Neurocomputing 142 (2014) 16–29

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

Design and construction of an autonomous underwater vehicle

Khairul Alam n, Tapabrata Ray, Sreenatha G. Anavatti
School of Engineering and Information Technology, University of New South Wales, UNSW Canberra, ACT 2600, Australia

art ic l e i nf o a b s t r a c t

Article history: Autonomous underwater vehicles (AUVs) are becoming increasingly popular for ocean exploration,
Received 30 June 2013 military and industrial applications. In particular, AUVs are becoming an attractive option for underwater
Received in revised form search and survey operations as they are inexpensive compared to manned vehicles. Previous attempts
10 November 2013
on AUV designs have focused primarily on functional designs while very little research has been directed
Accepted 9 December 2013
Available online 14 May 2014
to identify optimum designs. This paper presents an optimization framework for the design of AUVs
using two state-of-the-art population based optimization algorithms, namely non-dominated sorting
Keywords: genetic algorithm (NSGA-II) and infeasibility driven evolutionary algorithm (IDEA). The framework is
Evolutionary computation subsequently used to identify the optimal design of a torpedo-shaped AUV with an overall length of
Design optimization
1.3 m. The preliminary design identified through the process of optimization is further analyzed with the
Autonomous underwater vehicle
help of a computer-aided design tool, CATIA to generate a detailed design. The detailed design has since
then been built and is currently undergoing trials. The flexibility of the proposed framework and its
ability to identify optimum preliminary designs of AUVs with different sets of user requirements are also
demonstrated.
& 2014 Elsevier B.V. All rights reserved.

1. Introduction the context of AUV design as existing design tools are not
sufficiently robust and/or fast to be used within an optimization
Autonomous underwater vehicles (AUVs) have become an scheme [6]. For this reason, previous attempts on AUV designs
interesting area of oceanic research because of their promising reported in the literature have focused primarily on functional
uses in military applications, homeland security, hydrographic designs and non-optimal designs are often adopted and accepted
surveys, mineral field surveys, environment monitoring and ocea- as an option. As the desire/need to minimize the use of resources
nographic studies. Over the years, there have been intensive (e.g. fuel, building cost, and time for design) become increasingly
efforts toward the development of sea-worthy AUVs to meet the important, optimization approaches become increasingly popular
challenges of oceanographic exploration and exploitation pro- [7]. For example, Husaini et al. [2] have worked on the AUV hull
grams. Recently AUV development is focused on improving the design by using numerical method to reduce the drag. Yamaguchi
operation range and endurance for long term data collection in the et al. [3] have used numerical simulations based on the finite
fields of oceanography and coastal management [1]. The growth difference method to optimize the body shape of their vehicle in
of cooperative AUV application has increased the significance of order to reduce the resistance. Alvarez et al. [6] have investigated
optimum energy consumption of AUVs [2]. The AUV's mission is the optimum hull shape of an underwater vehicle moving near the
often compromised due to the on-board limited energy storage free surface using simulated annealing (SA) optimization techni-
capacity. The vehicle must save energy during the period of que while Joung et al. [8] have employed computational fluid
observation to be able to cruise longer distances over a period dynamics (CFD) analysis with the same goal. Beside that, Martz
of time. For this purpose, a body shape with low resistance is and Neu [9] have developed a design optimization process for
desirable [2,3]. Therefore, in the preliminary stage of AUV design, an AUV using a multiple objective genetic optimization (MOGO)
hull resistance reduction is one of the key design targets [4]. algorithm.
While optimization techniques have been applied successfully AUV design optimization is not a new concept, but it poses
to a wide range of applications spanning various fields of science difficult computational problems. The design space is often large
and engineering, there is a very limited literature on optimization with functions and constraints exhibiting severe nonlinearities. An
of AUV designs [5]. Hydrodynamic optimization is rarely applied in ad hoc process for making these critical design decisions is not
adequate for the design of highly integrated AUV systems. There-
fore, much work still needs to be done in terms of optimizing the
n
Corresponding author. Tel.: þ 61 2 6268 8479; fax: þ61 2 6268 8276. AUV hull form design to minimize drag and increase propulsion
E-mail address: [email protected] (K. Alam). efficiency [8].

http://dx.doi.org/10.1016/j.neucom.2013.12.055
0925-2312/& 2014 Elsevier B.V. All rights reserved.
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 17

While the major thrust has been on hull form optimization of The rest of the paper is organized as follows. The proposed
AUVs, limited attention has been paid on the arrangement strate- optimization framework is described in Section 2. In Section 3,
gies within the hull items, i.e. how to optimally place the internal the details of the numerical experiments are given, followed by
on-board components in a clash-free state while maintaining a discussion of results (optimum designs) obtained. Finally, the
appropriate clearance among them, and other factors that affect findings of this study are summarized in Section 5.
controllability, like the centre of gravity (CG) and the centre of
buoyancy (CB) effects. The term clash-free is defined here as the
placement of the internal on-board components without an over- 2. Optimization framework
lap and with appropriate clearance among them. AUVs use on-
board computers, power packs and vehicle payloads for automatic This work presents an optimization framework for the design
control, navigation and guidance. Also the AUVs can be equipped of AUVs based on the given design requirements. The framework
with state-of-the-art scientific sensors to measure oceanic proper- incorporates a geometry and configuration modules, a hydrody-
ties, or specialized biological and chemical payloads to detect namics module, several accepted maritime performance and
marine life when in motion [10]. The small AUV, REMUS [11] characteristics estimation methods of AUVs and a suite of optimi-
which has a low internal payload volume, requires small-sized zation algorithms. The design optimization process starts with the
sensors. The viability of using such AUVs is largely dependent on initialization of a set of solutions. Such solutions are randomly
the availability of small AUV compatible sensors [10]. The AUV generated using the variable bounds and fed into the geometry
should possess some free internal volume for adequate buoyancy and configuration modules. The geometry and configuration
[12]. Thus it is important to consider the internal arrangement of modules generate a design for each of these solutions using the
components on-board and their optimum placement to ensure catalogue information which in turn translates to a hull form
space, weight, appropriate location of the CG and CB of the AUV, in geometry and location of the components. These modules not only
addition to its external shape and size. These establish the need for generate the external hull geometry but also place the internal on-
the development of a framework that is capable of generating the board components in a clash-free state. Once the internal parts are
optimum design of AUVs by simultaneously considering both placed in a clash-free state, the parallel mid-body geometry is
internal clash-free arrangement of on-board components and generated to cover the internal arrangement, and then the nose
external size and shape for a given set of design requirements. and tail cone geometries are coupled along with the mid-body,
In summary, the AUV design still poses several challenges and thereby generating the complete vehicle geometry. The perfor-
the aim of the present work is to identify efficient and seaworthy mance of the candidate design is then evaluated. The optimization
AUVs for a given set of user requirements. To this end, this paper process runs till the assigned number of function (design) evalua-
focuses on the development of an optimization framework for tions is completed. The optimization process has an option to
representing various torpedo shaped AUV geometries through visualize every candidate design. The detailed flowchart of the
seamless integration of Matlab-CATIA and in-house performance optimization framework is presented in Fig. 1 with further
analysis codes for reduced design cycle time and added flexibility. discussion of its components in the subsequent sections.
The objective is to find an appropriate hull shape to minimize drag The framework consists of five applications namely Matlab,
and optimum clash-free placement of the internal objects for CATIA, Microsoft Excel, Text Document and VBScript. Matlab is
optimal CG–CB separation thereby ensuring better controllability used for numerical computation/optimization. CATIA (Computer
of a vehicle moving submerged near the free surface while Aided Three-dimensional Interactive Application) is a multi-
fulfilling the design constraints. Since during the phase of optimi- platform computer-aided design (CAD) software tool used to
zation, the components are represented as rectangular bounding model and visualize candidate designs. Microsoft Excel and Text
boxes, the vehicle dimensions are essentially preliminary esti- Document are used as a medium of communication between
mates. During the phase of detailed design, the orientation of the applications. VBScript (Visual Basic Scripting Edition) is used for
objects and their actual geometry are considered and modified. CATIA automation wherein the 3D model of the candidate design
The final design is then built to resolve remaining uncertainties in can be generated without user intervention. Fig. 2 shows a generic
the design process. sequence diagram to illustrate the work flow.

