Ocean Engineering 233 (2021) 109071

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Review

Finite-time velocity-observed based adaptive output-feedback trajectory

tracking formation control for underactuated unmanned underwater
vehicles with prescribed transient performance
Hongtao Liang a, Yanfang Fu b, Jie Gao a, *, Hui Cao a, **
a
School of Physics and Information Technology, Shaanxi Normal University, Xian, 710119. China
b
School of Computer Science & Engineering, Xi’an Technological University, Xi’an, 710032. China

A R T I C L E I N F O A B S T R A C T

Keywords: This paper presents a finite-time velocity-observed based adaptive output-feedback trajectory tracking formation
Unmanned underwater vehicle control for underactuated unmanned underwater vehicles (UUVs) with prescribed transient performance. For
Leader-follower realistic dynamical UUV model, we assume that the mass and damping matrices are off-diagonal and only partial
Finite-time velocity observer
information about hydrodynamics is known, and we simultaneously consider the problems of input saturation,
Transient prescribed performance
unmeasured velocity, external disturbance, and prescribed performance to design formation controller. For this
Output-feedback trajectory tracking
purpose, a leader-follower pattern is employed, where each follower tracks its reference trajectory defined by the
leader position and predetermined formation without requirement of leader velocity. Moreover, a finite-time
velocity observer is constructed by utilizing only position and attitude information in a short time for
increasing the robustness, where the estimates are substituted for unmeasured velocities in output feedback.
Then, to guarantee that all output-states do not violate prescribed transient performances, a proper error system
is established by non-logarithmic error transformation function. Based on these designs, an output-feedback
adaptive formation scheme is proposed, where neural networks are utilized for identifying lumped nonlinear
terms composed of uncertain hydrodynamics and external disturbances, and auxiliary variables and their
additional control terms are presented to handle input saturation and underactuated problems simultaneously.
Using Lyapunov stability theorem, it is proved that all closed-loop signals are uniformly ultimately bounded
(UUB), and tracking errors can be guaranteed within prescribed performance bounds. Simulation results
demonstrate effectiveness of the proposed formation control scheme.

1. Introduce behavior-based strategy (Liang et al., 2020a,b; Balch et al., 1998) and
artificial potential field method (Fiorelli et al., 2006). In particalar,
The trajectory tracking problem of unmanned underwater vehicles leader-follower has achieved a significant theoretical formalization due
(UUVs) has aroused more compelling interest in ocean engineering over to the simplicity and reliability (Cui et al., 2010; Wang et al., 2019a,b,
the past decades, including marine environment monitoring, underwa­ 2020; Qi and Cai, 2018; Liu et al., 2019), which is essentially to ensure
ter topograph explorating, hydrological date collection, and marine that all followers track the orientation and distance of a real or virtuel
rescue (Sahoo et al., 2019). In order to exhibit its greater efficiency in leader with prescribed offset.
contrast to a single one, formation control for multiple UUVs has been However, accurate tracking control of UUVs is a major challenge in
more constructive in the real tracking applications (Wu et al., 2020; real oceanic engineering and application, since full knowledge of UUV
Saback et al., 2020; Sarda and Dhanak, 2019). Aimed at achieving ex­ dynamic model cannot be exactly obtained due to its inherent proper­
pected fomation pattern, various remarkable results have been pro­ ties, i.e., high nonlinearities and parameter perturbations. Therefore, the
posed, i.e., leader-follower pattern (Cui et al., 2010; Wang et al., 2020a, control approach designed for tracking control should be able to alle­
b), virtual structure (Makavita et al., 2019; Beard et al., 2001), viate the adverse effects caused by uncertainties. In this respect,

* Corresponding author.
** Corresponding author.
E-mail addresses: [email protected] (J. Gao), [email protected] (H. Cao).

https://doi.org/10.1016/j.oceaneng.2021.109071
Received 15 January 2021; Received in revised form 31 March 2021; Accepted 22 April 2021
Available online 30 May 2021
0029-8018/© 2021 Elsevier Ltd. All rights reserved.
H. Liang et al. Ocean Engineering 233 (2021) 109071

numerous control techniques have been proposed, including adaptive to simultaneously identify the unmeasured velocities and uncertainties,
control (Makavita et al., 2019; Antonelli et al., 2003), fuzzy control but its convergence was not assured. Among the above results, existing
(Sahu and Subudhi, 2018), sliding mode control (Qiao and Zhang, 2019, observers are mostly infinite-time (Qu et al., 2020). Accordingly, a
2020), dynamic surface control (Wang et al., 2019a,b), neural network finite-time velocity observer involved with output-feedback tracking
(NN) (Lu et al., 2018; Park, 2015), model predictive control (Saback control for underactuated UUV formation is still an open issue.
et al., 2020; Li et al., 2017), data-driven control (Peng et al., 2019a,b), Besides output-feedback tracking control, some works can only
extended-state-observer (ESO) (Peng et al., 2019a,b), and guarantee uniformly ultimate boundedness while the transient perfor­
Lyapunov-direct function (Liu and Chen, 2020). These results have been mance, such as convergence rate, maximum overshoot, or steady error
reported to achieve robustness against uncertainties, and where bounds, was not fully considered. From a practical perspective, transient
NN-based technique has been becoming one of the most promising tools performance can not only improve control performance, but also ensure
to compensate for uncertain terms (Cui et al., 2010; Wang et al., 2020a, safety when vehicles influence by external disturbances induced by
b; Lu et al., 2018; Park, 2015; Peng et al., 2019a, 2019b). oceanic waves, currents and tides. Therefore, it is necessary to satisfy
Unfortunately, some of the aforementioned studies only consider the requirements of prescribed transient-state and steady-state perfor­
case that the UUV is fully actuated. In practice, most torpedo-like UUVs mance. To end this, a remarkable prescribed performance was originally
are underactuated (Cui et al., 2010; Liu et al., 2019), which generally proposed for nonlinear system (Bechlioulis and Rovithakis, 2008),
have three freedom degrees but only two control signals, i.e., surge force which can guarantee that all errors converge to a prescribed arbitrary
and yaw moment, are available for control design. It is clear that an small residual set while transient performance and overshoot are smaller
underactuated UUV invariably suffers from variations in sway direction, than a sufficiently small preassigned constant. This method has been
thereby resulting in non-integral constraints. In addition, an accompa­ efficiently utilized in several applications, i.e., robots, spacecraft, and
nied problem of input saturation may be easily triggered due to its visual-servo. For the tracking control of underwater vehicle with pre­
restricted physical actuators (Li et al., 2019; Omid and Khoshnam, scribed performance, a control scheme with prescribed performance for
2020). Therefore, comparing with control approaches involving with underactuated UUVs was firstly proposed by (Bechlioulis et al., 2017). In
fully-actuated UUVs, complicated uncertainty to underactuated UUVs (Omid and Khoshnam, 2018), a neural network-based tracking control
has attracted a considerable amount of attentions on attempting to with prescribed performance was investigated to design a full-state
design high-performance tracking control approaches. feedback controller for underactuated autonomous underwater vehi­
Because of the design simplification and ease of implementation, cles (AUVs). In (Liu et al., 2020), an adaptive thruster-fault tracking
several traditional approaches, such as local linearization and input- control problem was addressed with prescribed transient performance.
output decoupling, have been employed to design tracking controllers In addition, Dai et al. (2018) had developed a full-state feedback platoon
(Fossen, 1994). However, it has already been pointed in (Bechlioulis formation controller using the benefit of prescribed performance. To
et al., 2017) that closed-loop performance is poor. An alternative improve convergence rate, finite-time performance function (Wang
approach involved with backstepping technique (Cui et al., 2010), et al., 2021) and non-logarithmic piecewise error mapping function (Li
which, however, suffered from an explosion of complexity problem et al., 2020) have also been introduced into tracking control problem.
caused by repeated differentiations (Peng et al., 2017a,b). To end this, Apparently, all aforementioned works are essentially a path-
dynamic surface control (DSC) techniques, i.e., neural-DSC(Li et al., following problem where the reference path is generated by a virtual
2019) and fuzzy-DSC (Liang et al., 2020a,b), were reported for un­ point with an independent variable, rather than the more difficult tra­
certainties of underactuated UUV. In the aim of handling uncertainties, a jectory tracking, since the controller requires to drive the UUV reaching
neural-based adaptive robust formation control strategy was proposed and tracking a time-varying trajectory on time (Jia et al., 2019).
in (Peng et al., 2011). In the presence of unknown hydrodynamic Moreover, it is turned out that these controllers are all based on an
damping, an adaptive neural-based leader-following formation assumption that off-diagonal terms in mass and damping matrices of
controller was presented for multiple UUVs (Park, 2015). It was underactuated UUV are zero. Indeed, this restrictive assumption is not
worthwhile mentioning that a novel approach angle method to track any tenable, the reason is that most underactuated UUVs are not a
reference paths, i.e., curve-path and straight-path, was introduced into semi-submerged sphere. Therefore, yaw moment control effects directly
control input to solve underactuation, which only required the position on sway dynamics, which results in the difficulty in control design. In
information, unlike the approach angle only suitable for case of order to overcome this problem, Do et al. (2005) firstly proposed a
straight-path but requiring both position and velocity information in nonlinear coordinate transformation with diagonal form to relax this
(Lionel and Didik, 2007). Based on previous works, a robust sliding restrictive assumption. Furthermore, based on the coordinate trans­
mode formation control strategy of underactuated UUV was presented formation with diagonal form, containment control and formation
by utilizing backstepping, adaptive neural network and DSC technique, control design (Sung and Bong, 2015) can be achieved. Although all
but its obtained results should hold an important assumption that the these works benefit from the above assumption, it also requires accurate
sway velocity was passive bounded (Wang et al., 2020a,b). model parameters on hydrodynamic damping.
It is noted that all aforementioned results are obtained under an In general, viscous hydrodynamic damping of UUV is mainly caused
assumption that the exact full-state information, including the linear and by potential damping, possible skin friction, wave drift damping, vortex
angular velocities in body-fixed reference frame and the positions and shedding damping and lifting forces. Different damping terms can be
attitude in inertial reference frame, is available for feedback control divided into both linear and nonlinear damping (Fossen, 2011). The
design. However, velocity information may not be readily acquired due linear damping coefficients, i.e., potential damping and possible skin
to the considering equipment cost and the risk of sensors damage, even if friction, can be calculated by semi-exprical method or computation
it can be acquired, certain signal noises will be brought. Therefore, it is programs. However, nonlinear damping coefficients, such as wave drift
necessary to design an output-feedback control, and some control damping, vortex shedding damping and lifting forces, are difficult to
schemes have been reported for nonlinear system based on observers acquire in most cases due to their nonlinear characteristics (Jia et al.,
(Chen et al., 2020; Li et al., 2015; Peng et al., 2017a,b). In order to 2019). Therefore, discrepancies between the nominal and actual model
recover unavailable velocity, high-gain observers (Chen et al., 2020) are utilized to design the controller with only partial information (Chen
were commonly developed with only position and yaw angle informa­ et al., 2020).
tion whereas uncertainties cannot be estimated. A backstepping-based Motivated by aforementioned observations, challenges on accurate
observer (Li et al., 2015) was designed to reconstruct velocity, but all control for UUV are mainly derived from uncertainties, i.e., parameter
parameters were required to be known. Moreover, a perturbation, input saturation, underactuated constricts, and prescribed
echo-state-network-based observer (Peng et al., 2017a,b) was designed transient and steady-state performance, as well as external disturbance.

