You are on page 1of 18

Original Article

Multi-input submarine control via L1


adaptive feedback despite
uncertainties

Proc IMechE Part I:


J Systems and Control Engineering
2014, Vol. 228(5) 330347
IMechE 2014
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0959651813515206
pii.sagepub.com

Keum W Lee1 and Sahjendra N Singh2

Abstract
The development of a novel adaptive autopilot for the dive-plane control of multi-input multi-output submarines with
unmodeled dynamics, based on the L1 adaptive control theory, is the subject of this article. An L1 adaptive autopilot is
designed for the trajectory control of the depth and pitch angle using bow and stern hydroplanes. Interestingly, the
structure of the adaptive controller remains fixed, regardless of the nonlinearities and external disturbance inputs,
retained in the model of the submarine. Unlike the traditional adaptive control laws, the L1 adaptive control input is generated by filtering the estimated control signal. A nice feature of the control law is that it is possible to achieve fast adaptation and desirable performance bounds in the closed-loop system by the choice of large adaptation gains. Simulation
results are presented, which show that the autopilot accomplishes precise trajectory control in the dive plane, despite
parametric uncertainties, unmodeled nonlinearities, and random disturbance inputs.

Keywords
L1 adaptive submarine control, autonomous underwater vehicle dive-plane control, nonlinear submarine control, uncertain autonomous underwater vehicle control, multi-input control

Date received: 13 August 2013; accepted: 8 November 2013

Introduction
The design of control systems for high-performance
autonomous underwater vehicles (AUVs), which are
capable of executing precise maneuvers, is an important
problem from the theoretical as well as practical point
of view. But the design of autopilots for AUVs is indeed
challenging. The complexity in the design is caused by
the nonlinear dynamics of AUVs. Also, the control system must attenuate the effect of disturbance forces and
moments due to sea waves.
Researchers have made significant progress in
exploring the nonlinear dynamical behavior of submersibles and designing feedback control systems for performing maneuvers. Based on modern control
techniques, linear and nonlinear control systems have
been proposed for AUVs.1 The design of autopilot for
lateral control has been presented based on H# control
method. An application of the multi-input multi-output
(MIMO) H control design technique for the robust
control of an AUV that suppresses pitch and yaw coupling has been provided.3 A suboptimal control law for
the dive-plane control of an AUV has been derived
using the state-dependent Riccati equation method.4

A robust and decoupled cascaded control system for


underwater vehicles has been designed.5 Based on the
Lyapunov method and nonlinear optimization
approach, controllers for an underactuated underwater
vehicle have been developed.6 For uncertain models of
AUVs, variable structure control (VSC) systems have
been designed.79 A VSC system for remotely operated
vehicles has been designed, which avoids actuator chattering.10 A simplified fuzzy logic controller (termed single input fuzzy controller) for a deep submerged rescue
vehicle has been proposed.11 Based on fluid compensation approach, a controller that compensates for the
hydrodynamic loads has been synthesized.12

Division of Electronic Information and Communication, University of


Kwandong, Gangneung, South Korea
2
Department of Electrical and Computer Engineering, University of
Nevada, Las Vegas, Las Vegas, NV, USA
Corresponding author:
Sahjendra N Singh, Department of Electrical and Computer Engineering,
University of Nevada, Las Vegas, 4505 Maryland Pky, Las Vegas, NV
89154-4026, USA.
Email: sahjendra.singh@unlv.edu

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

331

Researchers have developed a variety of adaptive


control systems for the control of AUVs. Do et al.13
have designed an adaptive path following control system for underactuated AUVs. A nonlinear adaptive law
for the control of AUVs in the dive plane has been proposed.14 Based on the Lyapunov theory and backstepping techniques, a robust adaptive dive-plane control
system has been designed.15 The depth control of an
underwater vehicle based on inputoutput linearization,
sliding mode, and adaptive fuzzy control techniques has
been considered.16 A neural network-based adaptive
law for AUVs has been proposed.17
An adaptive sliding mode controller based on recurrent neural networks for an underwater robot has been
proposed.18 The authors have developed an adaptive
output feedback control system for the control of an
AUV using dynamic recurrent fuzzy neural network.19
In general, adaptive laws using neural networks must
adapt a large number of parameters for the nonlinear
function approximation, and therefore, the controllers
have complicated structure. A self-adaptive recurrent
neuro-fuzzy controller for an AUV in unstructured
environment has been designed.20 An adaptive control
law for AUVs using quaternions to represent attitude
errors has been developed.21 An indirect adaptive control method for an AUV based on an extended Kalman
filter has been proposed.22 Control systems for biorobotic autonomous underwater vehicles (BAUVs) using
pectoral fins2325 have also been designed. These adaptive sampled-data pectoral fin control systems are for
yaw-plane as well as dive-plane control of BAUVs.
Control of submarines, equipped with bow and stern
hydroplanes, based on the classical (Bode diagrams and
root locus) techniques has been considered.26 A multimodel gain scheduled optimal controller for this submarine, under the assumption that its parameters are
known, has been presented.27 Later, a reconfigurable
sliding mode controller for this model was designed.28
For this submarine model, autopilots based on adaptive
backstepping control technique and the noncertaintyequivalence principle have also been designed.29,30 In
these articles, the decomposition of the input matrix as
a product of a symmetric positive definite matrix, a
diagonal matrix, and an upper triangular matrix has
been used for avoiding singularity in the control law.
The design algorithm of Lee and Singh30 is essentially
based on the modified version of the immersion and
invariance approach and uses filtered signals for
synthesis.
Specific assignments of submarines require precise
control of the depth and pitch angle in the dive plane.
The key technical issues for the control of an MIMO
submarine are to deal with the dynamic interactions
between the control inputs produced by the hydroplanes and the outputs (depth and pitch angle), uncertainties in its dynamics, and perturbing forces due to
the sea waves. Therefore, the design of a robust control
system for a fine control of the depth trajectory, with
small excursion in the pitch angle around 0, is of