Optimization
Design Design NSGA-II
Requirements Catalogue IDEA

Geometry and
Configuration Module

Candidate Design
Internal Arrangement
External Geometry

No Check No
Visualize
Performance Estimation Termination
Design
Criteria

Yes Yes

Best
3D Model Building
Design

Fig. 1. Detailed flowchart of the optimization framework.

18 K. Alam et al. / Neurocomputing 142 (2014) 16–29

Optimization Communication 3D Model Building xt

xn
Platform: Matlab
Medium: Excel, Platform: CATIA nn
Text Document Automation: VBScript yt
yn d
Fig. 2. Inter-process communication flow among applications.

2.1. Design requirements

ln lm lt
The basic design concept of the AUV employed herein is based l
on the development of a small vehicle that can be easily launched, Fig. 3. Parameterization of the hull geometry.
recovered and operated without special handling equipments
capable of working at shallow water depth. The design process
relies heavily on the use of off-the-shelf components in an attempt Two bilge pumps for Two bilge pumps for
Propeller unit for
pitch movement yaw movement
to contain cost. In the development stage, the performance testing Forward motion
of the AUV will be conducted in UNSW Canberra indoor swimming
pool of dimensions of 25 m and at a depth of 1.2 m. The major
potential areas of application of such class of AUVs include
environmental monitoring, scientific data collection, information
operations, and oceanographic survey. Based on the potential
applications and operability, the design requirements used in this
Payload section Battery unit
study are the following:
Fig. 4. General arrangement of the AUV system.
 Operating speed of the AUV is to be 2 m/s which is commonly
used for surveying [13]. most on-board components and/or payloads and batteries are
 Length of the AUV must be no more than 1.3 m so that it can rectangular in shape. In terms of dynamics, long cylindrical objects
cruise within the indoor swimming pool and can be handled by exhibit low drag compared to other geometries of similar length to
two individuals. The downsized design and the cruising ability diameter ratios [15]. From the simple perspective of drag reduc-
in a bounded space enable fast experimental feedback on tion, a ‘torpedo body’ that has a nose cone followed by a parallel
control, navigation and guidance system design of AUVs [14]. cylindrical mid-section and a tapered tail section is likely to be the
 Total weight of the AUV (including payload) must be less than best choice.
or equal to 30 kg which is the limit imposed by safe carrying The effects of nose shape on body drag have been calculated
load of two individuals [14]. numerically for a family of shapes [16]. Illustrations of the family
 The vehicle is to be propelled by one rear propeller, and the chosen for this study are presented in Fig. 3, where the body is
yaw and pitch movements will be achieved through the use seen to comprise a nose-section of variable length ln, a mid-section
of four bilge pumps. This is a typical example of a combined of variable length lm, and a tail-section of variable length lt, making
propulsion system (combination of jet-pumps and propeller) up the total body length of l units.
that avoids complex rudder control system for steering and The curve shapes of the nose and tail sections are determined
elevation. respectively from the Eqs. (1) and (2):
 Should have enough space to carry a controller unit, two 
n 1=nn
batteries and sensor payloads. 1 ln xn n
yn ¼ d 1 ð1Þ
 The AUV must be able to store lead shot for ballast and weight 2 ln
in order to balance the buoyant force for underwater use.
 where yn is the radius of the nose in metres, d is the maximum
Should maintain a modular configuration for easy access to the
body diameter in metres, which may be varied, ln is the length of
internal components and be reconfigured to suit various mis-
the nose in metres, xn is the reference length (in metres) that
sion requirements.
varies from 0 to ln, and nn is the shape variation coefficient of the
nose which may also be varied to give different shapes of the nose.
2.2. Geometry and configuration modules 2yt ¼ 9:6556x3t 4:9x2t þ 0:05xt þ d ð2Þ

2.2.1. Hull geometry where yt is the radius of the tail in metres and xt is the reference
There are many aspects to be considered when developing a length (in metres) that varies from 0 to lt.
hull design. First and the foremost is that the AUV must be capable The mid-body will hold all internal components in a clash-free
of housing and maintaining a watertight environment for the state and thrust motor in the tail. The general preliminary arrange-
integrity of its on-board internal systems. It is also required that ment of the parts of the AUV is shown in Fig. 4.
this environment be easily accessible for maintenance and in most
cases, the hull must also allow for modularity. The hull weight and
the drag on it should be low. 2.2.2. Geometry of the payload section
Inspired by torpedoes and submarines [15], most AUVs tend to The AUVs use on-board computers, sensors and vehicle pay-
opt for a cylindrical hull shape. The cylinder is considered to be the loads for automatic control, navigation and guidance. Therefore,
optimum geometry because of its ability to handle the pressure adequate free space should be allocated while designing the
and dynamic forces applied to the hull. A round hull presents no AUV hull form. Fig. 5 shows the parameterization of the payload
stress concentrations and thus for a certain material a cylindrical section. The AUV considered in the present study should be able to
object is able to withstand more pressure when compared to house a 150  150  200 mm3 payload box.
other shapes, except for the sphere. Even though a spherical form Subject to preliminary design requirements, the optimization
provides the best structural efficiency, it is considered as ineffi- framework developed herein produces AUV hull shape in a way
cient in terms of usable internal space. This is due to the fact that that can accommodate a variable position payload. The capability
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 19

Chamber for two Chamber for two

bilge pumps for Payload bilge pumps for
pitch movement compartment yaw movement

d
a b
Wet Dry Wet

lm
Fig. 5. Parameterization of the payload section. Nose Chamber Tail
section separating walls section

Two bilge pumps for Two bilge pumps for

Fig. 7. Typical configuration of the internal chambers.
pitch movement yaw movement Propeller unitf or
Forward motion
Due to simplicity and commercial availability, we limit our selec-
tion to the consideration of primary and secondary batteries as the
power source. After a careful review of the available battery
systems presented by Bradley et al. [18], the lead acid batteries
with nominal cell voltage of 12 V are proven to be the best option
for the power source of the present application. These batteries are
Flow direction
the simplest to use, cheap, small and usually safe. They also meet
Fig. 6. Configuration of the propulsion system. the vehicle size constraints.