2
H. Liang et al. Ocean Engineering 233 (2021) 109071

To our best knowledge, there is no work on accommodating these issues Lemma 2. (Khalil, 2002) Young’ Inequality is expressed for any
simultaneously in leader-follower formation tracking control for positive∀a, b ∈ R
underactuated UUVs. Therefore, a finite-time velocity-observed based ⃒ ⃒
kp ⃒ 1 ⃒
adaptive output-feedback formation tracking controller is investigated ab ≤ ⃒⃒a|p + q ⃒⃒b|q (3)
p qk
to achieve trajectory tracking, while ensuring the prescribed transient
and steady-state performance on output tracking errors. For a more
where k > 0, p > 1, q > 1, (p − 1)(q − 1) = 1.
general case, we design this controller under the assumption that mass
and hydrodynamic terms are off-diagonal, and prior information on Lemma 3. (Shtessel et al., 2007) Consider a nonlinear systemẋ(t) =
hydrodynamics is partial uncertain, meanwhile velocity information is f(x(t))
unmeasurable. Main contributions of this paper are summarized as
follows: ⎧ ẋ0 = z0 + x1
⎪
⎪
⎪
⎪ z0 = − ϖ 0 L1/n+1 [x0 ]n/n+1 sgn(x0 )
⎪
⎪
(1) In contrast to existing results (Cui et al., 2010; Dennis et al., 2019; ⎪
⎪
⎪
ẋ1 = z1 + x2
[ ]n− 1/n ( )
⎪
⎪
Li et al., 2020; Wang et al., 2020a, 2020b), the proposed forma­ ⎪
⎪ z1 = − ϖ 1 L1/n x1 − ẋ0
⎪ sgn x1 − ẋ0
⎪
tion control scheme based on leader-follower only relies on both ⎪
⎪
⎨⋮
positions and orientations available to track both straight-line (4)
⎪ ẋn− 1 = zn− 1 + xn
path and curve-line path, where required velocities for output ⎪
⎪
⎪ [ ]1/2 ( )
⎪
⎪
feedback control designing are recovered by using finite-time ⎪
⎪
⎪
zn− 1 = − ϖ n− 1 L1/2 xn− 1 − ẋn− 2 sgn xn− 1 − ẋn− 2
⎪
velocity observer, rather than the most commonly used ⎪
⎪
⎪ ẋn = zn + [ − L,(L]
⎪
⎪ )
infinite-time high-gain observers (Chen et al., 2020). It is more ⎪
⎪
⎩ z = − ϖ Lsgn x − ẋ
convenient to implement this in practice. n n n n− 1

(2) In order to achieve the tracking with prescribed transient per­

formance despite the presence of parameter uncertainty and where ϖ i (i = 0, 1, ⋯, m) and L are positive constants. Hence, the original
external disturbances, a novel output-feedback trajectory system ẋ(t) = f(x(t))is finite-time stable by aiding of setting suitable
tracking formation controller with prescribed performance for ϖ i > 0 and L > 0.
underactuated UUVs is firstly developed, where tracking errors in
Remark 1. Due to the expensive cost and simplifying mechanism,
transformed coordinate can be guaranteed to satisfy prescribed
velocity measuring instruments may not be equipped on UUV body, and
transient and steady state behaviors, i.e., convergence rate and
velocity information is unavailable for designing output feedback con­
maximum overshoot, by means of the non-logarithmic error
trol. Such that, Lemma 3 can be employed to design a nonlinear finite-
transformation functions.
time velocity observer for estimating unmeasurable velocities.
(3) Simultaneously increasing greatly robustness against input satu­
ration and underactuated constricts, and discarding the assump­
tion that sway velocity is passive bounded in previous works 2.2. Modeling UUV formation system
(Wang et al., 2019a,b, 2020; Omid and Khoshnam, 2018; Jia
et al., 2019), three auxiliary virtual variables and their additional Consider a leader-follower formation system consisting of N under­
control terms are designed based on filter-backstepping and actuated UUVs with index set N = {1, 2, ⋯, N}. To accurately illustrate
adaptive neural network, and all of which are incorporated into the motion of each UUV, earth-fixed frame {E} and body-fixed frame{B}
the proposed output-feedback tracking controller. Moreover, all are formulated in Fig. 1.
closed-loop signals are uniformly ultimately bounded within the The kinematic and dynamic model of the i-th UUV is provided as (Do
given prescribed performance bounds. and Pan, 2009; Fossen, 2011)
{
The reminder of this paper is organized as follows. Section 2 in­ η̇i = J(ψ i )(νi ) ( )
(5)
troduces the preliminaries and problem formulation. Section 3 details Mi ν̇i + Ci νi νi + Di νi νi = δi (τ) + wi
the adaptive robust output-feedback control design. Section 4 gives
main results. Section 5 illustrates simulation experiments. Finally, Sec­ where ηi = [xi , yi , ψ i ]T ∈ ℝ3 is a vector of position (xi , yi ) and yaw angle
tion 6 concludes this paper and discusses the future work. ψ i ∈ [0, 2π) in earth-fixed frame, respectively; νi = [ui , vi , ri ]T ∈ ℝ3 is a
velocity vector in surge ui , sway vi , and yaw ri in body-fixed frame; wi =
2. Preliminaries and problem formulation

2.1. Preliminaries

Proposition 1. (Ryan, 1991) For a general initial-value problem

ω̇(t) = F[ω(t)], ​ ω(0) = ω0 , ​ ω0 ∈ Rm , ​ ω(t) ∈ Rm (1)

where F[ω(t)] is a continuous function on Rm , if the solution for (1) is

bounded, then t→∞.
Lemma 1. (Wen et al., 2018) Consider a smooth function V(t), and its
initial stateV(0)is bounded. If there exists an inequalityV̇(t) ≤ − 2μV +
Cholding withC > 0andμ > 0for any time, this inequality can be obtained
( )
C C
V̇(t) ≤ V(0) − exp( − μt) + (2)
2μ 2μ

Fig. 1. Illustrating UUV formation with {B} and {E} frames.