considerable importance. Apparently, the hydrofoils


must generate only a net vertical depressing lift while
maneuvering if the pitch angle is restricted to 0 value.
It is pointed out that this kind of operation has vital
importance for conventional submarines to use their
periscope for searching targets and threats on the surface of water or air and charge batteries while cruising
despite external disturbances. Also, large pitch angle
deviations of vehicles cruising below the mean sea level
are undesirable because they may induce the propellers
or the bow to rise over the water. Furthermore, a precise control over vertical movement of underwater vehicles, when it is moving in a close proximity of a terrain
covered with soft sediment, is important. Recently, for
the control of uncertain nonautonomous nonlinear systems, a robust L1 adaptive control theory has been
developed by Hovakimyan and Cao31 and Cao and
Hovakimyan.32 This design approach has been successfully used for the control of NASA AirSTAR vehicle,33
aircraft, missiles, and other dynamical systems.
(Readers may refer to Hovakimyan and Cao31 for
additional references.) Adaptive schemes proposed in
Nambisan and Singh29 and Lee and Singh30 are based
on the linear parameterization of uncertain functions.
It is important to note that the L1 adaptive theory is
applicable to systems with unmodeled dynamics.
Because the modeling of the hydrodynamic nonlinearities of underwater vehicles is not easy and the forces
induced by the sea waves are unpredictable, the design
of an L1 adaptive autopilot for the multi-input AUVs
is of considerable interest.
The contribution of this article lies in the design of a
new adaptive autopilot for the dive-plane control of
MIMO submarines based on the L1 adaptive control
theory. It is assumed that the uncertain dynamical
model of the submarine includes unknown parameters
and unmodeled time-varying nonlinear functions. For
the purpose of design, the submarine model of Dumlu
and Istefanopulos27 and Demirci and Kerestecioglu,28
including uncertain nonlinear time-varying functions, is
considered. This submarine is equipped with a bow and
a stern hydroplane for the depth and pitch angle control. A nonlinear L1 adaptive control law is designed
for the trajectory control of the depth and pitch angle,
based on an equivalent time-varying nonlinear state
variable representation of the submarine. For the derivation of the adaptation law, a state predictor is constructed. For limiting the parameter estimates within
their expected bounds, a smooth projection algorithm
is used in the parameter update law. Simulation results
are obtained, which show that the designed autopilot
accomplishes precise dive-plane maneuvers, despite
large uncertainties in the model parameters, unmodeled
nonlinear hydrodynamic force and moment, and persistently acting random disturbance force and moment.
The control law derived here differs from the traditional adaptive laws in important ways. First of all, L1
adaptive system is applicable to nonlinear time-varying
systems with unmodeled dynamics (unstructured

Downloaded from pii.sagepub.com by guest on January 20, 2015

332

Proc IMechE Part I: J Systems and Control Engineering 228(5)

dynamics). Unlike this, the adaptive systems of


Nambisan and Singh29 and Lee and Singh30 assume
that the uncertain hydrodynamic nonlinearities are linearly parameterized. Interestingly, the structure of the
L1 adaptive controller remains fixed, regardless of the
nonlinearities and external disturbance inputs, retained
in the model of the submarine. For the synthesis of the
control law, only two unknown parameters and four
unknown functions are required to be estimated. This
is important because the hydrodynamic force and
moment are complex functions of the state variables,
and their representation in a precise analytical form is
difficult. It is noted that unlike the controller of this
article, the number of unknown parameters in the linearly parameterized models for design, used in
Nambisan and Singh29 and Lee and Singh,30 increases
with the nonlinear hydrodynamic terms retained in the
model, and their controller requires higher order adaptation law. Of course, the structure of their control systems changes with the form of nonlinearity in the
model, and the computational burden increases as the
number of estimated parameters increases.
Unlike the traditional adaptive schemes for submarines, here the actuating control input is obtained by filtering the estimated control signal through a low-pass
filter. A nice feature of the L1 adaptive autopilot
designed here is that it is possible to obtain fast adaptation, desirable performance bounds and guaranteed
robustness in the closed-loop system by the choice of
large adaptation gains. The choice of large adaptation
gain cannot cause instability in the system because the
high-frequency components, generated by the fast
adaptation law, are attenuated by the low-pass filter,
and only the control signals with low frequencies are
applied to the submarine for the purpose of maneuver.
Of course, the cutoff frequency of the low-pass filter
can be adjusted to an appropriate value for achieving
desirable transient performance by the choice of its
parameter, and the adaptation gain can be chosen sufficiently large for fast adaptation. The published traditional adaptive laws (without low-pass filter in the
loop) for AUVs do not have this property. In fact, it
has been shown that very large adaptation gains reduce
the stability margins of the traditional adaptive
systems.31
The organization of the article is as follows: Section
Submarine model presents the mathematical model
and the control problem. The L1 adaptive control system is derived in section L1 adaptive autopilot
design. Section Simulation results presents the simulation results, and conclusions are provided in section
Conclusion.

Submarine model
For this study, the dive-plane nonlinear time-varying
dynamics of submarines (as shown in Figure 1) are considered. This model is similar to the neutrally buoyant

Figure 1. Submersible with bow and stern hydroplanes.

model of the submarine investigated earlier,2628 but it


includes hydrodynamic nonlinearities as well for a practical representation. Of course, the design approach
considered here is applicable to submarines that are not
necessarily neutrally buoyant. The equations of motion,
describing the depth variable along the body-fixed zaxis and the pitch angle along the y-axis, are given by
Z9 L
Z9w U
1
_
_ + Q_ Q(t)
w(t) +
(Z9u_ + m9)Uu(t)
Lm93
m93
m93
Z9dB U2
Z9dS U2
Zd (t)
dB(t) +
dS(t) +
+
0:5rL3 m93
m93 L
m93 L
+ Zn (w, q)
1

_ =
w(t)

and
M9u_ U _
M9w U
_ = M9w_ w(t)
_ + 2
u(t)
w(t) +
Q(t)
L I92
LI92
LI92
M9dB U2
M9dS U2
+
dB(t)
+
dS(t)
L2 I92
L2 I92
2mg(zG  zB )
Md (t)
+
u(t) +
+ Mn (w, q)
rL5 I92
0:5rL5 I92
2

where w is the velocity along the z-axis, h is the depth of


the vehicle, u is the pitch angle, Q = u_ is the pitch rate,
and dB and dS are the hydroplane deflections in the
bow and stern planes, respectively. The time-varying
functions Zd (t) and Md (t) denote bounded disturbance
inputs due to the sea waves. The effects of nonlinear
hydrodynamic forces and moments are included in the
nonlinear functions Zn (w, Q) and Mn (w, Q). For example, these include the vehicle cross-flow drag, which is a
function of second-order terms wjwj and QjQj, and
other higher order functions of w and Q. In fact, it will
be seen later that for the derivation of the control law,
the analytical expressions of the nonlinear functions
Zn (w, Q) and Mn (w, Q) are not essential, and as such
these will be treated as unmodeled functions. (For notations, see Babaoglu26 and Dumlu and Istefanopulos.27)
In this model, we assume that the forward velocity
U = U0 is a constant.
Solving equations (1) and (2) gives the differential
equations for w and Q of the form

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh


  

w_
fw (w, u, Q, t)
...
................ + B0 uc
=
fu (w, u, Q, t)
Q_

333

and control inputs can be completed following the derivation of Hovakimyan and Cao.31)
In view of the interest in the depth trajectory control,
where uc = (dB, dS)T 2 R2 is the control input vector, it will be convenient to consider the design problem in
and the state space of variables h, u, h,
and
the
nonlinear
functions
(fw , fu )
_ and Q,
_ instead of w,
B0 = BT01 , BT02 T 2 R2 3 2 are given by
_ Differentiating equation (5) gives
_ and Q.
u, w,
2
3


Z9w U
_ + Zd (t)
w(t) + m91 3 (Z9u_ + m9)Uu(t)
3 m9 + Zn (w, q)
fw (w, u, Q, t)
Lm9
0:5rL
3
3
5
= M1 4
M9u_ U _
2mg(zG zB )
Md (t)
M9w U
fu (w, u, Q, t)
u(t)
+
w(t)
+
u(t)
+
+
M
(w,
q)
5
5
2
n
LI92
rL I92
0:5rL I92
L I92
2
3
"
#
Z9dS U2
Z9dB U2
1
ZQ9
_ L=m93
m9 L
m93 L 5
M=
4
; B0 = M1 4 3 2
2
1
M9
M9
U
dS U
dB
Mw9
1
_ (LI92 )
2
2
L I9
L I9
3

(Often the arguments of functions are suppressed.)


