Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

ANFN Controller Based On Differential Evolution

for Autonomous Underwater Vehicles

O. Hassanein*1, S. A. Salman **2, Sreenatha G. Anavatti*3, T. Ray*4


*
School of Engineering and Information Technology, UNSW@ADFA, Canberra, ACT, Australia
**
Mechanical Engineering Department, Assiut University, Assiut, Egypt
1
o.hassan@adfa.edu.au
2
salman@aun.edu.eg
3
s.anavatti@adfa.edu.au
4
t.ray@adfa.edu.au

Abstract- The Autonomous Underwater Vehicles (AUVs) sources limit the mission time of the vehicle and autonomy
dynamics have six degrees of freedom and are highly nonlinear limits the degree to which an AUV can be left unattended by
and time varying and the hydrodynamic coefficients of vehicles human operators.
are difficult to estimate accurately because of the variations of PID controller is one of the most widely used control
these coefficients with different navigation conditions and system in real industry control field and has the advantages of
external disturbances such as currents and waves. The path rapid response, good accuracy and stabilization. PID controller
controller of the AUV is a challenging problem due to the [3] is used to control AUV’s heading traditionally. PID
nonlinearities and uncertainties of the AUV dynamics. Thus, the controller is able to achieve control targets for whom can get
controller should be adaptive to handle variations in the accurate model of definite system in math, with permanent
dynamics of the AUV at different maneuvering regimes and work condition and small disturbances. Choosing a proper PID
disturbances arising from both the internal and external sources. parameters becomes a problem here since PID controller
In the present paper Adaptive Neural Fuzzy Network (ANFN)
design is based on a certain vehicle model. Most of the
commercial autopilots for AUVs have PID controllers [4-6]
controller is designed and applied to guide and control the AUV.
due to their simplicity in structure, reliability and ease of
Initially, the controller parameters are generated randomly and design. However, their performance is not all that satisfactory
tuned by Differential Evolution algorithm (DE). The back due to the presence of nonlinearities, disturbances and time
propagation algorithm based upon the error between the actual varying parameters [7,8]. In practice, if a variety of
outputs of the plant and the desired values is then used to adopt immeasurable disturbances are present due to multi-directional
the controller parameters online. The proposed ANFN controller currents or vehicle velocity changes during the operation which
adopts a functional link neural network (FLNN) as the change the hydrodynamic coefficients of the vehicle, the
consequent part of the fuzzy rules. Thus, the consequent part of robustness of PID controller will become worse. Hence, there
controller is a nonlinear combination of input variables. The is a necessity of a computationally efficient controller in the
results show that the performance of the AUV with the ANFN presence of these undesirable characteristics. Intelligent and
controller is having better dynamic performance as compared to adaptive control systems are the most appropriate control
strategies for the AUV where an accurate model of the
the conventional PID, even in the presence of noise and
controlled process is vaguely known.
parameter variations
A neural fuzzy system for nonlinear system control is
Keywords-AUV; AUV dynamic model; neuro-fuzzy control;
neuro-fuzzy mdelling; 6 DOF.
discussed in [9]. This controller is applied to the planetary-
train-type inverted pendulum system and the magnetic
I. INTRODUCTION levitation system in the VisSim solving nonlinear control
problems. The FLNN is a single-layer neural structure capable
The AUV is an underwater system that contains its own of forming arbitrarily complex decision regions by generating
power and is controlled by an onboard computer. The nonlinear decision boundaries with nonlinear functional
fundamental task for these devices is fairly well defined: The expansion. The FLNN [10] was conveniently used for function
vehicle is able to follow a predefined trajectory to reach the approximation and pattern classification with faster
final destination. The main advantage of an AUV is that it convergence rate and less computational loading than a
does not need a human operator. Therefore it is capable of multilayer neural network. Recently, genetic fuzzy systems
doing operations that are too dangerous for a person. They [11,12] have received increasing attention mainly because they
operate in conditions and perform task that humans are not able combine the approximate reasoning method of fuzzy systems
to do efficiently or at all [1]. with the learning capabilities. However, the search is extremely
According to [2] the two most significant technological time-consuming, which is one of the basic disadvantages of all
challenges in AUV’s design are power and autonomy. Power genetic algorithms (GAs).