3
H. Liang et al. Ocean Engineering 233 (2021) 109071

[wui , wvi , wri ]T denotes the external unknown disturbances; J(ψ i ) ∈ ℝ3×3 is χ vi ] = τi − δ(τi ) is defined to express mismatch between the control with
the rotation matrix, and М i ∈ ℝ3×3 , Ci ∈ ℝ3×3 andDi ∈ ℝ3×3 are the non- saturation and without saturation. Hence, saturation control in (5) can
diagonal inertia matrix, Coriolis matrix and damping matrix, respec­ be given by δ(τi ) = τi − χ i .
tively, all of which are given by Remark 2. The partial information about underactuated UUV in (7)
⎡
cos(ψ i ) − sin(ψ i ) 0
⎤ are available in nominal part, which can be utilized for control design
⎣
J(ψ i ) = sin(ψ i ) cos(ψ i ) 0⎦ (6a) purpose, unlike the previous works either considering full prior
0 0 1 knowledge (Do and Pan, 2005) or without any prior knowledge (Wang
et al., 2017) on model parameters. The two cases are both not real from a
⎡ ⎤
m11i 0 0 practical point of view. Moreover, shift part between normal and actual
Мi = ⎣ 0 m22i m23i ⎦ (6b) model can be approximated by NNs.
0 m23i m33i
⎡ ⎤ 2.3. Coordinate transformation
0 0 − m22i vi − m23i ri
Ci = ⎣ 0 0 m11i ui ⎦ (6c) Most like-torpedo UUVs are not submerged sphere (Bechlioulis et al.,
m22i vi + m23 ri − m11i ui 0 2017), so mass matrices М i cannot be assumed as diagonal due to its
⎡ ⎤ non-diagonal item m23i . This means that the yaw moment τr can influ­
d11 (ui ) 0 0
ence directly on sway dynamics in (5). In such case, it is difficult to
Di = ⎣ 0 d22 (vi , ri ) d23 (vi , ri ) ⎦ (6d)
formulate the control design. Therefore, we introduce a coordinate
0 d32 (vi , ri ) d33 (vi , ri )
transformation (Do and Pan, 2005) to make the non-diagonal mass
where m11i = mi − Xu̇i ,m22i = mi − Yvi˙ ,m23i = mi xgi − Yri˙ ,m33i = matrices as a diagonal form
⎧
Izi − Nri˙ ,d11 (ui ) = − (Xui + X|ui|ui |ui |),d22 (vi ,ri ) = − (Yvi + Y|vi|vi |vi | + + ⎪
⎪
⎨ xi = xi + ξi sin(ψ i )
Y|ri|vi |ri |), d23 (vi , ri ) = − (Yri + Y|vi|ri |ri | + + Y|ri|ri |ri |),
yi = yi + ξi cos(ψ i ) (11)
d32 (vi , ri ) = − (Nvi + N|vi|vi |vi | + + N|ri|vi |ri |), d33 (vi , ri ) = − (Nri + ⎪
⎪
⎩ vi = vri + ξi ri
N|vi|ri |vi | + + N|ri|ri |ri |). Here, mi denotes the mass of UUV, xgi is the
center of gravity in body-fixed frame, and Izi is the moment of inertia
with respect to yaw; Xu̇i ,Yv̇i , Yri˙ and Nri˙ are the added masses; Xui , X|ui|ui , where ξi = m*23i /m*22i . Then, the UUV mathematical model can be
Yvi , Y|vi|vi , Y|ri|vi , Y|vi|ri , Y|ri|ri , Nvi , Nri , N|vi|vi , N|ri|vi , N|vi|ri and N|ri|ri , are the rewritten as
⎧
hydrodynamic derivatives. τi = [τui , 0, τri ]T ∈ ℝ3 denotes the control ⎪
⎪
⎨ ẋi = ui cos(ψ i ) − vi sin(ψ i )
input vector consisting of surge forceτui and yaw moment τri .
ẏi = ui sin(ψ i ) + vi cos(ψ i ) (12)
Suffering from the unmodelled dynamics and parametric perturba­ ⎪
⎪
⎩ ψ̇ i = ri
tions, the precise modeling on hydrodynamic is difficult to be obtained,
such that the dynamics matrices can be divided into the nominal part
⎧ /
and shift part (Jia et al., 2019; Chen et al., 2020), i.e.,Mi = M*i + ΔMi , ⎪
⎪
⎪
⎪ u̇ i = φui + ϕui + (τ ui − χ ) m*11i
Ci = C*i + ΔCi , and Di = D*i + ΔDi , where (⋅)*i represents the nominal
ui
⎨
(13)
parameter, and Δ(⋅)i describes a shifted difference between real and ⎪ v̇i = φvi + ϕvi
⎪ /
⎪
nominal parameters. Therefore, the dynamics (5) can be formatted ⎪
⎩ ψ̇ = φ + ϕ + (τri − χ ) Δi
i ri ri ri
( ) ( )
М *i ν̇i + Ci* νi νi + D*i νi νi = δi (τ) + di (7)
m*22i m*23i 2 d*23i (uri ) m*11i d*22i (vri ,ri )
where φui = v r
m*11i ri i
+ r
m*11i i
− m*11i
uri ,φvi = − u r
m*22i ri i
− m*22i
vri −
where di = − ΔМ i ν̇i − ΔCi (νi )νi − ΔDi (νi )νi + wi . And the kinematic d*23i (vri ,ri )
ri ,
model for the i-th UUV can be formulated as m*22i
{
1 * *
⎧ φri = − Δi (m11i m22i − m*22i m*22i )uri vri + (m*11i m*23i −
*
d*32i (vri , ri )vri ) + m*23i (d*23i (vri ,
⎪
⎨ ẋi = ui cos(ψ i ) − vi sin(ψ i ) m*22i m*23i )uri ri − *
m22i (d33i (vri , ri )ri +
ẏi = ui sin(ψ i ) + vi cos(ψ i ) (8) − m*23i dvi +m*22i dri
⎪
⎩ ψ̇ = ri ri )ri + d*22i (vri , ri )vri )},ϕui = dui
m*11i
, ϕvi = dvi
m*22i
, ϕri = Δi ,Δi =
i

m*22i m*33i − m*23i m*23i ,j = u, v, r, i = 1, 2, ⋯, N. In addition, we define ηi =

[xi , yi , ψ i ]T and νi = [ui , vi , ri ]T .
Assumption 1. The unknown disturbance wi = [wui , wvi , wri ]T and its
Assumption 3. We assume that vi = [ui , vi , ri ]T is bounded due to
derivative are bounded and satisfy the following condition
saturated control input (10), and M*i is known, so φui , φvi , φvi are Lip­
⃦ ⃦ schitz with respect to velocity vi = [ui , vi , ri ]T , and ϕui ,ϕvi , ϕri are boun­
⃦ẇi (t)⃦ ≤ wi (9) ded and unknown (Chen et al., 2020).
Remark 3. The M*i , C*i and D*i can be calculated using semi-empirical
where wi is unknown positive constant. methods or hydrodynamic computation programs. However,φui , φvi ,
Assumption 2. The control input τi is determined by saturation φvi and ϕui , ϕvi , ϕri involved with the nonlinear damping terms are
function δ(τi ) = [δ(τui ), 0, δ(τri )]T ∈ ℝ3 as follows difficult to obtain accurately, and if such inaccurate information is uti­
lized, it may lead to undesired performance (Park, 2015). Such that we
⎧ will design the controller without nonlinear damping matrix
⎨ τji,max if τji > τji,max
( )
δ τji = τji, if τji,min < τji < τji,max , j = u, r ​ (10) information.
⎩
τji,min if τji < τji,min Remark 4. The coordinate transformation in (1) is equivalent to
moving the frame {B} forward a short distanceεi = m*23 /m*22 in surge
where τji,max and τji,min are upper and lower bounds of control forces or
direction, such that diagonal forms (12) and (13) derived from (5) can be
moments, respectively. In addition, a dead-zero function χ i = [χ ui , 0,
achieved.

4
H. Liang et al. Ocean Engineering 233 (2021) 109071

2.4. RBF neural networks Fig. 1. The predefined desired distancedif and angleϕif are introduced to
illustrate the relative position betweenηl andηir , which is formulated as
Due to the good approximation, RBF-NNs are employed to address follows
estimating uncertain hydrodynamic damping terms. In this paper, we ⎧
use the following RBF-NNs to approximate a continuous smooth ( )
functionh(Z) : Rq →R:
⎨x ir = xl + dif cos ψ l + ϕif
( )
(21)
h(Z) = W T Θ(Z) (14) ⎩ yir = yl + dif sin ψ l + ϕif
ψ ir = ψ l
where input vectorZ ∈ Ω⊂Rq , weighting vectorW ∈ Rn , ndenotes the
number of neurons, and Θ(Z) = [Θ1 (Z), Θ2 (Z), ⋯, Θn (Z)] ∈ Rn is a
wherexl = xl + ξl cos(ψ l ), yl = yl + ξl sin(ψ l ).
nonlinear activation vector, which is given by
Define the transformed positions and yaw error
[ ]
(Z − ci )T (Z − ci ) ⎧
Θi (Z) = exp − 2
, i = 1, 2, ⋯, n (15) ⎨ xei = xi − xir
σi y = yi − yir (22)
⎩ ei
ψ ei = ψ i − ψ ai
whereci andσi are the center and width of the i-th Gaussian function,
respectively.
whereψ ai is an approach angle to track desired leader for the i-th UUV
It has been proved that the aforesaid RBF-NNs can approximate any
formulated as (Park, 2015)
continuous functionh(Z) over a compact setΩZ ⊂Rq to any arbitrary de­
( ) ( / ) (( ( / ))
gree of accuracy as ψ ai = atan yei / xei tanh H 2 a0 + ψ l 1 − tanh H 2 a0 (23)

h(Z) = W *T Θ(Z) + ε(Z) ∈ ∀Z ∈ Ω⊂Rq (16) √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅

wherea0 is a positive constant, andH = x2ei + y2ei .
whereε(Z)is a bounded approximation error satisfying|ε| ≤ ε* , and the
Assumption 6. Theηi = [xi , yi , ψ i ]T is available andνi = [ui , vi , ri ]T is
unknown constantε* > 0is for allZ ∈ ΩZ . AndW* is an optimal weight
unmeasured.
vector, and it is defined as
{ } Assumption 7. The leader positionηl = [xl , yl , ψ l ]T is bounded and
⃒ ⃒
W * = arg min sup ⃒h(Z) − W T Θ(Z)⃒ (17) available only for followers who are within communicating region.
W∈Rn Z∈ΩZ
Assumption 8. The time-varying unknownψ ai is bounded such
It is noted that bothW* andε* are unknown constants, which are only ⃒ ⃒
that⃒ψ̇ ai ⃒ ≤ ψ *ai , whereψ *ai is positive constant.
⌢
used for analytical propose. For real scenarios, an estimationWis used
for practical approximation instead of theW* . Hence, an estimation Remark 5. As stated in Assumptions 1, 2, 3, 6, 7 and 8, model un­
ofh(Z)can be given by certainty and external disturbance and their bounds are unknown in this
paper. The most commonly used methods such as function approx­
⌢T
imators and adaptive techniques for UUV need great computation and
⌢
h (Z) = W Θ(Z) (18)
time for estimating uncertainties and their bounds, especially for
unmeasuredνi = [ui , vi , ri ]T .
⌢
whereWcan be estimated online by designing adaptive law.
Assumption 4. TheW* is bounded, i.e.,||W* || ≤ W* , whereW* is a pos­
3. Adaptive robust output-feedback control design
itive constant.
⌢
Assumption 5. (Ge and Wang, 2002) For‖W‖ ≤ W* andW
̃ = W − W* , The section is devoted to designing the adaptive robust output
an error can be formulated by feedback control scheme for the leader-following formation.