The superscript T denotes matrix transposition. For
simplicity in notation, the dependence of the functions
fw and fu on the disturbance inputs Zd (t) and Md (t) has
been denoted by the argument t.
Resolving the normal velocity (w) and the longitudinal velocity (Ut U0 ) along the vertical reference
axis, one obtains the nonlinear differential equation for
the depth h of the form
h_ = w cos u  U0 sin u

It is assumed that the matrix B0 is not precisely


known. Moreover, the nonlinear time-varying functions
fw (w, u, Q, t) and fu (w, u, Q, t) are treated as unmodeled
(unstructured) functions for the purpose of design.
For the control of the submarine in the dive plane,
the controlled output vector y associated with equation
(3) is selected as
y = (h, u)T

Consider a reference output vector


yr (t) = (hr (t), ur (t))T 2 R2

where hr (t) and ur (t) are certain smooth depth and pitch
angle trajectories. The objective here is to design an L1
adaptive control law so that the submarine follows the
given reference depth command hr (t) and attains the
specified depth, and furthermore, the deviation in the
pitch angle remains small around the commanded null
trajectory ur (t) = 0 for all t 5 0 despite parametric
uncertainties, unmodeled nonlinearities, and random
disturbance inputs.

adaptive autopilot design

In this section, an adaptive control system is designed


for the trajectory control of the depth and pitch angle
trajectories. The derivation of the control law is based
on the L1 adaptive control theory for uncertain nonlinear time-varying systems, developed in Hovakimyan
and Cao.31 In this section, the control law derivation is
outlined briefly. (The proof of stability and the computation of performance bounds on the system trajectories

h = w_ cos u  wQ sin u  U0 Q cos u

Solving for w from equation (5), one has


w = ( cos u)1 (h_ + U0 sin u)

Then using equations (3), (8), and (9), one obtains h


_ and Q of the form
and Q_ as functions of only h, u, h,
  

 
_ Q, t)
B01 cos u
h
gh (u, h,
10
=
uc
+
_ Q, t)
B02
Q_
gu (u, h,
where
_ Q, t) = fu (w, u, Q, t)
gu (u, h,
_ Q, t) = fw (w, u, Q, t) cos u  wQ sin u
gh (u, h,

11

 U0 Q cos u

Note that w from equation (9) is substituted in equation (11) to obtain gh and gu in equation (10).
_ Q)T 2 R4 . Then
Define a state vector as x = (h, u, h,
a state variable representation of the AUV obtained
from equation (10) is given by
2
3
2
3
h_
02 3 2
6
7
Q
7 4
5
12
x_ = 6
4 gh (u, h,
_ Q, t) 5 + B01 cos u uc
B
02
_ Q, t)
gu (u, h,
where the null matrix 02 3 2 has two rows and two columns. The controlled output vector is
y = I2 3 2 , 02 3 2 x = C0 x

13

where the identity matrix I2 3 2 has two rows and two


columns. In view of the singularity at u = 6p=2 in
equation (9), the region of interest in the state space is
chosen such that u does not exceed rx \ p=2, where rx
is a positive number. (Of course, it will be seen later
that in the closed-loop system, the pitch angle remains
small.) Also, it is assumed that the initial condition
x(0) = x0 satisfies jjx0 jj 4 r0 \ rx , where r0 is a
constant.
To this end, for the derivation of the control law, a
simplification in the product term B01 cos u is made in
equation (12). Because u will be regulated so that it
remains small during the maneuver, the coefficient

Downloaded from pii.sagepub.com by guest on January 20, 2015

334

Proc IMechE Part I: J Systems and Control Engineering 228(5)

B01 cos u of uc is replaced by B01 . With this simplification, equation (12) can be written as




02 3 2 I 2 3 2
02 3 2
_ Q, t) + B0 uc )
x_ =
x+
(g(u, h,
02 3 2 02 3 2
I2 3 2
14

where the nonlinear vector function g 2 R2 is


_ Q, t) = gh (u, h,
_ Q, t), gu (u, h,
_ Q, t)T
g(u, h,

_ Q, t) is an unmodeled vector
It is assumed that g(u, h,
function. For the design of the controller, the uncertain
input matrix is assumed to be of the form
B0 = B0m V; V = diagfv11 , v22 g

15

where v11 . 0 and v22 . 0 are unknown constants and


the matrix B0m 2 R2 3 2 is a nominal matrix. Note that
the parameters v11 and v22 represent uncertainty in the
control signals produced by the bow hydroplane and
stern hydroplane, respectively. It may be pointed out
that the derivation presented here is also applicable for
submarine models in which the uncertainty matrix V is
not necessarily diagonal, but it is required to be only
strictly row dominant with positive diagonal elements.
Define a Hurwitz matrix Am of the form


02 3 2 I 2 3 2
Am =
16
Am1 Am2
where Am2 , Am1  is a 2 3 4 matrix, and the matrices Am1
and Am2 are
Am1 = diagfv21 ,  v22 g
Am2 = diagf2z1 v1 ,  2z2 v2 g

where zi . 0 and vi . 0 are design parameters. By adding and subtracting Am1 , Am2 x, equation (14) can be
expressed as
x_ =Am x+ Bm
h
i
T
_ T
_
Vuc + B1
0m fAm1 (h, u)  Am2 (h,Q) +g(u, h, Q, t)g
:
Am x+ Bm Vuc +f(x,t)
17

where Bm = 02 3 2 , BT0m T 2 R4 3 2 and the nonlinear


vector function is
T
T
_
_
f(x, t) = B1
0m fAm1 (h, u)  Am2 (h, Q) + g(u, h, Q, t)g

18

Because the disturbance inputs Zd (t) and Md (t) due


to the sea waves are bounded, one has jjf(0, t)jj 4 b0
for all t 5 0, for some positive number b0 . For the
design of the controller, the vector function f(x, t) is
assumed to be unmodeled.
The control system design begins with an equivalent
representation of the time-varying nonlinear system
equation (17). For continuous control input uc (t), and
bounded trajectory x(t) over an interval 0, t, it has
been shown in Hovakimyan and Cao31 that there exist
continuous bounded vector functions m(t) = (m1 (t),
m2 (t))T 2 R2 and s(t) = (s1 (t), s2 (t))T 2 R2 , with piecewise continuous derivatives such that the system equation (17) is equivalent to the semi-linear time-varying
system given by
x_ = Am x + Bm Vuc + m(t)jjxjj + s(t)

19

It is interesting to note that unlike approximation of


nonlinear functions by neural networks using basis
functions, here the function f(x, t) has exact representation given by
f(x, t) = m(t)jjxjj + s(t)

For good approximation of nonlinear functions,


neural networks must use a large number of constant
weighting parameters; therefore, a large order adaptive
law is required for the purpose of control. Of course,
the vector functions m(t) and s(t) are not known
because f(x, t) is unknown. Note that V has been
assumed to be unknown as well.
The L1 adaptive control system includes a state predictor, an adaptation law, and a control law. The block
diagram of the closed-loop system is shown in Figure 2.
The state predictor is used for the derivation of the
adaptation law based on the Lyapunov analysis. Now
the design of each component of the control system is

Figure 2. A block diagram representation of the closed-loop system including the L1 adaptive autopilot.