978-1-4673-4439-5/12/$31.00 2012
c IEEE 184
Since 1995, Differential evolution (DE) has drawn the terms. IJ is a 6x1 vector including the control forces and
attention of many researchers all over the world resulting in moments. . For more details, readers are referred to [16].
many variants of the basic algorithm with improved
performance. DE is arguably one of the most powerful
stochastic real-parameter optimization algorithms. However,
unlike traditional Evolution Algorithms (EAs), the DE-variants
perturb the current generation population members with the
scaled differences of randomly selected and distinct population
members [13]. Reference [14] proposed a generalized hybrid
generation scheme to enhance the exploitation and accelerate
the convergence velocity of the original DE algorithm. Figure 1. UNSW@ADFA AUV.

This study proposes an Adaptive Neural Fuzzy Network


(ANFN) controller is designed and applied to guide and control
the AUV and compared with the conventional PID controller.
The powerful tools of DE are combined with the non-linear
mapping properties of ANFN to design suitable controller for
the AUV. In the present work AUV has been developed and
built at UNSW@ADFA. The hydrodynamic coefficients of the
vehicle are calculated under certain conditions using Figure 2. Six degrees of freedom of an AUV.
computational fluid dynamics (CFD) method to derive the
exact mathematical model. Fig. 1 shows a typical III. PID CONTROLLER DESIGN
UNSW@ADFA AUV during one of the experiments. One Initially, a conventional PID controller is designed to
electrical thruster powers the vehicle for forward motion. Two control the AUV. The block diagram for this is shown in Fig.
electrical pumps used for maneuvering in the horizontal plane. 3. Even though PID controllers are simple, tuning the gains to
In addition, two electrical pumps help the AUV to navigate in achieve the desired performance is a cumbersome process due
the vertical plane. The middle box is used for carrying the to the trial and error nature. For the present AUV, there are
sensors, battery and the electronic accessories. three-feedback control loops that represent surge, pitch and
This paper is organized as follows. The general dynamic yaw motions. For these controllers, there are nine gains that
equation describing the mathematical model of the AUV is need to be selected, i.e. tuned, to achieve the desired response.
provided in Section 2. Section 3 gives the brief description of Those three AUV controllers are then the control actions, the
the PID controller designed. The design details of the controller actions power all the pumps at right, left, up and down of the
of AUV using ANFN technique are presented in section 4 vehicle and the propeller that makes it move forwards. The up
whereas the numerical simulation results are discussed in and the down pumps indicate the pitch angle. The right and the
section 5. Finally, the paper is concluded in Section 6. left pumps determine the yaw angle of the body. And finally,
the surge represents the angular speed of the propeller, which
II. AUV’S MATHEMATICAL MODEL will form the mechanical propulsion of the vehicle, so it
Analyzing the dynamics of an AUV, including regulates the forward velocity, u. The PID controller may be
hydrodynamic parameter uncertainties, is a difficult task due to implemented in continuous or discrete time, in a number of
the fact that they are highly nonlinear, coupled and time controller structures [17,18]. The ideal continuous time PID
varying. The AUV can be described by 6 degrees of freedom controller is expressed in Laplace form as follows:
(DOF), which are defined as x (surge), y (sway), z (heave), ĭ ª 1 º
(roll), ș (pitch) and ȥ (yaw). These characteristics determine Gc ( S ) = Kp «1 + + T dS » (2)
¬ T iS ¼
the position and orientation of the AUV. u, v and w represent
the forward, lateral and vertical velocities along x, y and z axes where, Kp = proportional gain, Ti = integral time constant and
respectively. Similarly, the angular rates will be denoted by p, Td = derivative time constant. The gains of the PID controllers
q and r along x, y and z axes respectively. The motion have been tuned by trial and error. Due to its nature, the time
equations of the vehicle are defined with respect to two consumed in the development of the PID controllers has been
coordinate systems [15] as shown in Fig. 2 and can be written significantly high.
as follows [15]:
IV. ANFN CONTROLLER DESIGN
‫ݍܯ‬ሷ ൅ ‫ܥ‬ሺ‫ݍ‬ሶ ሻ‫ݍ‬ሶ  ൅ ‫ܦ‬ሺ‫ݍ‬ሶ ሻ‫ݍ‬ሶ  ൅ ‫ܩ‬ሺ‫ݍ‬ሻ  ൌ ߬ (1) The ANFN uses a nonlinear combination of input variables
where M is a 6x6 inertia matrix as a sum of the rigid body (FLNN) [19] with the fuzzy system. Each fuzzy rule
inertia matrix, MR and the hydrodynamic virtual inertia (added corresponds to the FLNN, comprising a functional expansion
mass) MA. C (q ) is a 6x6 Coriolis and centripetal matrix of input variables. The FLNN, initially proposed by [20], is a
single-layer ANN structure capable of forming complex
including rigid body terms C RB (q ) and terms C A (q ) due to
decision regions by generating nonlinear decision boundaries.
added mass D(q ) is a 6x 6 damping matrix including terms due The functional expansion block comprises of a subset of
to drag forces. G(q) is a 6x1 vector containing the restoring orthogonal polynomials basis function. The FLNN has been
terms formed by the vehicle’s buoyancy and gravitational inserted to the consequent part of the fuzzy rules.