⌢T
̃ T Θ(Z) = W Θ(Z) − W * T Θ(Z)
W (19)
3.1. Finite-time velocity observer
and then following inequalities can hold
(⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ) In practical application, UUV positionη = [x, y, ψ ]T can be readily
̃ T W ≤ χ ⃒⃒⃒⃒W * ⃒⃒⃒⃒ − ⃒⃒⃒⃒W measured via sonars, whereas its velocityν = [u, v, r]T may be unmea­
⌢T
− χW ̃ ⃒⃒⃒⃒ (20a)
2
sured due to the expensive cost of measured devices, i.e., doppler ve­
||Θi (Z)|| ≤ ci (20b) locity log (DVL) and inertial navigation system (INS). To identify the
velocity for output-feedback control design, we develop a finite-time
whereχ andci are positive constants. velocity observer to exactly estimate the velocity by aiding of only po­
sition information. Hence, observer for the i-th UUV is constructed based
2.5. Problem formulation on kinematics (5) as follows:
⎧⌢
⎪ η = z0,i
The objective of this paper is to design an adaptive output-feedback ⎪
⎪
⎪ i
⎨ ⃒ ( )
1/2 ⃒⌢
formation controller for N underactuated UUVs such that each vehicle
⌢ ⌢
z0,i = − ϖ 1,i Li ⃒ η i − ηi |1/2 sgn η i − ηi + J ν i (24)
can track a desired leader with the prescribed performance, while the ⎪
⎪
⎪
⎪
( )
⎩ J⌢ ⌢
ν i = − ϖ 2,i Li sgn J ν i − z0,i
tracking errors are confirmed to the predefined performance bounds in
the presence of model uncertainties and external disturbances.
Letηl = [xl , yl , ψ l ]T andηir = [xir , yir , ψ ir ]T be the desired leader and whereϖ 1,i > 0andϖ 2,i > 0are designing observer coefficients, Li =
reference trajectory of the i-th follower, respectively, as illustrated in diag[lui , lvi , lri ] ∈ ℝ3 is gain with lui > 0, lvi > 0, lri > 0, and η i and ν i are the
⌢ ⌢

5
H. Liang et al. Ocean Engineering 233 (2021) 109071

estimations forηi andνi , respectively. maximum overshot is prescribed less thanαj ρij (0)ifei (0) > 0or less
Consider observation errors between estimation and true value as thanβj ρij (0)ifei (0) < 0, which may even become zero by settingαij =
{
⌢ 0orβij = 0, respectively. Such that, ei can be imposed performance
εiη = η i − ηi
⌢ (25) bounds by elaborately selecting parameters.
εiν = J ν i − J νi
However, it is a major challenge to ensure that (30) is always satis­
Taking the derivative of (25) along (24), we have fied. Hence, an error transformation is defined to achieve the system
incorporated with the performance constraints (29) as follows

⎧ ⃒ ⃒1/2 ( ) ⃒ ⃒1/2 ( )
⎪ 1/2 ⃒⌢ ⃒ ⌢ ⌢ 1/2 ⃒ ⃒
⎨ ε̇iη = − ϖ 1,i Li ⃒ η i − ηi ⃒ sgn η i − ηi + J ν i − J νi = − ϖ 1,i Li ⃒εiη ⃒ sgn εiη + εiν
( ⃒ ⃒1/2 ( )) ( ) [ ] (26)
⎪
⎩ ε̇ = − ϖ L sgn ϖ L1/2 ⃒⃒ε ⃒⃒ sgn ε
iν 2,i i 1,i i iη iη − J ν̇i = − ϖ 2,i Li sgn εiν − ε̇iη + − D,D

Applying the Lemma 3, it is indicated from (26) that velocity

observation errorεiν can be ensured to converge to zero within a finite xei (t) = ρi1 (t)Si1 (e
( xi ))
time, i.e., existing0 < T < ∞such that yei (t) = ρi2 (t)Si2 eyi (31)
⃒⃒ ⃒⃒ ⃒ ⃒⃒ ψ ei (t) = ρi3 (t)Si3 (eψ i )
⃒⃒ ⃒⃒ ⃒ ⃒⃒
⃒⃒εiν (t)⃒⃒ + ⃒ε̇iν (t)⃒⃒ ≤ εe ν , t < T (27)
whereEi = [exi , eyi , eψ i ]T ∈ ℝ3 is the transformed error vector, andSij ( ⋅)is
⌢ ⌢ ⌢
u (t) ≡ u(t), v (t) ≡ v(t), r (t) ≡ r(t), T ≤ t (28) defined as strictly monotonically increasing functions

whereεe ν > 0.
Hence, we can conclude thatνi can be accurately estimated by
observers ν i .
⌢

Remark 6. To facilitate the output-feedback controller designing, we

replace an unmeasurable velocityν = [u, v, r]T with its estimations ν =
⌢

⌢ ⌢ ⌢T
[u , v , r ] provided by the observer (24). Furthermore, this observer can
simultaneously guarantee greater convergence and stronger anti-
disturbance, and all errors can exactly converge to zero within a short
time, rather than existing methods can only achieve a convergence
without considering the time.

3.2. Prescribed performance control

In the presence of input saturation and unknown disturbance, tran­

sient overshoot and convergence rate of tracking errors are easily trig­
gered to violate a prescribed boundness. To avoid such serious accidents,
we introduce a prescribed performance control (PPC) to strictly guar­
antee all tracking errors converging within a predefined region that is
bounded by decaying time-function. Consider a tracking error Fig. 2. Illustration of prescribed performance.
vectorei = [xei , yei , ψ ei ]T ∈ ℝ3 , and mathematical PPC is stated as

− αi1 ρi1 < xei (t) < βi1 ρi1 , ∀t ≥ 0

− αi2 ρi1 < yei (t) < βi2 ρi1 , ∀t ≥ 0 (29)
− αi3 ρi1 < ψ ei (t) < βi3 ρi1 , ∀t ≥ 0

where0 < αij , βij < 1are prescribed scale parameters,i = 1, 2, ⋯, N, j =

1, 2, 3, andρij (t) : ℝ+ →ℝ+ is defined as predefined bounded function
( ) ( )
ρij (t) = ρij,0 − ρij,∞ exp − lij t + ρij,∞ , i = 1, 2, ⋯, N, j = 1, 2, 3 (30)

whereρij,0 , ρij,∞ andlij are strictly positive constants,ρij,0 ≥ ρij,∞ ,

⃒ ⃒
⃒ ⃒
and⃒⃒ρ̇ij (t)⃒⃒ ≤ lij (ρij,0 − ρij,∞ ). Furthermore, this statement with a case

ofρij (t) = (20 − 2)exp( − 0.005t) + 2is illustrated in Fig. 2.

Remark 7. Theρij (i = 1, 2, ⋯, N, j = 1, 2, 3)is smooth, positive,

decreasing and bounded function, satisfying lim ρij (t) = ρij,∞ > 0.
x→∞
Thelij represents the minimum convergence rate forei , andρij,∞ describes
the maximum allowable size ofei at steady-state. Furthermore, the Fig. 3. Illustration of transformation function.