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

335

presented. First, in view of the equivalent system equation (19), consider a state predictor of the form
^ c +m
^ (t)jjx(t)jj + s
^ (t), x^(0) = x(0)
x^_ = Am x^ + Bm Vu
20

where x^ 2 R4 is the state vector of the predictor and


^
(V(t),
m
^ (t), s
^ (t)) is the estimate of (V, m(t), s(t)). This
state predictor is, in fact, a copy of the system equation
(19) in which an estimate of the unknown forcing function Vuc + m(t)jjxjj + s(t) has been used. Its purpose
is not to obtain the estimate of the state vector x like a
Kalman filter because the state vector is assumed to be
available for the synthesis of the control law. The predictor also uses x for its implementation, as seen in
equation (20). Define the state prediction error as
x~ = x^  x. Then subtracting equation (19) from equation (20) gives the state prediction dynamics of the
form
~ c +m
~ (t)jjxjj + s
~ (t)
x~_ = Am x~ + Bm Vu

21

where V~ = V^  V, m
~ =m
^  m, and s
~ =s
^  s are the
parameter and function errors.
For the derivation of the adaptation law, consider a
quadratic positive definite Lyapunov function
~ +m
~ m
~ Tm
W(~
x, V,
~, s
~ ) = x~T P~
x + G1 tr(V~T V)
~ +s
~Ts
~
22

where tr denotes the trace of a matrix, the adaptation


gain G . 0 and the positive definite symmetric matrix P
(denoted as P . 0) is the unique solution of the
Lyapunov equation
PAm + ATm P =  Q

23

for some Q . 0. Differentiating W along the solution of


equation (21) and using equation (23) gives
_ =  x~T Q~
~ c +m
W
x + 2~
xT PBm fVu
~ jjxjj + s
~g
_
1
T
T
T
^ +m
+ 2G tr(V~ V)
~ m
^_ + + s
~ s
^_ 

24

1

 2G ~
mT m_ + s
~ T s
_

In view of equation (24), one chooses the adaptation


law of the form
xjjxjj ), m
^ (0) = m
^0
m
^_ = GProj(^
m,  BTm P~
T
x), s
^ (0) = s
^0
s
^_ = GProj(^
s,  Bm P~
_ = GProj(V,
^  BT P~
^ = V^0
xuT ), V(0)
V^
i
m

^ c +m
uc =  Kf D(s)V(t)u
^ (t)jjx(t)jj
+s
^  Kg yr (t), uc (0) = 0

27

where Kf and Kg are 2 3 2 matrices and D(s) is a 2 3 2


strictly proper transfer matrix such that the strictly
proper matrix C(s) defined as
C(s) = VKf D(s)(I2 3 2 + VKf D(s))1 , C(0) = I

28

1
is stable and K1
g =  C0 Am Bm . Here, s denotes the
Laplace variable or a differential operator. Suppose the
initial conditions satisfy jjx0 jj 4 r0 . Then there exists
rin . r0 such that

jjeAm t jjL1 r0 4 rin

because Am is a Hurwitz matrix. (L1 denotes the function 1 norm.)


For stability in the closed-loop system, one needs to
ensure that for a given set of initial conditions
jjx0 jj 4 r0 , there exists rr . rin such that the following
norm condition holds31
jjGm (s)jjL1 \

rr  jjHm (s)C(s)Kg jjL1 jjyr jjL  rin


L1rr rr + b0

29

where Hm (s)= (sI  Am )1 Bm , Gm (s)= Hm (s)(I  C(s)),


rr (rr )), rr =
jjf(0, t)jj 4 b0 , and Lrr = rr (rr )r1
r dfx (
1 with g
1 . 0, an arbitrarily small number and
rr + g
dfx 5 jjf(x, t)=xjj . (L denotes the function norm.)
For examining the performance characteristics of the
adaptive system, consider a reference closed-loop system for the submarine of the form
x_ ref = Am xref + Bm (Vucref + f(xref , t))
ucref =  V1 C(s)(f(xref , t)  Kg yr )
yref = C0 xref

30

25

where Proj denotes a smooth projection operator


(defined later) to limit each estimated parameter within
certain expected lower and upper bounds. In view of
equation (25), one finds that the choice of large adaptation gain G yields rapid changes in the estimated parameters (^
m, s
^ , V^i ), and thereby achieves fast adaptation.
Substituting equation (25) in equation (24) gives
_ 4  x~Q~
W
x + 2G1 j~
mT m_ + s
~ T sj
_

It is noted that because m and s are functions of


time, it is not possible to select the adaptation law to
completely eliminate the uncertain functions in the
derivative of W. However, in view of the boundedness
of m(t) and s(t) in equation (19), it can be established
that W(t) is bounded over 0, t and jj~
xt jjL is bounded
by a function, which is proportional to G1=2 .31 (The
function x~t denotes the truncated function over 0, t
obtained from x~(t).)
The control law is chosen as

26

The reference control law ucref is obtained by filtering


the feedback linearizing control input through the lowpass filter C(s). The adaptive law is essentially an adaptive version of ucref . Note that if C(s) = I2 3 2 , then the
system equation (30) becomes a linear stable system
given by
x_ ref = Am xref + Bm Kg yr

Of course, the DC gain of C(s) is C(0) = I2 3 2 . It has


been shown that for sufficiency large adaptation gain G,
subject to the L1 norm condition equation (29), all the

Downloaded from pii.sagepub.com by guest on January 20, 2015

336

Proc IMechE Part I: J Systems and Control Engineering 228(5)

signals in the closed-loop system are bounded and the


trajectory and control input errors satisfy31

where
c
4 p  p
(p) =
p
e pm

jjxref (t)  x(t)jjL 4 g1


jjucref (t)  x(t)jjL 4 g 2

31

jjycref (t)  y(t)jjL 4 jjC(s)jj g1

where g 1 and g 2 are computable bounds. (For a complete derivation, readers may refer to Hovakimyan and
Cao.31) Interestingly, it has been established that the
tracking error between y(t) and yref (t), and uc (t) and
ucref (t), is uniformly
p bounded by a constant inversely
proportional to G. Therefore, it is possible to achieve
desirable tracking transient and steady-state performance by the choice of large adaptation gains. It is
pointed out that such kind of performance bounds cannot be obtained using traditional adaptive laws.
For the AUV model, the matrix D(s) is selected as
D = s1 I2 3 2 and the feedback matrix is set to
Kf = kf I2 3 2 for simplicity, where kf is a positive real
number. Then the control law equation (27) simplifies to
^ c+m
^ jjx(t)jj + s
^  Kg yr , uc (0) = 0 32
u_ c =  kf Vu
This completes the design.

