2


b-spline
GA-SVMR-1 2131 350 0.05 1

28
SUMMARY OF WORK (Contd.)
 Performance of GA-SVMR(Patil et al 2012) and GA-SVMR-1 model
for wave transmission (Kt) prediction of HIMMFPB

Test Data

Model
CC Test RMSE SI

GA -SVMR (spline) (Patil et al 2012) 0.9735 0.03671 0.06549

GA –SVMR-1 (spline) 0.9748 0.03477 0.05933

GA-SVMR (b-spline) (Patil et al 2012) 0.9741 0.03629 0.06474

GA-SVMR-1 (b-spline) 0.9749 0.03468 0.05918

29
SUMMARY OF WORK (Contd.)
 Comparison of predicted and measured Kt for GA-SVMR-1 (spline) model
1
CC Test = 0.9748
0.9

0.8

0.7

0.6
Predicted Kt

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Me asured Kt

30
SUMMARY OF WORK (Contd.)
 Comparison of predicted and measured Kt for GA-SVMR-1 (b-spline) model
1
CC Test = 0.9749
0.9

0.8

0.7

0.6
Predicted Kt

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Measured Kt

31
CONCLUSIONS

 Presently developed models ANN-1 and GA-SVMR-1 with


spline and b-spline kernel function has shown better
generalisation capability than ANN and GA-SVMR models.

 Replacing incident wave steepness (Hi/L) with relative


depth of water (d/L) has improved the performance of
ANN-1 and GA-SVMR-1 models.

 It is also observed that input parameter selection in


developing ANN and GA-SVMR models has a significant
effect on the performance

32
CONCLUSIONS

 The performance of CI models depends on choice of kernel


function and good setting of SVM and kernel parameters.
The b-spline kernel function performed better than other
kernel function.

33
SELECTED REFERENCES
 Adee, B.H. and Martin, W. (1974). Theoretical analysis of floating breakwater
performance. Proc. of the floating breakwater conference, Rhode Island.:21-
40.
 Briggs, M., Ye, M., Demiribilek, Z. and Zhang, J. (2002). Field and numerical
comparisons of the RIBS floating breakwater. Journal of Hydraulic Research.
40(3):289 -301.
 Chen, K. and Wiegel, R.L. (1970). Floating breakwaters for reservoir marinas.
Proc. 12th Coastal Engineering Conference, Washington D.C.:647–1666.
 Drimer, N., Agnon, Y. and Stiassnie, M. (1992).
A simplified analytical model for a floating breakwater in water of finite depth
. Applied Ocean Research. 14:33-41.
 Gunn, S.R. (1998). Support vector machines for classification and regression.
University of Southampton, Technical report, Image speech and intelligent
Systems group
 Hegde, A.V., Kamath, K., Magadum, A.S. (2007). Performance characteristics
of horizontal interlaced multilayer moored floating pipe breakwater. Journal
of Waterways, Port , Coastal and Ocean Engineering. ASCE. 133(4):275-286.

34
SELECTED REFERENCES
 Ito, Y. and Chiba, S. (1972). An approximate theory of floating breakwaters.
Report of the Port and Harbor Research Institute, Ministry of Transport,
Japan.11(2):138-213. (cross reference from Tsinker).
 Jain, P. and Deo, M.C.(2006). Neural networks in ocean engineering, SAOS.
1(1):25-35.
 Koraim, A., S. (2013). Hydrodynamic efficiency of suspended horizontal
rows of half pipes used as a new type breakwater. Ocean Engineering. 64:1-
22.
 Loukogeorgaki, E. and Agenlides, D.C. (2005). Performance of Moored
Floating Breakwaters. International Journal of Offshore and Polar
Engineering. 15(4):264-273.
 Murlikrishna, I.N., Raman, H. and Sundar, V. (1987). Performance of rigid
floating breakwaters. Indian National Conf. on Harbour and Ocean
Engineering.: 67-79.
 Patil, S.G., Mandal, S., Hegde, A. V. (2012). Genetic algorithm based
support vector machine regression in predicting wave transmission of
horizontally interlaced multi-layer moored floating pipe breakwater.
Advances in Engineering Software. 45:203-212.
35
THANK YOU

36
Additional slides on
SVMR

37
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR)
(Contd.,)
y  f ( x)  wii ( x)  b
i (x)
wi b
 Where is the non linear function in feature of inputs , and
both and are coefficients which are estimated by
minimizing the regularized risk function:
1 1 N

 L (d
2
Minimize : R (C )  w C i , yi )
2 N i 1

Where
 d i  yi   . d i  y i   ,
L ( d i , y i )   
 0, others, 

38
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR)
(Contd.,)
i  i*
 Introducing two positive slack variables and it is possible to
transform it into a primal objective function given by:
1 N
 C  ( i   i* )
* 2
Minimize : R ( w,  i ,  i )  w
2 i 1

d i  wi ( xi )  b     i ,
wi ( xi )  b  d i     i*
where  i ,  i*  0

39
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR) (Contd.,)

The loss margin setting corresponds to one dimensional linear SV machine

40
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR)
(Contd.,)
 To optimize the primal objective function Lagrange function is
constructed by introducing the dual set of variables
1 N N
L( wi ,  i ,  ,  i ,  ,  i ,  )  w  C  ( i   i )    i wi ( xi )  b  d i     i 
* * * 2 *
i i i
2 i 1 i 1

   (  
N N
   d i  wi ( xi )  b    
*
i i
*
i i   i* i* )
i 1 i 1

 Above equation is minimized with respect to primal variables and


maximize with respect to nonnegative Lagrangian multipliers

41
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR)
(Contd.,)
 Finally Karush- Kuhn - Tucker Condition is applied to the
regression
N
and thus
N
* yields the dual 1Lagrangian
N N
J ( i ,  i )   d i ( i   i )     i   i )   ( i   i* )( j   *j ) K ( x, xi )
* *

i 1 i 1 2 i 1 j 1
N

 (
i 1
i   i* )  0, 0   i i*  C

 i   i*  0
*
 i and multipliers
 In above equation Lagrangian i must satisfy
 After calculating an optimal desired weights
N
vectorw
of the *

 regression
( hyper
i 1
 * )-plane
K ( x is, x
i i
represented
) i
as
j

42
SVMR Model
 Fundamentals of Support Vector Machine Regression (SVMR) (Contd.,)

 Therefore , the regression function could be

N
*
f ( x,  ,  )   i i ) K ( xi , x j )  b
(
i 1
  *

 The role of the kernel function simplifies the learning process by


changing the representation of the data in the input space to a linear
representation in a higher-dimensional space called a feature space.

 A suitable choice of kernel allows the data to become separable in the


feature space despite being non-separable in the original input space.

43

You might also like