2012 First International Conference on Innovative Engineering Systems 185


A. Functional link neural network structure (FLNN) mathematical model, set of real input-output data or a real
The FLNN is a single-layer network while the input system.
variables generated by the linear links of neural networks are
linearly weighted, the functional link acts on an element of
input variables by generating a set of linearly independent
functions, orthogonal polynomials for a functional expansion,
and then evaluating these functions with the variables as the
arguments. Therefore, the FLNN structure considers
trigonometric functions. In the FLNN structure as shown in
Fig. 4, a set of basis functions ĭ and a fixed number of weight
parameters W represent ݂ܹሺ‫ݔ‬ሻ. The theory behind the FLNN
for multidimensional function approximation has been Figure 3. The block diagram for the AUV control system.
discussed in [20]. The linear sum of the ݆‫ ݄ݐ‬node is given by
‫ܯ‬

‫ݕ‬ො݆ ൌ  ෍ ‫ ݇߮ ݆݇ݓ‬ሺܺሻ ሺ͵ሻ


݇ൌͳ 
where X ∈ A ⊂ ℜ N , X = [ x1 ,..., x N ]T is the input vector and
ܹ݆ ൌ ሾ‫ ͳ݆ݓ‬ǡ ǥ ǥ Ǥ ǡ ‫ ܯ݆ݓ‬ሿܶ is the weight vector associated with Figure 4. FLNN structure.

the ݆‫ ݄ݐ‬output of the FLNN. ‫ݕ‬ො݆  denotes the local output of the
FLNN structure and hence, the consequent part of the ݆‫݄ݐ‬
fuzzy rule in the ANFN model. The m-dimensional linear
output is given by ŷ = WΦ , where ‫ݕ‬ො ൌ ሾ‫ݕ‬ොͳ ǡ ‫ݕ‬ොʹ ǡ ǥ Ǥ ǡ ‫ݕ‬ො݉ ሿܶ , m
denotes the number of functional link bases, which equals the
number of fuzzy rules in the ANFN model.
B. ANFN controller structure
The structure of the ANFN model is presented in Fig. 5.
The ANFN model changes a fuzzy IF–THEN rule in the
following form. Figure 5. NFN structure
ሺŽሻ ǣ ൫šͳ ‹• ͳŽ ƒ† ǥ ǥ ǥ Ǥ ƒ† š ‹• Ž ൯  A. OFF-LINE procedure