Simulation results
This section presents the simulation results. The complete closed-loop system including the submarine model
equations (1) and (2), adaptation law (equation (25)),
and the control law (equation (32)) is simulated. The
parameters of the model given in Dumlu and
Istefanopulos27 are used for simulation. (These are collected in Appendix 1 for convenience.) Of course, here
results for the model including nonlinear terms will also
be obtained. The initial state of the vehicle is x(0) = 0.
Numerical results for various cases for the model
including random disturbances, unmodeled nonlinearities, and parameter uncertainties are obtained. A
smooth projection algorithm is used for updating the
^ i , and s
^ i , (i = 1, 2). Suppose a paraparameters v^ii , m
meter p is to be confined to a set P defined by
P = fp : jp  pj \ pm g

where p and pm are given constants. Define a function c


of p as
"
#

2 p  p 2
c(p) =
 1+e
e
pm
where e is a constant satisfying 0 \ e \ 1. Then a
smooth projection (termed Proj) is given by34
Proj(p, f) =
8
f
>
>
>
>
f
>
>
>
<
c
(p)f
c(p)
>
p
>
>
f 
(p)
2 c
>
>
 p

>
c
>
 (p)
:
 p 

if c(p) 4 0
if c(p) 5 0 and

c
p (p)f 4 0

The input matrix B0 is not known. For the purpose


of simulation, it is assumed that the nominal matrix
B0m for the controller design is B0m = au B0 . Therefore,
one has au = v1
ii , i = 1, 2. Results are presented for
several values of au including au = 1:6 and au = 0:4.
Note that for the choice of the uncertainty factor
au = 1:6, the chosen matrix B0m for design is + 60%
higher, but for au = 0:4, it is 60% smaller than the
actual value B0 . Of course, vii , i = 1, 2 are not known.
For the purpose of illustration, the initial estimate of
the unknown diagonal matrix V(t) is arbitrarily set as
^ = diagf104 , 104 g, which is extremely small. As
V(0)
such, it is seen that there is a large uncertainty in the
coefficient matrix of uc in the predictor equation (20).
The initial values of the vector functions m
^ and s
^ are
^ = 0, 0T .
selected as m
^ = 0, 0T and s
The matrix Q is 500I4 3 4 . The adaptation gain G is
1000 and the gain kf is 5. The value of e in the projection algorithms is 0.0001. The chosen parameter bounding sets for the projection algorithm are such that
v^ii 2 104 , 20, m
^ i 2 20, 20, and s
^ i 2 20, 20; one
has zi = 1 and vi = 0:1, i = 1, 2. A smooth reference
depth trajectory is generated by a fourth-order command generator of the form
(s2 + 2zc1 vc2 s + v2c1 )(s2 + 2zc2 vc2 s + v2c2 )hr (t)
= v2c1 v2c2 hr

33

The reference depth trajectory generator is used here


to provide flexibility in shaping the transient response
of the depth trajectory (h). The parameters (zci , vci ) in
equation (33) can be chosen to obtain desirable reference trajectory for tracking. The input (hr ) of the command generator is chosen to produce reference
trajectory converging to the desired target depth. The
initial values of hr (t) and its derivatives at t = 0 are 0,
and one has zci = 1 and vci = 0:1. The several values of
hr are considered for simulation. Since it is desired to
keep u close to 0, the reference pitch angle trajectory ur
is assumed to be 0. Simulation results are obtained by
introducing saturation so as to limit the control magnitude within 30.
It is pointed out that the norm inequality equation
(29) only provides a sufficient condition for stability in
the closed-loop system. As such for obtaining satisfactory transient performance, here the feedback gain kf ;
the adaptation gain G; the parameters vi and zi of the
matrix Ami ; the projection bounds for vii , mi , and si ;
and the command generator parameters vci and zci
have been selected after observing the simulated closedloop responses.

if not

1.

Adaptive control: hr = 10, uncertainty factor


au = 1:6, disturbance Zd = 0 and Md = 0

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

337

Figure 3. Adaptive control: hr = 10 ft, uncertainty factor au = 1:6, disturbance Zd = 0 and Md = 0: (a) depth h; (b) heave velocity w;
(c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
^1, m
^ 2 ; and
(h) parameter estimate s
^1 , s
^2.

It is desired to steer the submarine to a depth of 10


ft. It is assumed that the disturbance input is Zd = 0
and Md = 0, and the uncertainty factor is au = 1:6 (i.e.
B0m = 1:6B0 ). In this case, one has v11 = v22 = (1=1:6)
and G = 1000. Note the chosen initial values
v^11 (0) = 104 and v^22 (0) = 104 are quite small compared to the actual value of 1/1.6. The selected
responses are shown in Figure 3. It is observed that the
submarine smoothly attains the specified depth in
about 130 s. (Actually, the settling time Ts required for
the tracking error to be within 2% of the target value is
120.3 s.) The pitch angle and the normal velocity w
converge to 0. The maximum value of the pitch angle
(denoted as um ) is 0.003, which is insignificant. The
maximum control magnitude of uc (denoted as ucm ) is

(9:62, 1:25)T (8). It is seen that the parameter estimates


v^ii , m
^ i , and s
^ i converge to some constant values.
Simulation has been done for several values of the
adaptation gain. Table 1 shows the effect of G on the performance. It is observed that for smaller values of G, the
settling time is almost the same (about 120 s). (Ts denotes
the settling time when the depth tracking error is within
2% of the target value.) But for a large value G = 5000,
Ts is about 98 s. This is a significant reduction. With
larger G, the pitch angle deviation gets smaller, but it is
not decreasing monotonically. Of course, use of extremely
large G is not preferred because it causes numerical difficulty while integrating the stiff differential equation (25).
2. Adaptive control: hr = 10 ft, uncertainty factor
au = 0:4 and 1:6, random disturbance Zd and Md

Downloaded from pii.sagepub.com by guest on January 20, 2015

338

Proc IMechE Part I: J Systems and Control Engineering 228(5)

Table 1. Effect of adaptation gain G on performance: hr = 10 ft, au = 1:6, Zd = 0, Md = 0.


Adaptive gain G

Ts (s)

ucm ()

um

0.1
1
10
100
500
800
1000
5000

120.2224
120.3567
120.3032
120.3046
120.3007
120.2999
120.2985
98.3757

(30.0000, 30.0000)
(14.8440, 5.8910)
(9.6143, 1.2550)
(9.6133, 1.2527)
(9.6134, 1.2524)
(9.6133, 1.2523)
(9.6232, 1.2490)
(13.9122, 2.0753)

0.3406
0.0140
0.0028
0.0029
0.0029
0.0029
0.0029
0.0006

()

To examine the robustness of the control system,


simulation is done in the presence of random disturbance force Zd (t) and moment Md (t) for au = 0:4
(60% lower uncertainty) and au = 1:6 (60% higher
uncertainty), respectively. Distinct disturbance force Zd
and moment Md are generated using band-limited
white noise that is passed through a filter, whose transfer function is given by 1/(s + 5). The design parameters and initial conditions of Case 1 are retained.
The magnitudes of Zd and Md are of the order of 1 3
104 slug ft/s2 and 1 3 104 slug ft2/s2, respectively (see
Figure 4(i) and (j)). The waveforms of Zd (t) and Md (t)
(Figure 4) show random fluctuations and are reasonable for examining sensitivity of the designed controller
to external disturbances. It is noted that the control signal is filtered by the low-pass filter, but the disturbance
inputs Zd and Md act independently on the submarine.
Selected responses for 60% lower uncertainty are
shown in Figure 4. It is observed that similar to Case 1,
the desired depth is attained in about 130 s, and the
pitch angle remains very small (within 3 3 103 (8))
despite the presence of persistent random disturbance
inputs. The rapid fluctuations in the control inputs do
not represent control chattering. In fact, irregular
waveforms of control inputs are essential to cancel the
effect of the random disturbance force and moment for
precise control. The maximum control magnitude ucm
is (30, 29:9)T (8). Note for this case dB saturates.
Figure 5 shows the responses for 60% higher uncertainty (au = 1:6) in the presence of the random disturbance inputs, as shown in Figure 4. Again depth
control is accomplished in about 130 s. The maximum
control magnitude is uc = (21:7, 4:6)T (8). The peak
value of the pitch angle is small (less than 0.003).
Furthermore, unlike Figure 4, the control magnitude is
smaller and there is no control saturation in this case.
As such in the presence of disturbance input, it is preferable to design the controller using the overestimated
value of the nominal matrix B0m . The estimated parameters remain bounded.