6
H. Liang et al. Ocean Engineering 233 (2021) 109071

βi1 exp(exi ) − αi1 exp( − exi ) performance of the corresponding tracking errors can be guaranteed
Si1 (exi ) =
exp(exi ) + exp( − exi ) within boundness
( ) ( ) { }
( ) βi2 exp eyi − αi2 exp − eyi Ωij = ei ∈ ℝ : |xei (t)| ≤ ρi1,∞ , |yei (t)| ≤ ρi1,∞ , |ψ ei (t)| ≤ ρi3,∞ (38)
Si2 eyi = ( ) ( ) (32)
exp eyi + exp − eyi
Therefore, it proves that prescribed transient performance can be
Si3 (eψ i ) =
βi3 exp(eψ i ) − αi3 exp( − eψ i ) satisfied for allt > 0. The proof is completed here.
exp(eψ i ) + exp( − eψ i )
3.3. Filter-backstepping for virtual velocity design
wherei = 1, 2, ⋯, N, j = 1, 2, 3,αij < Sij ( ⋅) < βij .
Then, we can obtain To stabilize the transformed tracking errors (30), we employ the
() ( ) filter-backstepping technique to design the appropriate virtual velocities
xei (t) 1 αi1 + zi1
exi t = S−i11 = ln in this subsection.
ρi1 (t) 2 βi1 − zii1
The virtual velocityγi = [γ i1 , γ i2 , γi3 ]T for observed velocity ν i =
⌢
() ( )
y (t) 1 αi2 + zi2
(33) [u i , v i , r i ]T is designed by
⌢ ⌢ ⌢
eyi t = S−i21 ei = ln
ρi2 (t) 2 βi2 − zi2
⌢ ⌢
() ( ) γi1 = − li1(exi cos(ψ) li1 eyi sin(ψ i ) +)u l cos(ψ i (
i) − ( − ψ l ) + v l cos()ψ i − ψ l )
ψ (t) 1 αi3 + zi3
eψ i t = S−i31 ei = ln ⌢
ρi3 (t) 2 βi3 − zi3 − dil ϕ̇il + r l sin ϕil + ψ l − ψ i + ḋil cos ϕil + ψ l − ψ i

(39a)
whereS−ij 1 is essentially a mapping( − αij , βij ) ↦ ( − ∞, ∞). In addition,
the aforesaid transformation function is illustrated withαij = βij = 1in ⌢
γi2 = li1 exi(sin(ψ i ) −) li1 e(yi cos(ψ i ) − u l)cos(ψ i − ψ
⌢
(l ) + v l cos(ψ i)− ψ l )
the Fig. 3. ⌢
Meanwhile, taking the time derivatives of transformed error in (33), +dil ϕ̇il + r l sin ϕil + ψ l − ψ i + ḋil cos ϕil + ψ l − ψ i
and yielding (39b)
[ ] ( )
ėxi (t) =
1 1
−
1
żi1 = ϱi1
ρi1 ẋei − xei ρ̇i1
(34a) γ i3 = − li2 eψ i + ψ
̂ ai (39c)
2 αi1 + zi1 βi1 + zi1 ρi1
whereli1 andli2 are positive constants, andu l , v l and r l provided by veloc­
⌢ ⌢ ⌢
⎛ ⎞
1
[
1 1
]
ρ ẏ − yei ρ̇i2 ⎠ itey observers are estimates oful ,vl andrl , respectively. Moreover, ψ ̂ ai is an
ėyi (t) = − żi2 = ϱi2 ⎝ i2 ei (34b) unknown time-varying estimate ofψ ai provided by designing observer
2 αi2 + zi2 βi2 + zi2 ρi2
̂ ai = μi − li3 ψ ei andμ̇i = li3 (ri − μi + li3 ψ ei ), whereli3 is positive con­
as ψ
[ ] ( ) stant.
1 1 1 ρ ψ̇ − ψ ei ρ̇i3
(34c) It is noticed that derivatives of observedu l , v l and r l are unavailable
⌢ ⌢ ⌢
ėψ i (t) = − żi3 = ϱi3 i3 ei
2 αi3 + zi3 βi3 + zi3 ρi3
for output feedback according to the Assumption 7, which can lead to an
( ) explosion of complexity problem, due to the repeated differentiations of
xei yei
whereϱij = 21ρ αj +z 1
ij
− 1
βj − zij ,zi1 = ρi1 , zi2 = ρi2 , zi3 = ρi3 ,
ψ ei
and exist­ virtual velocity control signals. To avoid this problem in backstepping
ij
⃒ ⃒ procedure, DSC design is introduced, i.e. letγi pass through a first-order
ing⃒zij ⃒ ≤ 1,j = 1, 2, 3.
filter in following design
Assumption 9. The tracking errors satisfy −
ζij ς̇ij + ςij = γ ij , ςij (0) = γij (0), i = 1, 2, ⋯N, j = 1, 2, 3 (40)
αi1 ρi1 (0) < xei (0) < βi1 ρi1 (0), − αi2 ρi2 (0) < yei (0) < βi2 ρi2 (0), and −
αi3 ρi3 (0) < ψ ei (0) < βi3 ρi3 (0)whent = 0.
whereς i = [ςi1 , ςi2 , ςi3 ]T is a filtered virtual control input vector, andζ j =
Lemma 4. Consider the tracking error vectorei (t) = [xei , yei , ψ ei ]T satisfy­ [ζi1 , ζi2 , ζi3 ]T > 0is the designing time parameter. It can eliminate the
ing Assumption 9, and the transformed error (34). The prescribed transient analytic computation ofγ̇i .
performance satisfying (29) can be guaranteed for allt > 0, whenEi = Define the filter error vector as
[exi , eyi , eψ i ]T is bounded from (31) and (34).
εi = ς i − γ i (41)
Proof. Due to boundedEi = [exi , eyi , eψ i ]T , we assume that it should
satisfy with unknown constant vectors, i.e.E i andEi , such that whereεi1 = ςi1 − γ i1 ,εi2 = ςi2 − γ i2 , andεi3 = ςi3 − γi3 .
From the (40) and (41), we have the derivative ofεi as
E i < Ei < Ei , ∀t > 0 (35)
εi
ε̇i = − − γ̇i (42)
Employing the inverse transformationS−ij 1 into(35), we can obtain ζij
( ) ( ) Define the first-order new errors between filtered virtual and actual
Si1 exi ρi1 (t) < xei (t) < Si1 exi ρi1 (t), ∀t > 0 velocities as
( ) ( )
(43a)
⌢
Si2 eyi ρi2 (t) < yei (t) < Si2 eyi ρi2 (t), ∀t > 0 si1 = u i − ςi1 − πi1 tanh θi1
( ) ( )
(43b)
⌢
si2 = v i − ςi2 − πi2 tanh θi2
Si3 eψ i ρi3 (t) < ψ ei (t) < Si3 eψ i ρi3 (t), ∀t > 0 (36)
(43c)
⌢
si3 = r i − ςi3 − πi3 tanh θi3
Owing to the properties of(αij , βij ) ↦ ( − ∞, ∞), we can obtain
where ν i = [u i , v i , r i ]T can be estimated by the finite-time velocity
⌢ ⌢ ⌢ ⌢
− αi1 ρi1 (t) < xei (t) < βi1 ρi1 (t)
− αi2 ρi2 (t) < yei (t) < βi2 ρi2 (t) (37) observer. Andθi1 , θi2 andθi3 are the auxiliary control terms given by (Park
− αi3 ρi3 (t) < ψ ei (t) < βi3 ρi3 (t) et al., 2017)
Moreover, due to the positive and decreasingρij (t), prescribed

7
H. Liang et al. Ocean Engineering 233 (2021) 109071

{ )/ )}/
θ̇i1 = cosh 2 θi1 − ℓi1 θi1 − χ ui m*11,i πi1 (44a) 3.4. Neural network-based output feedback controller

{ The output-feedback control signals for leader-follower formation

⌢T
θ̇i2 = cosh 2 θi2 φvi − li4 si1 + li4 si2 − li4 si3 + W i2 Θi2 (Z) − xei sin(ψ i ) are designed as follows:

⎧ ( )
⌢T
*
⎨ τui = m11,i − φui − ℓi1 θi1 − li4 si1 − li4 si2 − W i1 Θi1 (Z) + ς̇i1 − xei sin ψ i − yei cos ψ i
⎪
( ) (46)
⎪
⎩τ = Δ − φ − ℓ θ − l s − l s − W ⌢T
*
ri ri i2 i3 i4 i2 i4 i3 i3 Θ i3 (Z) + ς̇ i3 − ψ ei m22,i

}
/
+ yei cos(ψ i ) πi2 (44b) with the adaptive law
( )
⌢ ⌢
{ / )}/ W ij = Γij Θi (Z)sij − oi W ij , j = 1, 2, 3 (47)
θ̇i3 = cosh 2 θi3 − ℓi2 θi1 − m*22,i χ ri Δ π i3 (44c)

whereoi is positive control parameter.

whereπi1 , πi2 andπi3 are positive constants. Andli4 , ℓi1 andℓi2 are positive
Substituting (46) and (47) into (45) yields
constants.
⎧
Taking the time derivatives ofsi1 , si2 andsi3 along (44), we have ⎪
⎪
⎪ ṡ = − l s − l s − W ̃ Ti1 Θi1 (Z) − xei sin ψ i − yei cos ψ i + εi1
/ ⎪
⎪
ṡi1 = φui + ϕui − ς̇i1 + ℓi1 θi1 + τui m*11,i (45a) ⎨ i1 i4 i1 i4 i2

ṡi2 = li4 si1 − li4 si2 + li3 si3 − W ̃ T Θi2 (Z) + xei sin(ψ i ) − yei cos(ψ i ) + εi2
⎪
⎪ i2
⎪
⎪ ṡ = − l s − l s − W T
⌢T ⎪
⎩ i3 i4 i2 i4 i3
̃ i3 Θi3 (Z) − ψ ei + εi3
ṡi2 = ϕvi − ς̇i2 + li4 si1 − li4 si2 + li4 si3 − W i2 Θi2 (Z) + xei sin(ψ i ) − yei cos(ψ i )
(45b) (48)
)
/ T ⌢T
ṡi3 = φri + ϕri − ς̇i3 + ℓi2 θi1 + m*22,i τri Δ (45c) whereW
̃ = W −
ij ij Wij*T ,ϕui = Wi1
*T *T
Θi1 (Z) + εi1 ,ϕvi = Wi2 Θi2 (Z) +
*T
εi2 ,ϕri = Wi3 Θi3 (Z) + εi3 .
Taking the time derivatives both sides of (22) along the solutions of
Remark 8. The auxiliary control termsθi1 andθi3 are proposed to (12) as follows