of adjusting the position of the payload is an important design 2.2.5. Arrangement of internal components
consideration to account for CG–CB distance manipulations. The design optimization framework not only optimizes the hull
shape but also arranges its contents avoiding interference while
maintaining workable spaces around the components using ‘clash-
2.2.3. Propulsion system
free mechanism’. For efficient utilization of the available internal
Currently available AUVs are predominately propelled by three
volume, a careful arrangement needs to be achieved. While in
means: propeller, jet-pump and buoyancy driven (glider) propulsion
reality some components have irregular shapes, minimum bound-
systems. Although a majority of the current AUVs are equipped
ing box dimensions have been used to derive the clash-free
with a propeller and rudder mounted at the tail for propulsion
configuration. Details of the clash-free mechanism appear in Alam
and steering, a jet based steering system is used in this study to
et al. [19].
demonstrate jet-pumps as an option. Jet-pumps are more beneficial
than rudder in terms of mechanical design (absence of rotating parts
and transmission mechanisms), controlling complexity, cost realiza-
tion and robustness with respect to transportation and safety for 2.2.6. Waterproofing and manufacturing complexity
occasional swimmers in proximity of the vehicle. One problem commonly encountered in manufacturing AUVs
The forward motion of the AUV under consideration is achieved relate to ensuring watertightness between dry and wet sections.
through the use of a propeller unit located at the tail. There Fig. 7 shows the configuration of the dry and wet compartments of
are four bilge pumps of two different types mounted inside the the AUV considered in this study.
vehicle, two of which are used to steer the vehicle left and right, The term manufacturing complexity is defined as the number
and the remaining pumps are used to move the vehicle up and of watertight walls that are required to seal the dry sections from
down, as illustrated in Fig. 6. the wet ones. Ideally, the lower the number, the better it is from a
Considering the required speed of 2 m/s, optimization of the manufacturing point of view.
entire vehicle needs to be done in order to reduce drag to a
minimum while increasing thrust to a maximum. Therefore, the
2.3. Hydrodynamics module: drag estimation
position of the four bilge pumps for vertical and lateral move-
ments as illustrated in Fig. 6 is not fixed, but is rather free to move
A fundamental interest in the field of hydrodynamics is the
within the entire mid-section in order to maximize the lever arms.
reduction of submerged vehicle power requirements [20], and
The term ‘lever arm’ defined here as the longitudinal distance of
drag minimization is one of the most effective means of achieving
the pump from the CB of the vehicle. The higher value of lever arm
this. In this section, the details of the drag estimation are
produces higher pitching and turning moments that lead to better
discussed.
diving and heading changes. During the optimization process, the
The drag experienced by an axisymmetric underwater vehicle
optimizer chooses these pumps from a catalogue and the optimal
moving forward is the viscous drag comprising the skin friction
positions of those pumps while designing the AUV.
drag and the form drag [8,21]. The skin friction drag is due to the
viscous shear of the fluid flowing along the hull, and the form drag
2.2.4. Power source caused by development of a boundary layer and the resulting
Unlike tethered vehicles, AUV operations are limited by the difference of pressure distribution between the bow and the aft of
on board power [17]. The power determines the endurance and the vehicle.
operating range, and the mission characteristics of the AUV The bare hull skin friction drag coefficient (CF) as a function
accordingly. Most underwater vehicles in use today are powered of Reynolds number (Rn) is found using the 1957 ITTC correlation
by low cost rechargeable batteries. As the present application has line [22] as
several design constraints such as the vehicle should be no longer
than 1.3 m and total weight should be not more than 30 kg, other 0:075
CF ¼ ð3Þ
source of power such as fuel cells and solar power are not suitable. ðlog 10 Rn  2Þ2
20 K. Alam et al. / Neurocomputing 142 (2014) 16–29

where Rn is the Reynolds number and can be found as solutions is first done based on non-dominance, followed by
ρVl crowding distance. While the non-dominance based rank drives
Rn ¼ ð4Þ the population towards the Pareto Optimal Front, the crowding
μ
distance based rank aims to preserve the diversity among solu-
where ρ is the density of the fluid in kg/m3, V is the velocity in m/s, tions. For problems with constraints, the feasible solutions are
l is the overall length of the vehicle in metres and μ is the dynamic ranked above infeasible solutions. The infeasible solutions are
viscosity of the fluid in kg/(m s). ranked based on their constraint violations. For the single objec-
Although the empirical estimation of drag provides useful tive formulation, NSGA-II assumes the form of a real coded
information for the powering requirements at the early stage of evolutionary algorithm (EA) with SBX and polynomial mutation.
design [8], the classical formulas are often less reliable and have The main steps of NSGA-II are outlined in Algorithm 1. The
limitations regarding the vehicle geometries [23,24]. From the algorithm starts with an initial population (P1) of N candidate
available predictive methods reported in the literature, three solutions initialized by random sampling from the design (vari-
methods namely Virginia Tech (VT) [25], MIT [26] and G&J method able) space. Each candidate solution of the population is evaluated
[22] are employed in this study to compute the coefficient of to yield the corresponding values of the objective and the
viscous resistance (CV) in three different ways in order to increase constraint functions. Based on the objective and constraint func-
the reliability in empirical estimation of drag. tion values, the candidate solutions are ranked. The next steps
According to G&J method, the coefficient of the viscous resis- (lines 5–8) are repeated for NG generations. An offspring popula-
tance, CV, for the smooth bare hull can be found as tion, Ci, is created using a recombination operation from the
"

3 # current (or parent) population P i  1 . The new offspring solutions
d d
C V ¼ C F 1 þ 0:5 þ3 ð5Þ are evaluated and the combined set of parents and offspring
l l
solutions is ranked. Based on the ranks, the best solutions from
where d is the maximum body diameter in metres. the parent population P i  1 and the offspring population Ci
According to VT method, the CV is found from the modification are retained to form the population for the next generation Pi.
of the original Gillmer and Johnson equation, Eq. (5), to take into As the processes of evolution, evaluation, ranking and reduction
account the effects of nose and tail shape variation coefficients, nn are repeated, the population in successive generations contains
and nt, respectively as better and better solutions.
"

7  nn  nt =2 #
d d Algorithm 1. Non-dominated Sorting Genetic Algorithm II
C V ¼ C F 1 þ 0:5 þ3 ð6Þ
l l (NSGA-II).

In MIT method, the prismatic coefficient, CP, has been used to Require N; {Population size}
find the CV as Require N G r 1; {Number of generations}
V env 1: Initialize (P1); {Create an initial populations of solutions}
C P ¼  2 ð7Þ 2: Evaluate (P1);
πd
2 l
3: Rank (P1); {Assign ranks to each solution}
where Venv is the displacement volume of the vehicle (in m3). 4: for i¼ 2 to NG do
Then, 5: C i ¼ EvolveðP i  1 Þ; {Create child solutions from parents of
"
3=2
3 # previous generation}
d d 6: Evaluate Ci; {Compute the performance of the child
C V ¼ C F 1 þ 1:5 þ7 þ 0:002ðC P 0:6Þ ð8Þ
l l solutions}
Once the CV is known, then the vehicle drag, D, in Newtons (N) 7: Rank (P i  1 þ Ci); {Assign ranks to each solution}
can be calculated as 8: Pi ¼Reduce ðP i  1 þC i Þ; {Identify parents for the next
generation}
D ¼ 12 ρV 2 C V S ð9Þ 9: end for
where S is the wetted surface area of the vehicle (in m2).
As the primary objective of the proposed optimization appro- 2.4.2. Infeasibility Driven Evolutionary Algorithm (IDEA)
ach is to minimize the drag of the submerged vehicle, the Infeasibility Driven Evolutionary Algorithm (IDEA) was recently
maximum drag value out of the three methods is chosen as the proposed by Singh et al. [28]. Since the Pareto optimal solution for
objective function of the optimization exercise. a constrained problem usually lies on a constraint boundary, IDEA
tries to focus the search near the constraint boundaries by
2.4. Optimization modules maintaining a set of infeasible solutions (in addition to feasible
solutions). During the ranking process, a few marginally infeasible
The framework developed here uses two state-of-the-art opti- solutions (based on their constraint violations) are ranked above
mization algorithms, NSGA-II and IDEA. The algorithms are written the feasible solutions. The presence of infeasible solutions effec-
in Matlab and are integrated with CATIA along with VBScript tively translates to approaching the constraint boundary from both
to automate the whole AUV design process. It is worth noting feasible and infeasible search spaces, which helps in faster con-
that any optimization algorithm capable of solving single and vergence, as demonstrated in the earlier study [29].
multi-objective optimization problems can be used within the The key steps of IDEA are outlined in Algorithm 2. IDEA uses
framework. SBX and polynomial mutation operators [27] to generate offspring
from a pair of parents selected using binary tournament as in
2.4.1. Non-dominated Sorting Genetic Algorithm II (NSGA-II) NSGA-II. Individual solutions in the population are evaluated and
Non-dominated Sorting Genetic Algorithm II (NSGA-II) was infeasible solutions are identified. The solutions in the parent
proposed by Deb et al. [27], and remains one of the most widely and the offspring population are divided into a feasible set, Sf,
used population based algorithms for evolutionary optimization. and infeasible set, Sinf. The solutions in the feasible set and the
It uses Simulated Binary Crossover (SBX) and polynomial muta- infeasible set are both ranked separately using the non-dominated
tion for generating off-spring candidate designs. The ranking of sorting and the crowding distance sorting [27]. The solutions for
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 21