This section presents the off-line stage as a first step in
›ොŒ ൌ  ෍ ™‹Œ ij ሺሻ ሺͶሻ optimizing and tuning the controller. After the ANFN structure
‹ൌͳ  has been established, the controller enters the parameter-
ൌ  ™ͳŒ ijͳ ൅  ™ʹŒ ijʹ
൅ ‫ ڮ‬൅ ™Œ ij 
optimization phase based on Differential Evolution (DE) to
where x i and ‫ݕ‬ො݆  are the input and local output variables, adjust the parameters of the network optimally based on a
respectively; FNl is the linguistic term of the precondition part different input-output training data set. DE is known as a
powerful algorithm for real parameter optimization. In DE, an
with Gaussian membership function, N is the number of input
initial population is generated and, for each parent vector from
variables, wij is the link weight of the local output, ϕ M is the
the current population (target vector), a mutant vector (donor
basis trigonometric function of input variables, M is the vector) is obtained. Finally, an offspring is formed by
number of basis function, and rule j is the ݆‫ ݄ݐ‬fuzzy rule. combining the donor with the target vector. A tournament is
then held between each parent and its offspring with the better
In the precondition part, the input is represented by five
being copied to the next generation [21,22]. The DE learning
Gaussian membership functions. In the consequent part, the algorithm consists of four major steps as follows;
output is generated by FLNN. The function expansion in
FLNN uses trigonometric functions, given by a) Initialization step: The first step in DE is the coding
ሾͳǡ šොͳ ǡ •‹ሺʌšොͳ ሻ ǡ ‘•ሺʌšොͳ ሻሿ for one input variable. It leads of the neural fuzzy network parameters into an individual.
to the weight variables, w, is (4x5) matrix. Equation (5) shows the way of the individual coding of NFN
parameters, where i and j represent the i th input variable and
V. TUNING THE ANFN CONTROLLER the ݆‫ ݄ݐ‬rule, respectively. In this study, a Gaussian membership
The proposed technique consists of two phases. The first function is used with m ij and σ ij are the mean and variance of
phase is the off-line procedure and the second one is the on-line a Gaussian membership function, respectively, and w Mj
procedure. The target of the off-line phase is to tune and
optimize the controller’s parameters using an optimization represents the corresponding link weight of the consequent part
technique according to a set of input-output data. The on-line that is connected to the ݆‫ ݄ݐ‬rule node.
procedure aims to adapt the controller online by using the ‫ ݈ܽݑ݀݅ݒ݅݀݊ܫ‬ൌ ݉ͳ݆ ǡ ݉ʹ݆ ǡ ǥ ǥ Ǥ Ǥ ݆݉݅ ǡ ‫ ݆ͳݓ‬ǡ ǥ Ǥ ‫݆ܯݓ‬ ሺͷሻ
input-output measurements of the process which it can be the

186 2012 First International Conference on Innovative Engineering Systems


b) Evaluation step: The objective function is used to network enters the parameter-learning phase to adjust the
provide a measure of how individuals have performed in the parameters of the network based on a different real input-
problem domain. In the minimization problem, the fit output data, mathematical model simulation or real system. The
individuals will have the lowest numerical value of the final output of the ANFN is given by equation (10) with
associated problem objective function. Another Function, the cij and σ ij representing the centre and width of Gaussian
fitness function, is normally used to transform the objective memberships for input variable x i for the rule i, j and N being
function value into a measure of relative fitness. the number of rules of fuzzy model and number of inputs
‫ܨ‬ሺ‫ݔ‬ሻ ൌ ݃ሺ݂ሺ‫ݔ‬ሻሻ  respectively.
ሺ͸ሻ
‫ ݉ݕ‬ሺ݇ ൅ ͳሻ
where f (x) is the objective function, g transforms the value of ݆ ʹ
the objective function to a non-negative number and F is the ‫ ݔ‬െܿ
σܴ݆ൌͳ ‫ݕ‬ො݆ ൥ς݊݅ൌͳ ݁‫ ݌ݔ‬൭െͲǤͷ ቆ ݅ ݆ ݅ ቇ ൱൩
resulting relative fitness. The cost function that attempts to ߪ݅ ሺͳͲሻ
optimize the whole ANFN parameters is ITSE (Integral Time ൌ
݆ ʹ
of Square Error) over the total simulation time. ‫ ݔ‬െܿ
σܴ݆ൌͳ ൥ς݊݅ൌͳ ݁‫ ݌ݔ‬൭െͲǤͷ ቆ ݅ ݆ ݅ ቇ ൱൩
‫ʹݐ‬ ߪ݅
‫ ܧܣܶܫ‬ൌ න ݁ ʹ ݀‫ݐ‬ The ANFN controller is a NFN controller with a tuning
‫ͳݐ‬  ሺ͹ሻ algorithm. As shown in Figure 6, The ANFN controller
‫ ͳܧܬ‬ൌ ‫ݔܽܯ‬ሺ‫ܧܣܶܫ‬ሻ  describes a control law in the form of an If-Then rule-based
F E1 = 1 / J E !  structure. For adjusting the parameters, it is accompanied with
a hybrid online adaptive algorithm. The adaptive algorithm
where J E1 is the objective function of each controller separately utilizes a combination of the error between the reference input
for all population. F E1 is the fitness function of that controller. and the process output.
The objective function using fitness produces distribution in ͳ ʹ
‫ܧ‬ሺ݇ሻ ൌ ቀ‫ ݉ݕ‬ሺ݇ ൅ ͳሻ െ ‫ ݂ݕ‬ሺ݇ ൅ ͳሻቁ ሺͳͳሻ
the range (0, 1). ʹ
where E(k) is error between the ANFN model and the actual
c) Mutation and crossover step: This operation enables plant outputs. If z(k) represents the parameter to be adapted at
DE to explore the search space and maintain diversity. The iteration k in the NFN model, the BP algorithm seeks to
simplest form of this operation is that a mutant vector is minimize the value of the objective function by [21].
generated by multiplying an amplification factor, H by the AN FN