nonlinearities on the performance of the L1 controller


is examined. The uncertainty factor is au = 1:6. The
disturbance inputs Zd and Md used for Case 2 are
retained for simulation. To examine the effect of unmodeled nonlinearities, nonlinear functions Zn =  1:2 3
and
Mn =  2:0 3
104 wjwj  1:6 3 104 QjQj
104 wjwj  3:0 3 104 QjQj are introduced in the differential equations for w and Q, respectively. But the
adaptive autopilot of Case 1 is used without any modification. Simulated responses are shown in Figure 6. It
is interesting to observe that in spite of the parameter
uncertainties, random disturbance, and unmodeled
nonlinearities, the depth trajectory control is accomplished smoothly in about 130 s. The normal velocity
remains close to 0 in the steady state. The pitch angle
remains within 0.001. Similar to Case 2, fluctuating
control signal is required for attenuating the effect of
the random disturbance inputs. The maximum control
magnitude is ucm = (21:1, 4:4)T (8).

3. Adaptive control: hr = 10 ft, uncertainty factor


au = 1:6, random Zd and Md , unmodeled nonlinearities (Zn , Mn )

Similar to Case 2, simulation is done for 30 ft depth


command for the uncertainty factor of au = 1:6 (60%
higher uncertainty), in the presence of the random disturbance Zd and Md given in Figure 4. Selected
responses are shown in Figure 8. It is observed that the
submarine reaches the target depth in about 130 s.

Now the combined effect of parameter uncertainties,


random
disturbance
inputs,
and
unmodeled

4. Adaptive control: hr = 30 ft, uncertainty factor


au = 1:6, disturbance Zd = 0 and Md = 0
It is desired to steer the submarine to a larger depth
(30 ft). The disturbance inputs are 0 and the uncertainty factor au is 1.6. The controller parameters used
for Case 1 are retained. Even for this larger depth command, smooth tracking of the depth trajectory hr is
observed. Simulated responses are shown in Figure 7.
Similar to Case 1, the response time is about 130 s. The
maximum value ucm of the control input vector is hr .
The pitch angle remains within 0.009. Compared to
Case 1 for 10 ft command, larger control magnitude is
observed in this case. Of course, this one would have
expected. The estimated parameters converge to some
constant values. The responses for w, Q, and u converge to 0.
5. Adaptive control: hr = 30 ft, uncertainty factor
au = 1:6, random disturbance Zd and Md

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

339

Figure 4. Adaptive control: hr = 10 ft, uncertainty factor au = 0:4, random disturbance Zd and Md : (a) depth h; (b) heave velocity w;
(c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
^1, m
^ 2;
^ 2 ; (i) disturbance Zd ; and (j) disturbance Md .
(h) parameter estimate s
^1 , s

Similar to Case 2, the control inputs remain fluctuating


over the whole period. We observe control saturation at
certain instants. The pitch angle remains within 0.012.
The normal velocity and u converge to 0. The maximum
values are ucm = (30:0, 6:0)T (8). As expected, compared
to the peak value of the control input in Case 2 for 10 ft
command, the control input is larger in this case.
hr

= 30 ft, uncertainty factor


6. Adaptive control:
au = 1:6, random Zd and Md , unmodeled nonlinearities (Zn , Mn )
Similar to Case 3, the performance of the controller
is examined in the presence of parameter uncertainties,
random inputs, and unmodeled nonlinearities. The
responses are collected in Figure 9. It is observed that

the submarine attains the desired depth (30 ft). The


maximum control magnitude is ucm = (30:0, 12:4)T (8).
The response time is about 130 s. The pitch angle
remains within 0.02. The control input saturated at
certain instants.
7. Adaptive control: effect of control saturation,
hr = 60 ft, uncertainty factor au = 1:6, (Zd , Md ) = 0
To examine the effect of control saturation, simulation was done using large depth command hr = 60 ft,
au = 1:6, G = 1000, (Zd , Md ) = 0, and vi = 0:07.
Because depth command is of larger value, a smaller
value of vi = 0:07 is used. It is observed in Figure 10
that the target depth is attained in about 171 s. The

Downloaded from pii.sagepub.com by guest on January 20, 2015

340

Proc IMechE Part I: J Systems and Control Engineering 228(5)

Figure 5. Adaptive control: hr = 10 ft, uncertainty factor au = 1:6, random disturbance Zd and Md : (a) depth h; (b) heave velocity w;
(c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
^1, m
^ 2 ; and
(h) parameter estimate s
^1 , s
^2.

control input saturates over a contiguous interval of


time, but this causes no difficulty in the trajectory control. Despite control saturation, control chattering does
not occur (Figure 10).
Simulation for other depth commands hr has also
been performed. The performance for hr = 50, 60,
70, 90 ft is summarized in Table 2. It has been found
that despite control signal saturation, larger target
depths can be attained by the choice of lower values of
the controller parameter vi . Interestingly, for all these
commands hr considered here, the control signal saturates only over certain contiguous interval of time
somewhat similar to Figure 10, before the target depth
is attained. (The responses for these are not shown here

in order to save space.) But the settling time is larger


for larger values of hr and one must choose smaller values of the design parameter vi for completing the maneuver. For a given command hr , the control saturation
period and Ts depend on the value of vi . Of course, the
necessity of larger settling time for attaining larger target depth with limited control magnitude is not
unreasonable.
Additional simulation results have been obtained to
examine the robustness of the controller with respect to
the uncertainty factor au and the strength of the disturbance input. These results are summarized in Tables
35. Table 3 shows the performance when the submarine (with (Zd , Md ) = 0) is commanded to different

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

341

Figure 6. Adaptive control: hr = 10 ft, uncertainty factor au = 1:6, random Zd and Md , unmodeled nonlinearities (Zn , Mn ): (a) depth
h; (b) heave velocity w; (c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter
estimate m
^1, m
^ 2 ; and (h) parameter estimate s
^1 , s
^2 .