the next generation are selected from the sets to maintain feasible requirements. Constraints have also applied on the lengths of the
solutions in the population. In addition, the infeasible solutions are dry chambers in order to ensure that there are enough rooms to
ranked higher than the feasible solutions to provide a selection place watertight walls between dry and wet sections. The objec-
pressure in order to create better infeasible solutions which will tive function, constraints and design variables are listed in Eq. (10)
result in an active search through the infeasible search space. as Formulation 1:

Algorithm 2. Infeasibility Driven Evolutionary Algorithm (IDEA). Minimize

f ð1Þ ¼ D
Require N {Population Size}
Require N G 41 {Number of Generations} Subject to
Require 0 o α o 1 {Proportion of infeasible solutions} gð1Þ ¼ l r 1300 mm; gð2Þ ¼ w r30 kg
1: Ninf ¼ αnN gð3Þ ¼ LA1 Z 250 mm; gð4Þ ¼ LA2 Z 250 mm
2: Nf ¼ N  N inf
gð5Þ ¼ s r 4 mm; gð6Þ ¼ ld1 Z40 mm
3: pop1 ¼Initialize()
4: Evaluate(pop1) gð7Þ ¼ ld2 Z 40 mm
5: for i¼2 to NG do Variable bounds :
6: childpopi  1 ¼Evolve ðpopi  1 Þ
0 r X C r 50 mm; 0 r Y C r50 mm; 0 r Z C r 800 mm
7: Evaluate(childpopi  1)
8: ðSf ; Sinf Þ ¼ Splitðpopi  1 þ childpopi  1 Þ 0 r X BL r 50 mm; 0 r Y BL r 50 mm; 0 r Z BL r 800 mm
9: Rank(Sf) 0 r X BS r 50 mm; 0 rY BS r 50 mm; 0 r Z BS r800 mm
10: Rank(Sinf) 0 r X P1 r50 mm; 0 r Y P1 r 50 mm;
11: popi ¼ Sinf ð1; Ninf Þ þ Sf ð1; N f Þ
0 r X P2 r50 mm; 0 r Y P2 r 50 mm; 0 rZ P r 800 mm
12: end for
0 r X Y1 r 50 mm; 0 rY Y1 r 50 mm;
0 r X Y2 r 50 mm; 0 rY Y2 r 50 mm; 0 rZ Y r 800 mm
A user-defined parameter, α, is used to maintain a set of infeasible 1:5 r nn r 3; 250 rln r 350 mm
solutions as a fraction of the size of the population. The numbers 280 rlt r 400 mm ð10Þ
Nf and Ninf denote the number of feasible and infeasible solutions
respectively determined by parameter α. If the infeasible set, Sinf, has The design variables of the problem formulation and the
more than Ninf solutions, the first Ninf solutions are selected based on derived variables for constraints checking are illustrated in Fig. 8
their ranks, otherwise all the solutions from Sinf are selected. The rest showing exactly which parts of the AUV geometry they define.
of the solutions are selected from the feasible set, Sf, provided that There are 22 design variables that define the external geometry
there are at least Nf feasible solutions. If Sf has fewer solutions, all the and locations of the on-board components of the AUV. Also, seven
feasible solutions are selected and the rest are filled with infeasible constraints have been imposed to meet the basic design require-
solutions from Sinf. The solutions are ranked from 1 to N in the order ments. The detail descriptions of the design variables and the
they are selected. Hence, the infeasible solutions selected first will be derived variables for constraints checking are presented respec-
ranked higher than the feasible solutions selected later. tively in Tables 1 and 2.
Among the 22 design variables, 19 of them are used to place the
on-board components in a clash-free state. The maximum body
3. Numerical experiments diameter (d) is derived from the parallel mid-body that is required
to cover the internal arrangements. After that, nose and tail cones
In this section, the formulations of the optimization problems are attached with the mid-body and the overall length (l) of
based on the design requirements are presented. While satisfying the AUV is computed. Once the complete shape of the AUV is
the design requirements, optimization of two kinds of objectives is modelled, then other parameters like surface area (S) and dis-
sought. First is the minimization of drag (D), which is important placement volume (Venv) are calculated for the overall geometry
because minimum drag leads to least power consumption for and the vehicle drag (D) is obtained using Eq. (9) by substituting
propulsion, and corresponding savings in the operating costs. the values of the required parameters.
Second is to maximize one of the lever arms, as a longer lever
arm assists in better diving and turning of the vehicle. Hence two
formulations are studied in this paper as described below. LA1 LA2
Ld1 Ld2
s
3.1. Single objective formulation nn

The single objective optimization problem posed can be d CB

described as the identification of a vehicle hull form with mini- Z CG
mum drag as well as clash-free optimal placement of the internal X
objects subject to the constraints on length and weight of the
vehicle, lever arms and CG–CB separation. The first ðLA1 Þ and ln lm lt
second ðLA2 Þ lever arms are the longitudinal distances of the ZP
pumps from the centre of buoyancy respectively for pitch and ZC
yaw movements. The higher value of lever arm produces higher ZBL
pitching and turning moments that lead to better diving and ZBS
heading changes. The lower the value of CG–CB separation (s), the ZY
closer the position of the CG and the CB that leads to better
l
stability of the vehicle. Constraints on overall length (l) and weight
(w) of the vehicle are also important to meet the basic design Fig. 8. The constraints and design variables for problem formulation.
22 K. Alam et al. / Neurocomputing 142 (2014) 16–29

Table 1
Design variables for formulating the problem.

Variable Description

XC, YC, ZC Coordinates of the controller unit along X-, Y- and Z-axis respectively
XBL, YBL, ZBL Coordinates of the large battery unit along X-, Y- and Z-axis respectively
XBS, YBS, ZBS Coordinates of the small battery unit along X-, Y- and Z-axis respectively
XP 1, YP 1 Coordinates of the first pump for pitch along X- and Y-axis respectively
XP 2, YP 2 Coordinates of the second pump for pitch along X- and Y-axis respectively
ZP Coordinate of the pumps for pitch along Z-axis
XY 1, YY 1 Coordinates of the first pump for yaw along X- and Y-axis respectively
XY 2, YY 2 Coordinates of the second pump for yaw along X- and Y-axis respectively
ZY Coordinate of the pumps for yaw along Z-axis
nn Shape variation coefficient of the nose
ln Length of the nose
lt Length of the tail

Table 2 Table 3
Derived variables for constraints checking. Parameters used for the experiment.

Variable Description Parameter Value

l Overall length of the AUV Population size 300

w Total weight of the vehicle Maximum function evaluations 60,000
LA1 Length of the first lever arm Crossover probability 0.9
LA2 Length of the second lever arm Crossover distribution index 10
s Longitudinal distance between CG Mutation probability 0.1
and CB Mutation distribution index 20
ld 1 Length of the first dry chamber Infeasibility ratio, α (only IDEA) 0.2
ld 2 Length of the second dry chamber

CG: centre of gravity; CB: centre of buoyancy.