difference between two random vectors and the result is added


to a third random vector (DE/rand/1) [21]as: A daptive Algorithm

ܸሬԦ‫ݖ‬ǡ‫ ݐ‬ൌ ‫ݔ‬Ԧ‫ ͳݎ‬ǡ‫ ݐ‬൅ ‫ ܪ‬ൈ ൫‫ݔ‬Ԧ‫ ʹݎ‬ǡ‫ ݐ‬െ ‫ݔ‬Ԧ‫ ͵ݎ‬ǡ‫ ݐ‬൯  ሺͺሻ
Autonom ous U nderw ater
R e N FN C ontroller u y
V ehicle
where r1 , r2 and r3 are random numbers (1, 2,..., PS), ?e

r1 ≠ r2 ≠ r3 ≠ z , x is a decision vector, PS is the population size,


H is a positive control parameter for scaling the DE and t the Figure 6. Adaptive Neural Fuzzy Network Controller
current generation. For more details, readers are referred to ߲‫ܧ‬
[22]. ‫ݖ‬ሺ݇ ൅ ͳሻ ൌ ‫ݖ‬ሺ݇ሻ െ ߙ ሺͳʹሻ
߲ܼ 
d) Reproduction and selection step: To keep the To train

୧;
population size constant over subsequent generations, the next ݆
ܿ݅ ሺ݇ ൅ ͳሻൌ
step of the algorithm calls for selection to determine whether
݆ ‫ ݉ݕ‬െ‫݂ ݕ‬ ʹሺ‫ ݅ ݔ‬െ‫ ݆ ݔ‬ሺ݇ሻሻʹ ሺͳ͵ሻ
the target or the trial vector survives to the next generation, i.e., ܿ݅ ሺ݇ሻ െ ߙ ሺ‫ݕ‬ො ݆ െ ‫ ݉ݕ‬ሻ‫݈ ݖ‬ ݆ʹ
ܾ ߪ݅ ሺ݇ሻ
at G = G + 1. The selection operation is described as ୨
Ԧ ሬԦ ሬԦ Ԧ To train ı୧ ;
 ܺ݅ǡ‫ܩ‬൅ͳ ൌ ܷ݅ǡ‫݂݂݅ ܩ‬൫ܷ݅ǡ‫ ܩ‬൯ ൑ ݂൫ܺ݅ǡ‫ ܩ‬൯  ሺͻሻ ݆
ߪ݅ ሺ݇ ൅ ͳሻൌ
ൌ  ܺԦ݅ǡ‫݂݂݅ ܩ‬൫ܷ
ሬԦ݅ǡ‫ ܩ‬൯ ൐ ݂൫ܺԦ݅ǡ‫ ܩ‬൯
 ‫ ݉ݕ‬െ‫݂ ݕ‬ ʹሺ‫ ݅ ݔ‬െ‫ ݆ ݔ‬ሺ݇ሻሻʹ ሺͳͶሻ
where f (x) is the objective function to be minimized. ߪ ݆ ሺ݇ሻ െ ߙ ܾ
ሺ‫ݕ‬ො ݆ െ ‫ ݉ݕ‬ሻ‫݈ ݖ‬ ݆͵ 
ߪ݅ ሺ݇ሻ
ʹ
B. ON-LINE procedure ‫ ݅ ݔ‬െ݈ܿ݅
ܾ ൌ σ‫ܯ‬ ݈
݅ൌͳ ܼ Ǣ ܼ ݈ ൌ ς݊݅ൌͳ ݁‫ ݌ݔ‬ቆെͲǤͷ ൬ ൰ ቇ ሺͳͷሻ
ߪ݈݅
The learning algorithm is using the back propagation
technique, BP. The Bp minimizes a given cost function, Eq Similarly, the equation for adapting parameter w is derived
(11), by adjusting the parameter of the membership in the as shown in the following equations.
antecedent part of the ANFN system as mentioned in Eq (13, μ
14) and the link weight parameters of the FLNN in the ™‹Œ ሺ ൅ ͳሻ ൌ ™‹Œ െ Į™ ሺͳ͸ሻ
μ™‹Œ 
consequent part of the ANFN system as in Eq (16, 17). After μ „ijŒ
the NFN parameters have been tuned in the optimization phase ൌ ‡ቆ Œ ቇ ሺͳ͹ሻ
according to a certain input-output training data set, the μ™‹Œ œ 

2012 First International Conference on Innovative Engineering Systems 187


where α w is the learning rate parameter of the FLNN weight. for yaw angle for PID and ANFN controllers, respectively.
The learning rate Įin equation (13, 14) [10] has a significant Generally, it is clear that the performance of the AUV with
effect on the stability and convergence of the system. A higher ANFN controller is more accurate and more effective.
learning rate may enhance the convergence rate but can reduce B. Robustness of controller
the stability of the system. A smaller value of the learning rate
guarantees the stability of the system but slows the In order to appreciate the robustness of the ANFN and PID
convergence. controllers, parameter variations and noise are considered.