⎧ ( ) ( ) ( )
⌢ ⌢ ⌢ ⌢ ⌢
⎪
⎪ ẋ ei = u i cosψ − v i sinψ − u l cosψ − +v l sinψ − ḋ il cos ψ + ϕ + d lf r l + ϕ̇ sin ψ + ϕ
⎪
⎨ i i l l l il il l il
( ) ( ) ( )
⌢ ⌢ ⌢ ⌢ ⌢ (49)
⎪ ẏei = u i sinψ i + v i cosψ i − u l sinψ l − v l cosψ l − ḋil sin ψ l + ϕil − dll r l + ϕ̇il cos ψ l + ϕil
⎪
⎪
⎩ ⌢ ⌢
ψ̇ ei = r i − ψ ai

compensate input saturation, and θi2 is proposed to handle the under­

actuation of UUV, which can relax the assumption that the sway velocity
is passive-bounded (Wang et al., 2019a,b, 2020; Omid and Khoshnam, Substituting (43) into (49) yields

( ) ( ) ( ) ( ) ( )
⎧ ⌢ ⌢ ⌢
⎪ ẋ ei = s i1 + εi1 + γ i1 + π i1 tanhθi1 cosψ i − s i2 + εi2 + γ i2 + π i2 tanhθi2 sinψ i − u l cosψ l + v l sinψ l − ḋ il cos ψ l + ϕil + d lf r l + ϕ̇il sin ψ l + ϕil
⎪
⎪
⎨ ( ) ( ) ( ) ( ) ( )
ẏ = s + ε + γ + π tanhθ sinψ + s + ε + γ + π tanhθ cos ψ −
⌢
u sin ψ −
⌢
v cos ψ − ḋ sin ψ + ϕ − d
⌢
r + ϕ̇ cos ψ + ϕ (50)
⎪
⎪ ei i1 i1 i1 i1 i1 i i2 i2 i2 i2 i2 i l l l l il l il lf l il l il
⎪
⎩ ( )
⌢
ψ̇ ei = si3 + εi3 + γi3 + πi3 tanhθi3 − ψ ai

2018; Jia et al., 2019).

And substituting (50) and into (34) yields

8
H. Liang et al. Ocean Engineering 233 (2021) 109071

⎧
⎪
⎪ ( ) ( ) ( )
( i1 cosψ i − si2 sinψ i + εi1 cosψ i − εi2 sinψ i − li1 exi + gi1 ))
⎨ ėxi (t) = ϱi1 (s
V̇ 1i ≤ −
5 2
exi − li1 −
li1 −
5 2
eyi − li2 −
3 2 1
e + s2i1 + s2i2 + s2i3 + ε2i1
ėyi (t) = ϱi2 si1 sinψ i + si2 cosψ i + εi1 sinψ i + εi2 cosψ i − li1 eyi + gi2 (51) 2 2 2 ψi 2
⎪
⎪
⎩ ėψ i (t) = ϱi3 (si3 + εi3 − li2 eψ i + gi3 ) ∑ 3
1 1
+ ε2i2 + ε2i3 + g2
2 2 j=1 ij
wheregi1 = πi1 tanhθi1 cosψ i − πi2 tanhθi2 sinψ i − zi1 ρ̇i1 ,gi2 =
π i1 tanhθi1 sinψ i + πi2 tanhθi2 cosψ i − zi2 ρ̇i2 , gi3 = ψ̂ ai + πi3 tanhθi3 − ψ ai −
⌢
(54)
zi3 ρ̇i3 . ⃒ ⃒
⃒ ⃒
⃒ ⃒
* ⃒ ⃒ *
⃒ ⃒
Recalling the Remark 9, existing gi1 ≤ gi1 , gi2 ≤ gi2 , and gi3 ≤ gi3
⃒ ⃒ *
,
Remark 9. Due to the |tanhθi1 | ≤ 1,|tanhθi2 | ≤ 1, |sinψ i | ≤ 1,|cosψ i | ≤ 1, we can obtain
⃒ ⃒
⃒ ⃒ ⃒ ⃒ ( ) ( ) ( )
⃒zij ⃒ ≤ 1and
⃒ρ̇ (t)⃒ ≤ lij (ρ − ρ ),j = 1, 2, 3, it is implied thatgij is 5 2 5 2 3 2 1
⃒ ij ⃒ ij,0 ij,∞ V̇ 1i = ≤ − li1 − exi − li1 − eyi − li2 − e + s2i1 + s2i2 + s2i3 + ε2i1
⃒ ⃒ ⃒ ⃒ ⃒ ⃒ 2 2 2 ψi 2
* ⃒ * ⃒ *
bounded such that gi1 ≤ gi1 , gi2 ≤ gi2
⃒ ⃒ ⃒ , gi3 ⃒ ≤ gi3 .
1 1 ∑3
+ ε2i2 + ε2i3 + g*2
2 2 j=1 ij
4. Main results
(55)
In this section, we summarize the main results of this paper in the
following theorem. Step2. Substituting (50) into the time derivative ofV2i , we have
Theorem 1. Consider the underactuated UUV model (7) in leader-
V̇ 2i = xei ẋei + yei ẏei + ψ ei ψ̇ ei
following tracking system (22) and (23) with external disturbances (9) and
= xei ((si1 + εi1 + γi1 + π i1 tanhθi1 )cosψ i − (si2 + εi2 + γ i2 + πi2 tanhθi2 )sinψ i
input saturation(10), and suppose that Assumptions 1-3 and 6-8 are satis­ ( ) ( ) ( ))
fied. If the error transformed function is applied by (33), the finite-time ve­ ⌢ ⌢ ⌢
− u l cosψ l + v l sinψ l − ḋil cos ψ l + ϕil + dlf r l + ϕ̇il sin ψ l + ϕil +
locity observer is calculated by (24), the filter-backstepping velocity
controller is designed by (44), and the output feedback controller for leader- yei ((si1 + εi1 + γ i1 + πi1 tanhθi1 )sinψ i + (si2 + εi2 + γ i2 + πi2 tanhθi2 )cosψ i
( ) ( ) ( ))
following formation is updated by (46) with adaptive law given by (47), then ⌢ ⌢ ⌢
− u l sinψ l − v l cosψ l − ḋil sin ψ l + ϕil − dlf r l + ϕ̇il cos ψ l + ϕil +
the output feedback controller can guarantee the following statements.
(( ) )
⌢
ψ ei si3 + εi3 + γi3 + πi3 tanhθi3 − ψ ai
(i) Thexei , yei andψ ei can be guaranteed within the given prescribed ⃒ ⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒
performance boundρij . ≤ − li1 x2ei − li1 y2ei − li2 ψ 2ei + ⃒si1 ⃒xei ⃒ + ⃒si2 ⃒⃒xei ⃒ + ⃒si1 ⃒⃒yei ⃒ + ⃒si2 ⃒⃒yei ⃒ + ⃒si3 ⃒⃒ψ ei ⃒
(ii) All control signals of the closed-loop system remain bounded. +|εi1 ||xei | + |πi1 ||xei | + |εi2 ||xei | + |πi2 ||xei | + |εi1 ||yei | + |π i1 ||yei |
⃒ ⃒
⃒⌢ ⃒
+|εi2 ||yei | + |πi2 ||yei | + |ψ ei ||εi3 | + |π i3 ||ψ ei | + |ψ ei |⃒ψ ai ⃒
Proof: Consider the following Lyapunov function candidate for the i-
(56)
th UUV as
Vi = V1i + V2i + V3i + V4i + V5i (52) By aiding of Lemma 2, V̇ 2i can be represented as
1
whereV1i = 2ϱ1 e2xi + 2ϱ1 e2yi + 2ϱ1 e2ψ i ,V2i = 12x2ei + 12y2ei + 12ψ 2ei ,V3i = 12s2i1 + V̇ 2i ≤ − (li1 − 3)x2ei − (li1 − 3)y2ei − (li2 − 2)ψ 2ei + s2i1 + s2i2 + s2i3 + ε2i1 + ε2i2
i1 i2 i3 2
1 2 1 2 ̃ T Γ− 1 W
= 12ε2i1 + 12ε2i2 + 12ε2i3 ,V5i = 12W ̃T − 1̃
̃ i1 + 1W 1 2 1 2 1⌢2
2si2 + 2si3 ,V4i 2 i2 Γ i2 Wi2 +
2 2
i1 i1 + εi3 + π i1 + πi2 + πi3 + ψ ai
2 2 2
1̃T − 1 ̃
2Wi3 Γ i3 Wi3 (57)

Step1. Substituting (51) into the time derivative ofV1i , we have

1 1 1
V̇ 1i = exi ėxi + eyi ėyi + eψ i ėψ i
ϱi1 ϱi2 ϱi3

= − li1 e2xi − li1 e2yi − li2 e2ψ i + exi (si1 cosψ i − si2 sinψ i ) + exi (εi1 cosψ i − εi2 sinψ i ) + exi gi1

+eyi (si1 sinψ i + si2 cosψ i ) + eyi (εi1 sinψ i + εi2 cosψ i ) + eyi gi2 + eψ i si3 + eψ i εi3 + eψ i gi3 (53)
⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒
⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒
≤ − li1 e2xi − li1 e2yi − li2 e2ψ i + ⃒exi ⃒⃒si1 ⃒ + ⃒exi ⃒⃒si2 ⃒ + ⃒exi ⃒⃒εi1 ⃒ + ⃒exi ⃒⃒εi2 ⃒ + ⃒exi ⃒⃒gi1 ⃒ + ⃒eyi ⃒⃒si1 ⃒
⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒ ⃒ ⃒⃒ ⃒
+⃒eyi ⃒⃒si2 ⃒ + ⃒eyi ⃒⃒εi1 ⃒ + ⃒eyi ⃒⃒εi2 ⃒ + ⃒eyi ⃒⃒gi2 ⃒ + ⃒eψ i ⃒⃒si3 ⃒ + ⃒eψ i ⃒⃒εi3 ⃒ + ⃒eψ i ⃒⃒gi3 ⃒