4. Results and discussion

The results obtained from the optimization exercises are pre-

3.2. Multi-objective formulation sented in this section. Firstly, the results of the single and multi-
objective optimization problems are summarized and an optimum
The multi-objective optimization problem is posed as the preliminary AUV design is identified. Then, a detailed CAD analysis
identification of an AUV hull form with minimum drag and has been carried out as a post-process in order to enhance the
maximum length of one of the lever arms as well as clash-free performance of the preliminary design and to make the design
optimal placement of the on-board components. The objective viable to be built. The fabrication process describes the develop-
functions are listed in the Eq. (11) as Formulation 2, while the ment of a tooling and manufacturing process aimed at ensuring
constraints and design variables are the same as in the single accurate, workable and efficient construction of the individual
objective case, except the total number of constraints for multi- modules. Finally, an in-depth 3D CFD analysis of the preliminary
objective formulation has reduced to five as the lever arms are and detailed designs is also carried out in order to assess the true
considered as objective functions. performance of the optimum designs.
In this study the overall numerical procedure is implemented
Minimize
in Matlab. The computing time per evaluation (without invoking
f ð1Þ ¼ D; f ð2Þ ¼  LA ð11Þ CATIA) is about 0.05 s in an Intel Xeon processor machine of
3.33 GHz with 6.00 GB memory.
where LA ¼ minðLA1 ; LA2 Þ.
It can be noted that the optimization algorithms employed in
this study are formulated to minimize the objective functions
4.1. Single objective optimization results
of a problem. Therefore, to maximize the length of lever arm in
Formulation 2, a negative sign has been placed; thereby formulat-
The single objective optimization problem (Eq. (10)) is a
ing both the objectives as a minimization problem.
constrained optimization of a continuous function. It is important
to assess the nature of the problem and to this effect 60,000
3.3. Computational setup random solutions were generated using the variable bounds.
It is clear from Fig. 9 that none of the randomly generated
For each formulation, 30 independent runs using NSGA-II and solutions were feasible. Since typical local search methods require
IDEA are performed, as carrying out multiple runs is a usual a feasible starting point, their use for this problem is not particu-
practice in the field of evolutionary computation. Results are larly attractive. It is clear that population based stochastic algo-
presented based on these multiple runs by varying the seed value rithms would fair well for these classes of problems. Hence,
while keeping the other parameters like crossover probability, two population-based stochastic evolutionary algorithms (global
mutation probability, crossover distribution index and mutation search methods) namely NSGA-II and IDEA have been chosen for
distribution index constant for all runs. The values of the para- the current problem.
meters are listed in Table 3. The values of the key parameters Fig. 10 shows the progress plots of the best and median runs for
(Table 3) used for NSGA-II and IDEA are based on the recom- the single objective drag minimization problem using NSGA-II
mended values from the literature [28–30]. and IDEA. Also, the standard error bars illustrating the difference
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 23

of standard deviation values between the two EA approaches are The performance metric averaged over multiple runs of final
presented in Fig. 11. solutions obtained using NSGA-II and IDEA is listed in Table 5. It is
It can be observed from Fig. 10a that both optimization seen that the best value of IDEA is marginally worse than that of
algorithms are able to minimize the drag in around 110 genera-
tions. The statistics of results computed across 30 runs for each
algorithm is reported in Table 4. It is seen that the best, the median 24
and the average objective values obtained by IDEA are better than NSGA−II
IDEA
NSGA-II. In addition, the standard deviation across the multiple 22
runs is much less for IDEA than NSGA-II, indicating that it is able
to achieve better objective values more consistently. This is also
20
reflected in the median runs as shown in Fig. 10b, where IDEA is
seen to converge faster than NSGA-II.
18

Drag (N)
4.2. Multi-objective optimization results
16

Multi-objective optimization i.e. minimization of drag and maxi-

mization of one of the lever arms is performed using NSGA-II and 14
IDEA. To quantify the performance of the AUV design problem, one
commonly used performance metric has been used to compare the 12
results obtained from NSGA-II and IDEA. This performance metric
is called the hypervolume [31] metric which is a measure of 10
volume (or area) dominated by the given non-dominated set with 0 50 100 150 200
respect to a reference point. The larger this dominated area, the Generations
better is the non-dominated set. In the present study, the refer-
Fig. 11. Standard error bars of the median designs for single objective drag
ence point is constructed by taking the maximum value of each minimization using NSGA-II and IDEA.
objective across all the runs.

100 96.61%
Table 4
89.36% Single objective drag minimization results.
90 87.9%
84.68%
81.74%
Design NSGA-II IDEA
80
72.63%
Number of violations (%)

Best (N) 10.2770 9.69995

70 64.92% Mean (N) 12.5926 10.8650
Median (N) 14.0680 10.3398
60 Worst (N) 14.6182 14.4393
SD (N) 1.9182 1.5916
50
SD: standard deviation.
40

30
Table 5
20 Performance metric for two objectives formulation.

10 Algorithm Hypervolume metric

0 Best Mean Median Worst

1 2 3 4 5 6 7
Constraint NSGA-II 759.60 308.23 249.56 56.15
IDEA 745.70 412.17 555.28  393.75
Fig. 9. Calculation of constraint violations for 60,000 random solutions.

24 19
NSGA−II NSGA−II
IDEA 18 IDEA
22
17
20
16
18
Drag (N)

Drag (N)

15
16
14
14
13
12
12
10 11

8 10
0 50 100 150 200 0 50 100 150 200
Generations Generations

Fig. 10. Progress plots of the best (a) and median (b) designs for single objective drag minimization using NSGA-II and IDEA.
24 K. Alam et al. / Neurocomputing 142 (2014) 16–29

−50 −50
Reference Reference
NSGA−II NSGA−II
−100 IDEA −100 IDEA

−150 −150

(−) Lever arm (mm)

(−) Lever arm (mm)

−200 −200

−250 −250

−300 −300

−350 −350

−400 −400
8 10 12 14 16 18 20 22 8 10 12 14 16 18 20 22
Drag (N) Drag (N)

Fig. 12. Final non-dominated fronts for the best and median runs of the AUV design problem using NSGA-II and IDEA.

Two bilge pumps for Two bilge pumps for

pitch movement Controller Unit yaw movement

Fig. 13. Bare hull of the preliminary design.