In this technique, the fuzzy network is adapted through two a) Parameter Varaition: Fig. 10 illustrates the system
ways. The BP algorithm is applied to tune the membership response when the hydrodynamic coefficients parameters are
function parameters in antecedent part. The consequent part of increased by 10%. As can be seen, the performance is much
the fuzzy rules is adapted through the FLNN. In addition, the better with the ANFN. Also, the PID controller cannot handle
BP algorithm is used for tuning the parameters of FLNN as larger mass variations than this.
well. b) External noise: The system response is investigated when
C. Convergence analysis of ANFN system external noise is applied as sensor noise and sea currents as
To ensure a quick and stable convergence of fuzzy shown in Fig. 11. It is observed that the system based on
controller parameters a convergence analysis of the learning ANFN controller has a significant reduction in the oscillation
rate Į will be considered according to the following theorem. with a 5% noise than the PID controller with practically no
loss of accuracy.
Theorem [24]: Let Į be the learning rate for the parameters
of fuzzy controller and g max be defined as g max := max k g (k ) VII. CONCLUDING REMARKS
where g (k ) = δy (k ) / δz (k ) and ||.|| is the usual Euclidean norm in The paper presents the numerical simulation results of the
online adaptive NFN and PID controllers with the
ℜ n and let S = δy m / δu . Then the convergence is guaranteed if
mathematical modeling of the AUV. ANFN controller is found
Į is chosen as; to be quite effective for the coupled nonlinear, six degree of
ʹ freedom dynamic system control. It is seen that the proposed
Ͳ ൏‫ן‬൏ ʹ ʹ ሺͳͺሻ
ܵ ݃݉ܽ‫ݔ‬ control system does a better job compared to the conventional
PID in terms of accuracy as well as the speed. The PID
VI. SIMULATION RESULTS controller is seen to have higher transients as well as steady
This study demonstrated the performance of the ANFN state error. The results with varying hydrodynamic coefficients
control for nonlinear system such as AUV. This section and noise indicate similar trends with the ANFN Controller
simulates the AUV system behavior and compares the doing a better job than the conventional PID.
performance of the ANFN controller and conventional PID 5
controller with the mathematical model of the AUV.
Fig. 7 shows the simulation result of the motion of the 2.5
AUV in square trajectory in XY plane, which means the yaw
angle control when ANFN and PID are applied to the AUV
Y-Direction

mathematical model. It is obviously seen that DE has a good 0

job to tune the controller’s parameter from initial values and it


makes the job easier for the online tuning to get the controller -2.5
control successfully the behavior of the vehicle. Desired Path
PID Controller