Table 2. Effect of control saturation: hr : hr = 50, 60, 70, 90, au = 1:6, Zd = 0, Md = 0.
Depth command hr (ft)

vi

Ts (s)

ucm ()

um ()

50
60
70
90

0.08
0.07
0.06
0.05

149.5369
170.9839
200.3521
240.2922

(30.0000, 7.9188)
(30.0000, 7.7124)
(30.0000, 2.9597)
(30.0000, 3.7158)

0.5866
0.6561
0.1555
0.3222

target depths (10, 20, or 30 ft) using the controller of


Case 1. It is seen that peak value of the control input
increases with the command level, but the pitch angle

remains close to 0. The settling time Ts for each depth


command is almost the same. Table 4 shows the effect
of the choice of uncertainty factor au on the

Downloaded from pii.sagepub.com by guest on January 20, 2015

342

Proc IMechE Part I: J Systems and Control Engineering 228(5)

Figure 7. Adaptive control: hr = 30 ft, uncertainty factor au = 1:6, disturbance Zd = 0, Md = 0: (a) depth h; (b) heave velocity w;
(c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
^1, m
^ 2 ; and
(h) parameter estimate s
^1 , s
^2.

Table 3. Performance under different depth commands hr : au = 1:6, Zd = 0, Md = 0.


Depth command hr (ft)

Ts (s)

ucm ()

um ()

10
20
30

120.2985
120.2975
120.2977

(9.6232, 1.2490)
(19.2265, 2.5046)
(28.8696, 3.7466)

0.0029
0.0058
0.0087

performance. It is seen that the controller provides


robust performance, and the effect of the choice of the
uncertainty factor on the peak values of the control

input and pitch angle excursion is not significant.


Table 5 provides the performance of the controller
when random disturbance inputs of different strength

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

343

Figure 8. Adaptive control: hr = 30 ft, uncertainty factor au = 1:6, random disturbance Zd and Md : (a) depth h; (b) heave velocity w;
(c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
^1, m
^ 2 ; and
^2.
(h) parameter estimate s
^1 , s

Table 4. Effect of uncertainty factor au on performance: hr = 30 ft, Zd = 0, Md = 0.


Uncertainty factor au

Ts (s)

ucm ()

um ()

1.8
1.6
1.4
1.2
0.8
0.6
0.4
0.3
0.2

120.3012
120.2977
120.2943
120.2907
120.2837
120.2804
120.2769
120.2766
120.2747

(28.8747, 3.7659)
(28,8696, 3.7466)
(28.8050, 3.7475)
(28.7705, 3.7385)
(28.7022, 3.7206)
(28.6687, 3.7117)
(28.6360, 3.7036)
(28.6205, 3.6997)
(28.6087, 6.6375)

0.0097
0.0081
0.0076
0.0065
0.0043
0.0032
0.0021
0.0016
0.0011

Downloaded from pii.sagepub.com by guest on January 20, 2015

344

Proc IMechE Part I: J Systems and Control Engineering 228(5)

Figure 9. Adaptive control: hr = 30 ft, uncertainty factor au = 1:6, random Zd and Md , unmodeled nonlinearity (Zn , Mn ): (a) depth h;
(b) heave velocity w; (c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter
estimate m
^1, m
^ 2 ; and (h) parameter estimate s
^1 , s
^2 .

Table 5. Effect of disturbance inputs Zd and Md on performance: hr = 30 ft, au = 1:6.


Disturbance input (Zd , Md )

Ts (s)

ucm ()

um ()

Case 6 (paper)
53 Case 6
103 Case 6

120.2393
120.0490
NC

(30.0000, 5.9972)
(30.0000, 22.6765)
NC

0.0117
0.1136
NC

are acting on the vehicle. For this purpose, the disturbance inputs used in Case 2 are amplified by a factor of
5 or 10. It is seen that although the controller succeeds

in completing the maneuver when the disturbance is


five times larger, it is unable to control the vehicle when
the disturbance magnitude is increased by 10-fold. NC

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

345

Figure 10. Adaptive control: effect of control saturation, hr = 60 ft, uncertainty factor au = 1:6, (Zn , Mn ): (a) depth h; (b) heave
^ 1,
velocity w; (c) pitch rate Q; (d) pitch angle u; (e) control signal uc (dB, dS); (f) parameter estimate ^v11 , ^v22 ; (g) parameter estimate m
m
^ 2 ; and (h) parameter estimate s
^1 , s
^2.

denotes not controlled. The settling time remains


almost the same in each of these tables, and the pitch
angle remains close to 0.
Similar to Demirci and Kerestecioglu,28 it is
observed that the response time for the depth control is
large. But this response time is reasonable in view of
the design objective. Here the design objective is to
maintain the pitch angle excursion close to 0 for all
time during the trajectory control of the depth in spite
of the uncertainties and disturbance inputs. For this
kind of maneuver, larger convergence time is required
because only the net vertical depressing lift is used for
the depth control. Of course, the response time will

depend on the effectiveness of the hydroplanes in producing the control forces and moments.
To this end, it is appropriate to provide a comparison of the adaptive control systems and the reconfiguring VSC system published for this submarine model.
For the model without hydrodynamic nonlinearities,
the sliding mode controller of Demirci and
Kerestecioglu,28 the adaptive controller of Nambisan
and Singh,29 the noncertainty-equivalent adaptive system of Lee and Singh,30 and L1 adaptive autopilot have
somewhat similar response characteristics. Each of
these controllers accomplishes depth and pitch angle
control. But it is important to note that unlike the L1

Downloaded from pii.sagepub.com by guest on January 20, 2015

346

Proc IMechE Part I: J Systems and Control Engineering 228(5)

adaptive controller, the adaptive laws of Nambisan and


Singh29 and Lee and Singh30 have been derived under
the stringent assumption that the model is free of unmodeled dynamics. This limits the application of those controllers. It has been assumed in Demirci and
Kerestecioglu28 that the parameters of the model are
known for design. For 30 ft depth command in the presence of disturbance, the response time in Demirci and
Kerestecioglu28 is more than 300 s compared to 130 s in
this article (Figure 9). For the 30 ft depth control in the
presence of random disturbance inputs, the
noncertainty-equivalent adaptive law of Lee and Singh30
has response time less than 100 s, but the pitch angle
deviation is larger (about 0.75) and the control input
saturates over several longer intervals. The adaptive law
of Nambisan and Singh29 has slow response time (about
400 s) even for smaller command of 10 ft for the model
with random disturbance inputs. Interestingly, the simulation results obtained here show that the response time
of the submarine model controlled by the L1 adaptive
autopilot remained identical (about 130 s) even if unmodeled nonlinearities and random disturbances are introduced in the model. Furthermore, the structure of the
adaptive controller remains fixed, regardless of the nonlinearities and external disturbance inputs, retained in
the model of the submarine.

3.

4.

5.

6.

7.

8.

9.

10.

Conclusion
In this article, the dive-plane control of MIMO submarine model including time-varying unmodeled dynamics,
based on the L1 adaptive control theory, was considered. An L1 adaptive law was designed for the trajectory control of the depth and pitch angle. Interestingly,
unlike the traditional adaptive laws, the L1 autopilot
provides flexibility in the selection of large adaptation
gains for fast adaptation and achieving desirable transient and steady-state performance bounds. Simulation
results showed that the designed L1 adaptive controller
accomplished the specified dive-plane maneuvers
despite large parameter uncertainties, unmodeled nonlinearities, and disturbance inputs.

11.

12.

13.

14.

15.

16.

Declaration of conflicting interests


The authors declare that there is no conflict of interest.
Funding

17.