NSGA-II. But on average, IDEA is able to achieve much higher

values of hypervolume as compared to NSGA-II. This indicates that
IDEA is able to achieve comparatively better non-dominated sets. Watertight wall Sliding board Battery Unit Main Motor
The non-dominated fronts for the best and median runs obtained Fig. 14. Internal configuration of the preliminary design.
using the two algorithms are shown in Fig. 12.
Once the optimization process is done, then the non-
dominated set of solutions can be presented to the decision maker
Table 6
(end user), who can choose a design most suited to his/her needs.
Performance criteria of the preliminary and detailed designs.
For example, in this case the decision maker can choose from a set
of optimal designs with minimum drag, or maximum lever arm Vehicle particulars Preliminary Detailed
length, or an intermediate value of both the objectives. Thus the
multi-objective formulation can cater to a number of end users, Nose length (mm) 250 250
Parallel middle body length (mm) 738 738
while ensuring that all the basic design criteria are met. Tail length (mm) 312 312
Length overall (mm) 1300 1300
4.3. Results of the preliminary AUV design Maximum diameter (mm) 252 250
Length to diameter ratio 5.2 5.2
Maximum dimension of the inner square (mm) 170 169
Based on the results obtained by carrying out an optimization
Wetted surface area (m2) 1.06939 1.06157
exercise, Figs. 13 and 14 show the best shape and the internal Displacement volume (m3) 0.029577 0.029194
configuration of the preliminary AUV design respectively. The Mass of the displaced water (kg) 29.60 29.20
preliminary design refers to the best final design obtained from Total mass of the vehicle (kg) 25 25
the single objective drag minimization exercise using NSGA-II Length of the first lever arm (mm) 366 369
Length of the second lever arm (mm) 250 369
and IDEA. X-coordinate of CG (mm) 18.4182 18.4624
The resulting performance criteria of the preliminary best Y-coordinate of CG (mm) 3.11559 3.12187
design are listed in Table 6. The values of the lever arms, CG-CB Z-coordinate of CG (mm) 679.297 682.555
separation, length and weight of the vehicle clearly indicate that X-coordinate of CB (mm) 0 0
Y-coordinate of CB (mm) 0 0
the design constraints are satisfied while achieving minimum
Z-coordinate of CB (mm) 678.098 679.235
drag. Longitudinal distance between CB and CG (mm) 1.199 3.32
Nominal speed (m/s) 2 2
4.4. Detailed design of the AUV Drag (VT method) (N) 9.04479 8.97321
Drag (G&J method) (N) 9.20864 9.13294
Drag (MIT method) (N) 9.69995 9.61129
The focus of the detailed design is to finalize the preliminary
design with the actual components to be built. In order to verify CG: centre of gravity; CB: centre of buoyancy; VT: Virginia Tech; G&J: Gillmer and
the preliminary design and move the data into a useful form, Johnson; MIT: Massachusetts Institute of Technology.
CATIA is used to show the location of the components and a visual
representation of the detailed design with all the actual on-board can be placed together. The detailed design demonstrates and
components is illustrated in Fig. 15. provides a tighter assembly yet clash-free. In the final design, the
It is important to note that the clash-free mechanism adopted in pumps for yaw motion are placed with pitch pumps in the bow
the present study, despite the fact that the internal components and the controller and battery units are shifted towards aft of the
are not necessarily rectangular, they have been represented as AUV in order to keep the CG–CB separation within 4 mm limit. The
the minimum bounding boxes. Another limitation of the current major consequence of these changes are the reduced diameter of
approach is the inadequacy of rotated placement of the on-board the final design, higher lever arms and a lower drag of the AUV as
components. These obviously limit the way in which the objects compared to the preliminary design that reported in Table 6.
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 25

Two bilge pumps for  Fittings: The completed vehicle along with its all on-board
pitch movement Controller Unit Watertight wall
components is shown in Fig. 16a. The four bilge pumps are
attached to the nose cone which is used to steer the vehicle.
Inside the dry section contains two batteries, control box,
sliding rail board and sufficient free space to store lead shot
for ballast and weight in order to balance the buoyant force.
The sliding board has holes along it so that the batteries can be
placed anywhere along it to shift the CG forward or backward
as needed. The propeller is mounted in the tail cone using four
Two bilge pumps for Sliding board Battery Unit Main Motor
yaw movement
aluminum strips and fastened by nuts and bolts to the PVC
collar to strength.
Fig. 15. Internal configuration of the final design.  Testing: Performance testing is conducted in a swimming pool
at UNSW Canberra and depicted in Fig. 16c. The vehicle is
It is worth highlighting that for a detailed design, no optimiza- trimmed to neutral buoyancy and balanced to sit horizontally
tion exercise has been carried out. The preliminary design identi- leveled in the water using small bags of lead shot. The vehicle
fied through the process of optimization provides the design proves to be precisely manoeuvrable, demonstrating operating
boundaries for the detailed design phase (length to diameter ratio, speed of approximately 2 m/s. After conducting a couple of
total weight, allowable distance between CB and CG, etc.). A tests, it is concluded that the AUV does work and meets all the
detailed CAD based analysis using CATIA is carried out in the design requirements.
detailed design phase. The actual shapes of the on-board compo-
nents are drawn prior to the analysis. The internal components are
placed in their respective locations as determined by the pre-
4.6. Design analysis using CFD
liminary estimates and found that the diameter can be reduced
further by 2 mm. Also the pumps for yaw can be placed along with
The Computational Fluid Dynamics (CFD) analysis is becoming
the pumps for pitch motion that eventually contributes to the
more and more popular in estimating the hydrodynamic para-
increased lever arm lengths. To compensate the CG to CB distance,
meters such as drag, although it is computationally more expen-
the rest of the components are moved towards the aft of the
sive than the empirical ones [9]. CFD can provide accurate
vehicle. To automate the CAD analysis, a macro is written using
simulations of the flow around the vehicle, and a useful under-
VBScript that takes the locations of the on-board components as
standing of the fluid–structure interactions [32]. Therefore 3D
input and provides an assembly design and the corresponding
CFD analysis of the preliminary and detailed designs for better
performance criteria for the detailed design as outputs. To com-
drag estimation has been carried out following the optimization
pute the drag of the final design, a Matlab function is used (the
approach.
same function used during the course of optimization).

4.5. Fabrication 4.6.1. Theoretical background of the CFD analysis

The numerical analysis of flow around submerged bodies
The manufacturing process is broken up into hull components is based on the incompressible Navier–Stokes equations that
fabrication, control system design, fittings and finally testing, as describe the flow properties such as the velocity, pressure,
described below. temperature and density. Discretizing the equations with respect
to time produces the Reynolds Averaged Navier–Stokes (RANS)
 Nose and tail cone: The first step is to take the best design of the equations that are commonly used in most commercially available
nose and the tail, and create a mould to the desired shape that CFD packages. The commercial CFD software package ANSYS 13.0
the nose and the tail cone have to follow. Then the layering of has been used in the current study to simulate the flow around the
fiberglass along with resin is done onto the mould. After the designed vehicles.
fiberglass and resin had set, both the nose and the tail cone are
coated with a special type of car putty. This is to ensure a
4.6.2. Model building and flow domain
smooth finish as well as its strength and waterproof integrity.
The bare hull shape is drawn in CATIA based on the hull
Finally, the nose and the tail cone are coated with a car primer
parameters and exported to ICEM CFD as a step file for meshing.
and then finished with a top coat of paint.
 As the object is axisymmetric, the quarter model of the bare hull is
Mid-body: The mid-body of the AUV is made from a 250 mm
considered for CFD analysis to reduce the computational costs.
poly vinyl chloride (PVC) water pipe as this material is easily
Fig. 17 shows the full scale and quarter models of the submarine
available, cheap, strong, easy to work with and offers good
bare hull. The computational far field is extended 16l upstream of
resistance to corrosion. The mid-body holds the two batteries,
the leading edge, 18l downstream from the trailing edge and 10l
payload and the control box. The remaining length is kept to fit
above the body surface, where l being the length of the vehicle (in
two 25.5 mm watertight walls and has 30 mm length of PVC
metres). Since the flow is incompressible, the considered solution
pipe clear for fittings to attach the nose and tail cones.
domain is found large enough to capture the entire viscous–
To seal the AUV, a special PVC piping glue is used that plastic
inviscid interaction and the wake development.
welds the collars and the pipe together to ensure a watertight
seal. A multipurpose co-polymer sealant is also used on either
side of the collars to ensure water security. Then two bulk 4.6.3. Grid generation
heads are screwed on to the collars in order to isolate the dry After defining the model and the far field, the solution domain
chamber from the wet ones. is decomposed into appropriate number of locations based on the
 Control system: With the prototype, the primary goal is to accuracy of the results required. Geometry meshing in this study is
create a control system that is simple to install and easy to use. generated by using ANSYS ICEM CFD Meshing software. The mesh
This led to the use of a 36 GHz, four channel remote control file is then exported to FLUENT for numerical study. The generated
system capable of working at the depth of 3–4 m. mesh is composed primarily of tetrahedral elements. Typically in
26 K. Alam et al. / Neurocomputing 142 (2014) 16–29

Fig. 16. Fabrication of the AUV: (a) AUV component layout; (b) the completed vehicle ready for testing; and (c) AUV in operation during testing.