A. Closed loop accuracy evaluation of the controllers -5


ANFN Controller

0 200 400 600 800


The most widely used method for measuring performance X-Direction

and the accuracy indicators of the AUV system is the root Figure 7. The AUV sqeuar motion in XY plane with ANFN and PID.
mean square error (RMSE) and is defined as:
ଵൗ 50
௡ ଶ PID Controller
ͳ
ܴ‫ ܧܵܯ‬ൌ  ൭ ෍ ݀௜ଶ ൱  ሺͳͻሻ 40
ANFN Controller
݊
௜ୀଵ
A n g le E r r o r , d e g / s e c

where n is the number of data pairs and di is the difference 30


between ith desired and measured values The RMSE is said to
20
provide information on the short-term performance of a model
by allowing a term-by-term comparison of the actual 10
difference between the desired value and the measured value.
Fig. 8 and Fig. 9 show the corresponding angle error 0

between the desired and the actual trajectory in pitch angle -10
and yaw angle, respectively. The RMSE values of the pitch 0 200 400
Time, sec
600 800

angle error of the system are 4.61 and 0.572 and 10.35 and 5.1 Figure 8. The error in pitch angle.

188 2012 First International Conference on Innovative Engineering Systems


100 Technologies for Small Fixed Wing UAVs,” AIAA Journal of
Aerospace Computing, Information, and Communication, vol. 2, no. 1,
pp. 92-108, 2005.
50 [6] UAV MarketSpace Inc., “Autopilot Navigation,” http://www.uavm.com/
uavsub system/ autopilotnavigation.html.
A n g le E r r o r , d e g /s e c

[7] Chao, H., Y. Cao and Y. Chen, “Autopilots for Small Fixed Wing
0 Unmanned Air Vehicles: a Survey,” Proceeding of IEEE Conference on
Mechatronics and Automation, pp. 3144-3149, 2007.
[8] Jang, J. S. and C. J. Tomlin, “Autopilot Design for the Stanford
-50 DragonFly UAV - Validation through Hardware in the Loop
PID Controller
Simulation,” Proceeding of AIAA Guidance, Navigation, and Control
ANFN Controller Conference and Exhibit, Montreal, Canada, 2001.
-100
0 200 400 600 800 [9] C. H. Chen, c. J. Lin, and c. T. Lin, “nonlinear system control using
Time, sec adaptive neural fuzzy networks based on a modified differential
Figure 9. The error in yaw angle. evolution” IEEE Transactions on systems, man, and cybernetics—part c:
applications and reviews, vol. 39, no. 4, pp. 459-473, July 2009.
5 [10] J.C. Patra, G. Panda, Adriaan van den Bos, Modeling of an intelligent
pressure sensor using functional link artificial neural networks, ISA
Transactions 39 (2000) 15-27.
2.5
[11] Hassanein, O.; Anavatti, S.G.; Ray, T.; , " Genetic Fuzzy Controller for
Robot Manipulator Position Control based upon Inverse Dynamic,"
2011 IEEE Congress on Evolutionary Computation (CEC 2011), June
Y - Direction