⃒ ⃒
⃒ ⃒
Applying the Assumption 8, existing⃒⃒ψ̇ ai ⃒⃒ ≤ ψ *ai , such that (57) can be
wherej = 1, 2, 3. represented as
By aiding of the Lemma 2, V̇ 1i can be represented as
1
V̇ 2i ≤ − (li1 − 3)x2ei − (li1 − 3)y2ei − (li2 − 2)ψ 2ei + s2i1 + s2i2 + s2i3 + ε2i1 + ε2i2
2
1 2 2 2 1 2 1 *2
+ εi3 + π i1 + πi2 + πi3 + ψ ai
2 2 2
(58)

9
H. Liang et al. Ocean Engineering 233 (2021) 109071

Step3. Substituting (48) into the time derivative ofV3i , We have Step4. Substituting (42) into the time derivative ofV4i , we have

( ( ) ) ( ( )
V̇ 3i = si1 ̃ T Θi1 Z − xei sin ψ i − yei cos ψ i + εi1 + si2 li4 si1 − li4 si2 + li3 si3 − W
− li4 si1 − li4 si2 − W ̃ T Θi2 Z
i1 i2
( ) ( ) ) ( ( ) )
T
+xei sin ψ i − yei cos ψ i + εi2 + si3 − li4 si2 − li4 si3 − W ̃ i3 Θi3 Z − ψ ei + εi3
T
̃ Θi1 (Z) − si1 xei sin ψ i − si1 yei cos ψ i + si1 εi1 − si2 W
= − li4 s2i1 − li4 s2i2 − li4 s2i3 − si1 W ̃ Θi2 (Z) T (59)
i1 i2

+si2 xei sin(ψ i ) − si2 yei cos(ψ i ) + si2 εi2 − ̃ Ti3 Θi3 (Z) −
si3 W si3 ψ ei + si3 εi3
≤− li4 s2i1 − li4 s2i2 − li4 s2i3 ̃ Ti1 Θi1 (Z) −
− si1 W ̃ Ti2 Θi2 (Z) −
si2 W ̃ Ti3 Θi3 (Z)
si3 W
+|si1 ||xei | + |si1 ||yei | + |si2 ||xei | + |si2 ||yei | + |si3 ||ψ ei | + |si1 ||εi1 | + |si2 ||εi2 | + |si3 ||εi3 |

( ) ( ) ( )
By aiding of Lemma 2, V̇ 3i can be represented as
V̇ 4i = εi1 ς̇i1 − γ̇i1 + εi2 ς̇i2 − γ̇i2 + εi3 ς̇i3 − γ̇i3
( ) ( ) ∑3
3 2 3 2 ̃ Tij Θi (Z) + x2ei
V̇ 3i ≤ − li4 − si1 − li4 − si2 − (li4 − 1)s2i3 − sij W
2 2 j=1
∑
3
ε2ij
=−
ζij
+ εi1 γ̇ i1 + εi2 γ̇i2 + εi3 γ̇ i3 (62)
1 1∑ 3
j=1
+ y2ei + ψ 2ei + ε2
2 2 j=1 ij ⃒ ⃒ ⃒ ⃒ ⃒ ⃒
∑
3
ε2ij ⃒ ⃒ ⃒ ⃒ ⃒ ⃒
(60) ≤− + |εi1 |⃒γ̇ i1 ⃒ + |εi2 |⃒γ̇i2 ⃒ + |εi3 |⃒γ̇i3 ⃒
j=1
ζij
Due to the existing inequality|ε| ≤ ε* , V̇ 3i can be rewritten as
By aiding of Lemma 2, V̇ 4i can be represented as
( ) ( ) ∑3
3 2 3 2 ̃ Tij Θi (Z) + x2ei ( )
V̇ 3i ≤ − li4 − si1 − li4 − s − (li4 − 1)s2i3 − sij W ∑ 3 2
2 2 i2 1 γ̇ij 2 3
j=1 V̇ 4i ≤ − − ε + (63)
j=1
ζij 2 ij 2
1 1∑ 3
+ y2ei + ψ 2ei + ε*2
2 2 j=1 ij It is noted that the derivative ofγi = [γ i1 , γi2 , γi3 ]T given by (41) can be
(61) represented asϒ i ( ⋅) = [ϒ i1 ( ⋅),ϒ i2 ( ⋅),ϒ i3 ( ⋅)] = γ̇ i ,ϒ i1 = γ̇ i1 (eix ,eiy ,si1 ,xei ,yei ,
si2 ,si3 , εi1 , εi2 , εi3 , η̇l ,η̈l ,dil , ḋil , d̈il ,ϕil , ϕ̇il , ϕ̈il ),ϒ i2 = γ̇i2 (eix ,eiy ,xei ,yei ,si1 ,si2 ,si3 , εi1 ,
εi2 , εi3 , η̇l , η̈l , dil , ḋil , d̈il , ϕil , ϕ̇il , ϕ̈il ),ϒ i3 = γ̇i3 (eψ i , ψ ei , si3 , εi3 , ψ ai ), whereϒ i1 ( ⋅),
⌢

ϒ i2 ( ⋅), andϒ i3 ( ⋅)are all continuous functions (Swaroop et al., 2000).

Furthermore, applying the Assumptions 6 and 7, ||ϒ i ( ⋅)||has the

Fig. 4. Case1: straight line.

Fig. 5. Case2: curve line.

10
H. Liang et al. Ocean Engineering 233 (2021) 109071

⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2

⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒
maximumϒ iM on Π i1 × Π i2 ,whereΠ i1 = {⃒⃒⃒⃒η̇l ⃒⃒⃒⃒ + ⃒⃒⃒⃒η̈l ⃒⃒⃒⃒ + ⃒⃒⃒⃒dil ⃒⃒⃒⃒ +
⌢ ⌢ ⌢
̃ Ti1 Γ −i11 W i1 + W
V̇ 5i = W ̃ Ti2 Γ−i21 W i2 + W
̃ Ti3 Γ −i31 W i3
⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ⃒⃒ ⃒⃒2 ∑
3 ∑
3
⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ T⌢
(67)
T
⃒⃒ḋil ⃒⃒ + ⃒⃒d̈il ⃒⃒ + ⃒⃒ϕil ⃒⃒ + ⃒⃒ϕ̇il ⃒⃒ + ⃒⃒ϕ̈il ⃒⃒ ≤ А0 , ∀А0 > 0},Π i2 = {(eix , = ̃ Θi (Z) −
sij W ̃ W ij
oi W
⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ⃒⃒ ij ij
j=1 j=1
eiy , eψ i , xei , yei , ψ ei , si1 , si2 , si3 , εi1 , εi2 , εi3 ) : V ≤ В0 , ∀В0 > 0}. Thus, we can
obtain By aiding of the Assumptions 4 and 5, we can obtain
( ) ∑
3 ( ) ∑3 ⃒⃒ ⃒⃒2 ∑ 3 ⃒⃒ ⃒⃒2
⃒⃒ ⃒⃒ ⃒⃒ ̃ ⃒⃒
̃ Tij Θi Z + oi ⃒⃒Wij* ⃒⃒ − (68)
∑ 3 2
1 ϒ ij,M 2 3 V̇ 5i ≤ sij W oi ⃒⃒W ij ⃒⃒
V̇ 4i ≤ − − εij + (64)
j=1
ζij 2 2 j=1 j=1 j=1

Then, substituting (55), (58), (61), (65) and (68) into the time de­
Choosingζ1i1 = 12ϒ 2i1,M + ζ*i1 , ζ1i2 = 12ϒ 2i2,M + ζ*i2 , ζ1i3 = 12ϒ 2i3,M + ζ*i3 , (64) can
rivative ofVi , we have
be rewritten as

( )( ) ( ) ( )( ) ( )
5 3 2 5 2
Vi ≤ − li1 −e2xi + e2yi − li2 − eψ i − li1 − 4 x2ei + y2ei − li2 − ψ
2 2 2 ei
( )( ) ( ) ( ) ( ) ( )
7 1 2
− li4 − s2i1 + s2i2 − li4 − 2 s2i3 − ζ*i1 − 1 ε2i1 − ζ*i2 − 1 ε2i2 − ζ*i3 − ε (69)
2 2 i3

∑ ⃒⃒ ⃒ ⃒⃒ ⃒
3 ⃒⃒ ⃒ 2 1 ∑ 3
1∑3
1 2 1 *2 ∑ 3
⃒⃒ ⃒ 3
− oi ⃒⃒⃒⃒W
̃ ij ⃒| +
⃒ g*2
ij + ε*2 2 2
ij + π i1 + π i2 + π i3 + ψ ai + oi ⃒⃒Wij* ⃒|2 +
j=1
2 2 j=1 j=1
2 2 2j=1

∑
3
3 Now, let us consider the Lyapunov function candidate for leader-
V̇ 4i ≤ − ε2ij ζ*ij + (65) following formation as follows
j=1
2
)
∑n

whereζ*ij is positive constant. V= (V1i + V2i + V3i + V4i + V5i (70)

i=1

Step5. Substituting adaptive law (47) into the time derivative ofV5i , Substituting (69) into the time derivative of (70), we have
we can obtain V̇ ≤ − 2μV + C (71)

where

{ ( )}
5 3 5 7 * * * 1 oi
μ = min li1 − , li2 − , li1 − 4, li2 − , li4 − , li4 − 2, ζi1 − 1, ζi2 − 1, ζi3 − , min ( )
2 2 2 2 2 λmax Γ −ij 1

Fig. 7. Tracking erroryei with Case1.