Fig. 17. CATIA model of the AUV bare hull: (a) full model and (b) quarter model.

total 2,747,167 regular elements were built with 707,331 nodes in enhanced wall treatment is applied to simulate the surrounding
the solution domain. flow of the vehicle. This model is quite robust, economic and
In flow simulations using the turbulence models, the computa- reasonably accurate for a wide range of turbulent flows [8].
tional grid should be in such a way that a sufficient number of grid
points lie within the laminar sublayer of the ensuing boundary
layer [33]. Therefore, ‘prism’ elements with 50 layers are selected 4.6.5. Results of CFD analysis
for generating meshes adjacent to the body surfaces as these The drag values obtained for the preliminary and detailed
are the most appropriate for a boundary layer mesh [8]. Also to designs computed using ANSYS (FLUENT) under turbulence flow
capture the wake, the dense mesh is selected near the stern of the regimes with flow speed of 2.0 m/s are presented in Table 7.
vehicle body as illustrated in Fig. 18. The values of drag force obtained from both empirical and CFD
methods as reported in Table 7 show good agreement. Thus the
ranking of the designs is consistent with the values obtained using
4.6.4. Problem setup and simulation the empirical relations.
When the computational domain is meshed, the flow is solved Other than drag estimation, the CFD analysis is also performed
using the software ANSYS FLUENT 13.0. The Reynolds numbers with the purpose of obtaining the pressure and velocity distribu-
of the surrounding flow of the vehicle are typically 2.6  106. From tions around the vehicle under the influence of streaming water. A
the various turbulence models, the realizable k–ϵ model with visualization of the pressure distribution of the flow around the
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 27

Y Y

Z X Z X

Z X

Fig. 18. Generated mesh highlighting the boundary layer grid adjacent to the vehicle body: (a) mesh of the flow domain around the vehicle; (b) enlarged view of mesh
adjacent to the vehicle body (the boundary layer); and (c) enlarged view of dense mesh near the stern side.

Table 7
 Optimum design generated while considering clash-free inter-
Comparison of drag values between empirical and CFD estimates. nal arrangement and effects of external shape.
 Coupled with a suite of state-of-the-art optimization algo-
Design Empirical method CFD analysis rithms e.g. NSGA-II and IDEA. Other optimizers can easily be
Preliminary 9.69995 N 9.85289 N
included.
Detailed 9.61129 N 9.44751 N  Seamless integration of Matlab and CATIA, and in-house
performance analysis codes.
 Ability to solve various forms of single and multi-objective,
body of the designed vehicles is shown in Fig. 19. An even constrained formulation of the AUV design problems based on
distribution of pressure can be seen along the main body except user requirements.
for the stagnation point at the nose-tip of the hull. The pressures at  Modularity and catalogue driven structure allows for design of
the nose and tail sections are higher compared to the pressure AUVs of various sizes, propulsion and power systems.
along the main body of the AUV.
The velocity contours of flow around the surfaces of the
designed vehicles are shown in Fig. 20. It can be seen that the The aspects of both single and multi-objective design processes
velocity around the nose is lower and the flow is accelerated as it are illustrated using an example. The complexity of the problem
reaches the stern. This can be explained from the conservation of stems largely from very low feasibility, wherein manual identifica-
energy, in a steady flow, an increase in the pressure of the fluid tion of a feasible design itself is non-trivial. Since the problem
occurs simultaneously with a decrease in the velocity. As the is highly constrained, population based stochastic optimization
pressure at the nose of the AUV is higher, the velocity is lower at algorithms are a preferred choice over conventional local search
this region. As the outlet gauge pressure is set to zero and also due approaches.
to the shape of tail, the flow-stream converges when it reaches the The ability of the proposed approach to represent and optimize
stern and velocity increases. a class of underwater vehicles is illustrated through designing a
torpedo-shaped AUV with an overall length of 1.3 m. Furthermore
the design process relies heavily on the use of off-the-shelf
5. Summary and conclusions components to contain both risk and cost. The current approach
is embedded with a clash-free mechanism that relies on the mini-
Presented in this paper is an efficient evolutionary approach for mum rectangular bounding box dimensions, despite the irregular
the preliminary design of AUVs. The framework is the first of its shapes of the internal on-board components. In an attempt to
kind to offer full multidisciplinary design optimization function- enhance the performance of the preliminary design, a detailed
alities to be considered in the design of AUVs and offers several CAD analysis has been carried out that overcomes the limitations
advantages over traditional functional design approaches. Further- of the preliminary approach. The final design is then built to
more, the modular framework allows the designer to incorporate resolve remaining uncertainties in the design and tested in an
additional modules and control the complexity of the optimization indoor swimming pool. The AUV met the specified design require-
problem definition. The key aspects of the proposed framework ments and successfully demonstrated the operability at shallow
are summarized as follows: water depth. The preliminary and detailed designs are also vali-
dated using CFD analyses to establish confidence on the empirical
 Able to represent various torpedo shaped AUV geometries. estimates used during the course of optimization. Future work will
28 K. Alam et al. / Neurocomputing 142 (2014) 16–29

Fig. 19. Pressure contours around the surfaces of the designed vehicles: (a) preliminary design and (b) detailed design.

Fig. 20. Velocity contours around the surfaces of the designed vehicles: (a) preliminary design and (b) detailed design.