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit
sectors.

18.

19.

References
1. Fossen TI. Guidance and control of ocean vehicles. New
York: Wiley Publications, 1994.
2. You SS, Lim TW and Jeong SK. General path-following
maneuvers for an underwater vehicle using robust con-

20.

trol synthesis. Proc IMechE, Part I: J Systems and Control Engineering 2010; 224: 960969.
Petrich J and Stilwell DJ. Robust control for an autonomous underwater vehicle that suppresses pitch and yaw
coupling. Ocean Eng 2011; 38(1): 197204.
Naik MS and Singh SN. State-dependent Riccati
equation-based robust dive plane control of AUV with
control constraints. Ocean Eng 2007; 34: 17111723.
Chin CS, Lau MWS, Low E, et al. Robust and
decoupled cascaded control system of underwater robotic
vehicle for stabilization and pipeline tracking. Proc
IMechE, Part I: J Systems and Control Engineering 2008;
222(4): 261278.
Lau MWS, Swei SSM, Seet G, et al. Control of an
underactuated remotely operated underwater vehicle.
Proc IMechE, Part I: J Systems and Control Engineering
2003; 217(5): 343358.
Healey AJ and Lienard D. Multivariable sliding mode
control for autonomous diving and steering of unmanned
underwater vehicles. IEEE J Oceanic Eng 1993; 18(3):
327339.
Yoerger DR and Slotine JE. Robust trajectory control of
underwater vehicles. IEEE J Oceanic Eng 1985; 10(4):
462470.
Chin CS and Lum SH. Rapid modeling and control systems prototyping of a marine robotic vehicle with model
uncertainties using xPC target system. Ocean Eng 2011;
38: 21282141.
Soylu S, Buckham BJ and Podhorodeski RP. A
chattering-free sliding-mode controller for underwater
vehicles with fault-tolerant infinity-norm thrust allocation. Ocean Eng 2008; 35(16): 16471659.
Ishaque K, Abdullah SS, Ayob SM, et al. A simplified
approach to design fuzzy logic controller for an underwater vehicle. Ocean Eng 2011; 38(1): 271284.
Levedahl BA and Silverberg L. Control of underwater
vehicles in full unsteady flow. IEEE J Oceanic Eng 2009;
34(4): 656668.
Do KD, Pan J and Jiang ZP. Robust and adaptive path
following for underactuated autonomous underwater
vehicles. Ocean Eng 2004; 31(16): 19671997.
Li J and Lee P. Design of an adaptive nonlinear controller for depth control of an autonomous underwater vehicle. Ocean Eng 2005; 32(1718): 21652181.
Lapierre L and Jouvencel B. Robust nonlinear pathfollowing control of an AUV. IEEE J Oceanic Eng 2008;
33(2): 89102.
Trebi-Ollennu A and White BA. Non-linear robust control designs for a remotely operated underwater vehicle
depth control system. Proc IMechE, Part I: J Systems
and Control Engineering 1996; 210(3): 201214.
Ishii K, Fujii T and Ura T. An on-line adaptation method
in a neural network based control system for AUVs.
IEEE J Oceanic Eng 1995; 20(3): 221228.
Zhang M-J and Chu Z-Z. Adaptive sliding mode control
based on local recurrent neural networks for underwater
robot. Ocean Eng 2012; 45: 5662.
Zhang L-J, Qi X and Pang Y-J. Adaptive output feedback control based on DRFNN for AUV. Ocean Eng
2009; 36: 716722.
Wang J-S and Lee CSG. Self-adaptive recurrent neurofuzzy control of an autonomous underwater vehicle.
IEEE T Robotic Autom 2003; 19(2): 283295.

Downloaded from pii.sagepub.com by guest on January 20, 2015

Lee and Singh

347

21. Antonelli G, Caccavale F, Chiaverini S, et al. A novel


adaptive control law for underwater vehicles. IEEE T
Contr Syst T 2003; 11(2): 221232.
22. Mohan S and Kim J. Indirect adaptive control of an
autonomous underwater vehicle-manipulator system for
underwater manipulation tasks. Ocean Eng 2012; 54:
233243.
23. Singh SN, Simha A and Mittal R. Biorobotic AUV maneuvering by pectoral fins: inverse control design based on
CFD parameterization. IEEE J Oceanic Eng 2004; 29(3):
777785.
24. Narasimhan M, Dong H, Mittal R, et al. Optimal yaw
regulation and trajectory control of biorobotic AUV
using mechanical fins based on CFD parameterization. J
Fluid Eng: T ASME 2006; 128(4): 687698.
25. Naik MS, Singh SN and Mittal R. Indirect adaptive output feedback control of a biorobotic AUV using pectorallike mechanical fins. Bioinspir Biomim 2009; 4(2): 026001
(11 pp.).
26. Babaoglu OK. Designing an automatic control system for
a submarine. Masters Thesis, Naval Postgraduate School,
Monterey, CA, 1988.
27. Dumlu D and Istefanopulos Y. Design of an adaptive
controller for submersibles via multimodel gain scheduling. Ocean Eng 1995; 22(6): 593614.

28. Demirci U and Kerestecioglu F. A re-configuring slidingmode controller with adjustable robustness. Ocean Eng
2004; 31(13): 16691682.
29. Nambisan PR and Singh SN. Multi-variable adaptive
back-stepping control of submersibles using SDU decomposition. Ocean Eng 2009; 36(2): 158167.
30. Lee KW and Singh SN. Noncertainty-equivalent multivariable adaptive control of submersibles using filtered
signals. Ocean Eng 2012; 53: 98110.
31. Hovakimyan N and Cao C. Adaptive control theory: guaranteed robustness with fast adaptation. Philadelphia, PA:
SIAM, 2010, pp.94105.
32. Cao C and Hovakimyan N. L1 adaptive controller for
nonlinear systems in the presence of unmodelled
dynamics: part II. In: Proceedings of American control
conference, Seattle, WA, 1113 June 2008, pp.40994104.
New York: IEEE.
33. Gregory IM, Cao C, Xargay E, et al. L1 adaptive control
design for NASA AirSTAR flight test vehicle. In: AIAA
guidance, navigation and control conference, Chicago, IL,
1013 August 2009, AIAA-2009-5738. Reston, VA:
AIAA.
34. Pomet J-B and Praly L. Adaptive nonlinear regulation:
estimation from the Lyapunov equation. IEEE T Automat Contr 1992; 37(6): 729740.

Appendix 1
Hydrodynamic coefficients27,28
Parameter value

Parameter value

Parameter value

Z9w =  0:0110
Z9u =  0:0002
M9w = 0:0030
M9u =  0:0004
I9y = 5:68673104
ZG  ZB =  1:5 ft
I92 = I9y  M9u

Z9w_ =  0:0075
Z9dB =  0:0025
M9w_ =  0:0002
M9dB = 0:0005
L = 286 ft
U = 8:43 ft=s
m9 = 2m=(rL3 )

Z9u_ =  0:0045
Z9dS =  0:0050
M9u_ =  0:0025
M9dS =  0:0025
m = 1:523105 slug
r = 2:0 slug=ft3
m93 = m9  Z9w_

Downloaded from pii.sagepub.com by guest on January 20, 2015

You might also like