0 2011
[12] O. Cordon, F. Herrera, F. Hoffmann, L. Magdalena, Genetic Fuzzy
Systems evolutionary Tuning and Learning of Fuzzy Knowledge Bases ,
-2.5
World Scientific, Singapore, 2001.
Dsired Path
ANFN Controller [13] S. Das, P.N. Suganthan, Differential Evolution: A Survey of the State-
-5
PID Controller
of-the-Art, IEEE Transactions on Evolutionary Computation, 15 (2011)
0 200 400
X - Direction
600 800
4-31.
[14] Wenyin Gong, et al, A Generalized Hybrid Generation Scheme Of
Figure 10. The AUV sqeuar motion in XY plan with with parameter variation. Differential Evolution For Global Numerical Optimization. Int. J. Comp.
Intel. Appl., (2011), 10, 35.
5
[15] Fossen, Thor I.: Guidance and control of ocean vehicles. Wiley , New
York (1994)
2.5
[16] Hassanein. O, S. Anavatti, T. Ray “Fuzzy Modeling and Control for
Autonomous Underwater Vehicle”, the Proceedings of the 5th
International Conference on Automation, Robotics and Applications
Y - D irection

0
(ICARA 2011), 169-174.
[17] Koivo, H. N., and J. T. Tanttu, “Tuning of PID Controllers : survey of
SISO and MIMO techniques”, IFAC Intelligent Tuning and Adaptive
-2.5 Control, Singapore, pp. 75-80, (1991)
Dsired Path [18] Ziegler, J. G., and Nicholas, N. B., “Optimum Setting for Automatic
PID Controller
ANFN Controller
Controllers” Trans. of ASME Vol. 65, pp 759-768, (1942)
-5
0 200 400 600 800 [19] C. H. Chen, C. J. Lin, and C. T. Lin, “A functional-link-based neuro-
X - Direction
fuzzy network for nonlinear system control,” IEEE Trans. Fuzzy Syst.,
Figure 11. The AUV sqeuar motion in XY plan with noise and sea currents. vol. 16, no. 5, pp. 1362–1378, Oct. 2008.
[20] Y. H. Pao, S. M. Phillips, and D. J. Sobajic, “Neural-net computing and
REFERENCES intelligent control systems,” Int. J. Control, vol. 56, no. 2, pp. 263–289,
1992.
[1] Smallwood, David A., and L. Louis Whitcomb. “Model-Based Dynamic
[21] R. Storn, K. Price, Differential Evolution - A simple and efficient
Positioning of Underwater Robotic Vehicles: Theory and Experiment.”
adaptive scheme for global optimization over continuous spaces, in:
IEEE JOURNAL OF OCEANIC ENGINEERING VOL. 29 (Jan. 2004):
Technical Report, International Computer Science Institute 1995.
169-186
[22] S.M. Elsayed, R.A. Sarker, D.L. Essam, A three-strategy based
[2] Holtzhausen, Servaas. Design of an AutonomousDesign of an
differential evolution algorithm for constrained optimization, in: the 17th
Autonomous Underwater Vehicle: Vehicle Tracking and Position
international conference on Neural information processing: theory and
Control. Master Thesis, KwaZulu-Natal: University of KwaZulu-Natal,
algorithms - Volume Part I, Springer-Verlag, Sydney, Australia, 2010,
2010.
pp. 585-592.
[3] Xia, Guoqing, Xiaocheng Shi, Mingyu Fu, Hongjian Wang, and Xinqian
[23] S. Das, P.N. Suganthan, Differential Evolution: A Survey of the State-
Bian. "Design of dynamic positioning systems using hybrid CMAC-
of-the-Art, IEEE Transactions on Evolutionary Computation, 15 (2011)
based PID controller for a ship." Mechatronics and Automation, 2005
4-31.
IEEE International Conference. Canada: IEEE, 29 July-1 Aug, 2005.
825 - 830. [24] Chen, Y.C. and C.C. Teng.: A Model Reference Control Structure Using
a Fuzzy Neural Network. J. Fuzzy Sets and Systems, vol.73, pp. 291-312
[4] Jang, J. S. and D. Liccardo, “Automation of Small UAVs Using a Low
(1995).
Cost MEMS Sensor and Embedded Computing Platform,” Proceeding of
25th IEEE/AIAA Conference on Digital Avionics Systems, pp. 1–9,
2006.
[5] Beard, R., D. Kingston, M. Quigley, D. Snyder, R. Christiansen, W.
Johnson, T. Mclain and M. Goodrich, “Autonomous Vehicle

2012 First International Conference on Innovative Engineering Systems 189

You might also like