address the problem of designing a full autonomous control [9] M. Martz, W.L. Neu, Multi-objective optimization of an autonomous under-
system for the final design. water vehicle, Mar. Technol. Soc. J. 43 (2) (2009) 1–9.
[10] E. Desa, R. Madhan, P. Maurya, Potential of autonomous underwater vehicles
as new generation ocean data platforms, Curr. Sci. 90 (9) (2006) 1202–1209.
[11] B. Allen, R. Stokey, T. Austin, N. Forrester, R. Goldsborough, M. Purcell,
Acknowledgements C. von Alt, REMUS: a small, low cost AUV; system description, field trials
and performance results, in: Oceans Conference Record (IEEE), vol. 2, 1997,
pp. 994–1000.
The first author would like to acknowledge the Research [12] D. Walker, Micro autonomous underwater vehicle concept for distributed data
Publication Fellowship he received from the University of New collection, in: OCEANS 2006 IEEE, 2006, pp. 1–4.
South Wales, Canberra, Australia. The second author would like [13] B. Anderson, J. Crowell, Workhorse AUV - A cost-sensible new autonomous
underwater vehicle for surveys/soundings, search & rescue, and research, in:
to acknowledge the support of Future Fellowship offered by the Proceedings of MTS/IEEE OCEANS, 2005.
Australian Research Council. The authors would also like to [14] B.-H. Jun, J.-Y. Park, F.-Y. Lee, P.-M. Lee, C.-M. Lee, K. Kim, Y.-K. Lim, J.-H. Oh,
acknowledge the suggestions and comments from the reviewers Development of the AUV ‘ISiMI’ and a free running test in an Ocean
Engineering Basin, Ocean Eng. 36 (2009) 2–14.
in improving this work.
[15] W. Wang, X. Chen, A. Marburg, J. Chase, C. Hann, A low-cost unmanned
underwater vehicle prototype for shallow water tasks, in: Proceedings of the
IEEE/ASME International Conference on Mechatronic and Embedded Systems
References and Applications, MESA 2008, Beijing, China, 2008, pp. 204–209.
[16] D.F. Myring, A theoretical study of body drag in subcritical axisymmetric flow,
[1] M. Husaini, Z. Samad, M.R. Arshad, Autonomous underwater vehicle propeller Technical Report, Aeronautical Quarterly 27, Royal Aircraft Establishment
simulation using computational fluid dynamic, in: Computational Fluid (RAE), Hants, UK, 1976, pp. 186–194.
Dynamics Technologies and Applications, InTech, 2011, pp. 293–314. [17] V.P. Shah, Design considerations for engineering autonomous underwater
[2] M. Husaini, Z. Samad, M.R. Arshad, CFD simulation of cooperative AUV motion, vehicles (Master's thesis), Massachusetts Institute of Technology, 2007.
Indian J. Mar. Sci. 38 (3) (2009) 346–351. [18] A.M. Bradley, M.D. Feezor, H. Singh, F.Y. Sorrell, Power systems for autono-
[3] S. Yamaguchi, T. Kawanami, W. Koterayama, A study on shape optimization for mous underwater vehicles, IEEE J. Ocean. Eng. 26 (4) (2001) 526–538.
an underwater vehicle based on numerical simulation, in: Proceedings of the [19] K. Alam, T. Ray, S.G. Anavatti, A new robust design optimization approach for
Fifth (2002) ISOPE Pacific/Asia Offshore Mechanics Symposium, Daejeon, unmanned underwater vehicle design, Proc. Inst. Mech. Eng. Part M: J. Eng.
Korea, 2002, pp. 43–48. Marit. Environ. 226 (3) (2012) 235–249.
[4] S. Fangxi, Z. Lianhong, W. Ziliang, W. Leping, On resistance calculation for [20] J.S. Parsons, R.E. Goodson, F.R. Goldschmied, Shaping of axisymmetric bodies for
autonomous underwater vehicles, Adv. Mater. Res. 189–193 (2011) 1745–1748. minimum drag in incompressible flow, J. Hydronautics 8 (3) (1974) 100–107.
[5] K. Alam, T. Ray, S. Anavatti, An evolutionary approach for the design of [21] A. Phillips, M. Furlong, S.R. Turnock, The use of computational fluid dynamics
autonomous underwater vehicles, in: Proceedings of the AI 2012: Advances in to assess the hull resistance of concept autonomous underwater vehicles, in:
Artificial Intelligence, Lecture Notes in Computer Science, vol. 7691, Springer, Oceans Conference Record (IEEE), 2007, pp. 1–6.
Berlin, Heidelberg, 2012, pp. 279–290. [22] T.C. Gillmer, B. Johnson, Introduction to Naval Architecture, 2nd Edition, US
[6] A. Alvarez, V. Bertram, L. Gualdesi, Hull hydrodynamic optimization of Naval Institute Press, Annapolis, MD, 1982.
autonomous underwater vehicles operating at snorkeling depth, Ocean Eng. [23] A. Phillips, M. Furlong, S.R. Turnock, The use of computational fluid dynamics
36 (2009) 105–112. to determine the dynamic stability of an autonomous underwater vehicle, in:
[7] A.F. Mohamad Ayob, T. Ray, W. Smith, A hydrodynamic preliminary design Proceedings of the 10th Numerical Towing Tank Symposium (NuTTS'07),
optimization framework for high speed planing craft, J. Ship Res. 56 (1) (2012) Hamburg, Germany, 2007, pp. 1–6.
35–47. [24] H. Zhang, Y.R. Xu, H.-P. Cai, Using CFD software to calculate hydrodynamic
[8] T. Joung, K. Sammut, F. He, S.-K. Lee, A study on the design optimization of an coefficients, J. Mar. Sci. Appl. 9 (2) (2010) 149–155.
AUV by using computational fluid dynamic analysis, in: Proceedings of the [25] M.A. Martz, Preliminary design of an autonomous underwater vehicle using a
Nineteenth (2009) International Offshore and Polar Engineering Conference, multiple-objective genetic optimizer (Master's thesis), Virginia Polytechnic
Osaka, Japan, 2009, pp. 696–702. Institute and State University, Blacksburg, VA, United States, 2008.
 K. Alam et al. / Neurocomputing 142 (2014) 16–29 29

[26] P.E.C.U.R. Jackson, H.A., MIT Professional Summer Course “Submarine Design Tapabrata Ray is an ARC Future Fellow with the School
Trends”, 1992. of Engineering and Information Technology, University
[27] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective of New South Wales, UNSW Canberra, Australia. He
genetic algorithm: NSGA-II, IEEE Trans. Evolut. Comput. 6 (2) (2002) 182–197. received his Ph.D. from the Indian Institute of Technol-
[28] H.K. Singh, A. Isaacs, T. Ray, W. Smith, Infeasibility driven evolutionary ogy (IIT) Kharagpur. He is the founder and leader of the
algorithm (IDEA) for engineering design optimization, in: 21st Australasian Multidisciplinary Design Optimization Research Group
Joint Conference on Artificial Intelligence: Advances in Artificial Intelligence, at UNSW, Canberra. He has published extensively in
Auckland, New Zealand, Lecture Notes in Artificial Intelligence, vol. 5360, the area of evolutionary computation and design opti-
Springer-Verlag, Berlin, Heidelberg, 2008, pp. 104–115. mization. His research areas include multidisciplinary
[29] T. Ray, H.K. Singh, A. Isaacs, W. Smith, Infeasibility driven evolutionary algor- design optimization, evolutionary computation, multi-
ithm for constrained optimization, in: E. Mezura-Montes (Ed.), Constraint- objective optimization, constrained optimization, robust
Handling in Evolutionary Optimization, Studies in Computational Intelligence, design optimization, surrogate assisted optimization,
vol. 198, Springer, Berlin, Heidelberg, 2009, pp. 145–165. shape representation and optimization, bioinspired
[30] H. Singh, A. Isaacs, T.T. Nguyen, T. Ray, X. Yao, Performance of infeasibility models for optimization.
driven evolutionary algorithm (IDEA) on constrained dynamic single objective
optimization problems, in: IEEE Congress on Evolutionary Computation, CEC
'09, 2009, pp. 3127–3134.
[31] J. Bader, E. Zitzler, HypE: an algorithm for fast hypervolume-based many- Sreenatha G. Anavatti is a Senior Lecturer with the
objective optimization, Evolut. Comput. 19 (1) (2011) 45–76. School of Engineering and Information Technology,
[32] R.P. Irwin, C. Chauvet, Quantifying hydrodynamic coefficients of complex University of New South Wales, UNSW Canberra,
structures, in: Proceedings of the OCEANS 2007, Aberdeen, UK, 2007, pp. 1–5. Australia. He received his Ph.D. from the Indian Insti-
[33] M.M. Karim, M.M. Rahman, M.A. Alim, Numerical computation of viscous drag tute of Science (IISc), India. His research areas include
for axisymmetric underwater vehicles, J. Mek. 26 (2008) 9–21. flight dynamics, design of autopilots for aircraft and
missiles, flexible spacecraft dynamics and control,
active vibration control, application of fuzzy and neural
network for practical applications, and unmanned
Khairul Alam is a Ph.D. candidate in the School of underwater vehicle design.
Engineering and Information Technology at the Uni-
versity of New South Wales, UNSW Canberra, Australia.
He obtained his Master of Science in Information
Technology and Mechatronics of Vehicles from the
University of Ulsan, Ulsan, Republic of Korea and
Bachelor of Science in Mechanical Engineering from
Chittagong University of Engineering and Technology,
Chittagong, Bangladesh. His research interests include
multidisciplinary design optimization, multi-fidelity
and surrogate assisted optimization, shape representa-
tion and optimization, and underwater vehicle design.