Fig. 6. Tracking errorxei with Case1.
11
H. Liang et al. Ocean Engineering 233 (2021) 109071

Fig. 8. Tracking errorψ ei with Case1. Fig. 11. Tracking errorψ ei with Case2.

Fig. 9. Tracking errorxei with Case2. Fig. 12. Velocity observer errors with Case1.

Fig. 10. Tracking erroryei with Case2. Fig. 13. Velocity observer errors with Case2.

12
H. Liang et al. Ocean Engineering 233 (2021) 109071

( ) ( )
1∑ n ∑
3 ∑3 ∑3 ⃒⃒ ⃒ ∑n
1 2 1 *2
⃒⃒ * ⃒ 2
C= g*2 + ε*2
+2 oi ⃒⃒W ⃒| + π 2
+ π 2
+ π + ψ
2 i=1 j=1 ij j=1
ij
j=1
ij
i=1
i1 i2
2 i3 2 ai
3n
+
2

with the designing parameters satisfying

5 7 1
li1 > 4, li2 > , li4 > , ζ*i1 > 1, ζ*i2 > 1, ζ*i3 > (72)
2 2 2

whereλmax denotes the maximum eigenvalues.

Integrating (71) over [0, t] yields
( )
C C − 2μt
0 ≤ V(t) ≤ + V(0) − e (73)
2μ 2μ
Fig. 14. Filter errors with Case1.
From the inequality (73), it is clearly indicated thatV(t)is uniformly
ultimately bounded for all V(0) ≤ В0 on[0, t], and by aiding of the
Proposition 1 thent→∞. Therefore, we can draw a conclusion that
tracking errorsxei , yei , ψ ei , transformed errorseix , eiy , eψ i , si1 , si2 , si3 , filtered
errors εi1 , εi2 , εi3 and RBFNN weighting errorsW ̃ i1 , W ̃ i2 , W ̃ i3 are all boun­
ded. Furthermore, it can be concluded from (29) and (33) that tracking
errorsxei , yei , ψ ei satisfy the prescribed transient performance without
violating anyρij for allt > 0. Thus, all error signals of closed-loop system
remain bounded and converge withinρij . The proof is completed.

5. Simulation experiment

In this section, numerical simulations are provided to demonstrate

performance of the designed control scheme for underactuated UUVs.
Consider a formation system consisting of two followers and one leader.
The dynamics of each underactuated UUV is adopted from (Chen et al.,
2020), and the nominal parameters are given as follows: (i)m*11i =
141.85,m*22i = 191.75,m*33i = 15.6, m*12i = m*21i = m*13i = m*31i = 0;
Fig. 15. Filter errors with Case2.
(ii)c*13i (v, r) = − 191.75v − 5.7r ,c*23i (u) = 141.85u,c*31i (v, r) =
191.75v + 5.7r,c*32i (u) = − 141.85u,andc*11i = c*12i = c*21i = c*22i =

Fig. 16. Norm of NN weights. (a) Follower1 with Case1. (b) Follower1 with Case2. (c) Follower2 with Case1.(d) Follower2 with Case2.

13
H. Liang et al. Ocean Engineering 233 (2021) 109071

Fig. 17. Control inputs. (a) τu with Case1. (b) τu with Case2. (c) τr with Case1. (d)τr with Case2.

Fig. 18. Results of tracking errors.xei

Fig. 19. Results of tracking errors.yei

c*33i = 0; (iii)d*11i (u) = 45.6 + 67.26|u| + 10u2 , d*22i (v, r) = 29.54 +

73.85|v| + 15|r|,d*23i (v, r) = − 2.5 + 2|v| + 10.71|r|,d*32i (v, r) = − 2.4 − For the finite-time velocities observer, the parameters are selected
13|v| − 0.2|r|,d*33i (v, r) = 5.59 + 10.71|r| − 0.07|r|, and d*12i = d*21i = asϖ 1,i = 0.2,ϖ 2,i = 0.02, and lui = lvi = lri = 200. For the predefined
d*13i = d*31i = 0. For the leader-follower formation, the predefined performance functions, following parameters are selected asρi1,0 =
desired distancedif and angleϕif of two followers are set as (5m, ρi2,0 = 20,ρi1,∞ = ρi2,∞ = 2,ρi3,0 = 3,ρi3,∞ = π/9,li1 = li2 = 0.005,
2π /3rad)and (5m, 4π /3rad), respectively. In addition, all UUVs are andli3 = 0.02. The parameters of virtual filter are chosen asζi1 = ζi2 =
affected by time-varying disturbance do (t)in the earth frame as ζi3 = 0.2. To estimate uncertain terms, RBF networks are employed.
⌢ ⌢
⎡
− 1 + 0.25sin(0.08t + 0.5π) − 0.05sin(0.04t + 0.125π)
⎤ There are 256 neuron nodes for eachΘi (Z ), where Z =
T
do (t) = − 1 + 0.15sin(0.04t + 0.5π) − 0.05sin(0.04t + 0.125π) ⎦
⎣ (74) [ui vi ri ςi1 ςi2 ςi3 eix eiy eiψ εi1 εi2 εi3 ] is the input vector of output feedback
0.1sin(0.5t + 0.2π) controller. The gain matrices are defined asΓ i1 = Γ i2 = Γi3 = 20I12×12 ,
and the initial weights are randomly given in the range of[− 1, 1]. The
such thatw(t)acting on UUV in the body frame can be represented centersci are evenly placed in [− 1.5,1.5] × [− 1,1] × [− 1,1] × [− 1.5,1.5] ×
asw(t) = J(ψ )do (t). [− 1, 1] × [− 1, 1] × [− 1, 1] × [− 1, 1] × [− 1, 1] × [− 1, 1] × [− 1, 1] × [− 1, 1].

14
H. Liang et al. Ocean Engineering 233 (2021) 109071

Table 1
Performance comparisons of different methods.
Method IAE ITAE ISA(⋅109 ) ISA(⋅109 )
∫t ∫t ∫t 2 ∫t 2
0 ||ei ||dτ 0 t||ei ||dτ 0 τui dτ 0 τri dτ

ACFC 132.26 1125.85 2.5209 2.5198

Proposed 102.97 581.23 2.3252 2.3188

capture the lumped nonlinear terms composed of model uncertainties

and external disturbance in (7). It is also evident that the formulated NN
adaptive controller can provide optimal solutions for surge force and
yaw moment within the finite time. Finally, Fig. 17 demonstrates cor­
responding control inputs including surge forceτu and yaw momentτr .
Therefore, from all above results, it is implied that the proposed scheme
is effective in achieving desired leader-following formation in spite of
the imposed uncertainties and disturbances.

Fig. 20. Results of tracking error.ψ ei 5.2. Comparative results

The widths are chosen asσ i = 5. The control parameters and gains of To further demonstrate advantages of our approach, comparison
adaptive NN output feedback are chosen asπi1 = 20, πi2 = 30, π i3 = 20, simulations are conducted by using adaptive cooperative formation
controller (ACFC) (Lu et al., 2018). For comparison, simulations studies
li1 = 5,li2 = 3,li3 = 2,li4 = 4,ζ*i1 = 2,ζ*i2 = 2,ζ*i3 = 1, θi1 (0) = 0,
are conducted between two types of controllers with the same control
θi2 (0) = 0,θi3 (0) = 0.
parameters and initial states as previous two cases. Simulation results
are shown in Figs. 18–20, it is observed that the proposed scheme can
5.1. Closed-loop performance provide the satisfied transient and steady performance of the tracking
errors, in which performance constraints are never violated the pre­
In order to demonstrate the arbitrary time-varying leader-follower scribed bounds. However, the controller in (Lu et al., 2018) clearly fails
formation achieved by the proposed control scheme, we carry out the to satisfy the prescribed bounds and provides low convergence.
following two cases, i.e., straight line and curve line, in the closed-loop Furthermore, detailed comparisons of these two methods are summa­
performance. rized in Table 1, in which integral absolute error (IAE, defined
∫t ∫t
Case1. The trajectory of the leader is defined as straight line using the as 0 ||ei ||dτ), integral time absolute error (ITAE, defined as 0 t||ei ||dτ),
∫t 2 ∫t 2
following velocities: 0 ≤ t,ul = 1.2m/s,vl = 0m/s,rl = 0rad/s.The and the integral of square error (ISA, defined as 0 τui dτand 0 τri dτ) are
initial condition of leader is defined as ηl (0) = [0m, 0m, 0rad]T . The introduced to evaluate transient performance, steady-state performance,
initial conditions of two followers are defined as η1 = − and control energy, whereei = [exi , eyi , eψ i ]T . From Table 1, it is evident
[10m, − 22.5m, 0rad] , T
ν1 = [0m/s, 0m/s, 0rad/s] T
and η2 = that the proposed method achieves a better dynamic performance in
[− 10m, − 2.5m, 0rad]T , ν2 = [0m/s, 0m/s, 0rad/s]T . comparison to the ACFC as it guarantees the lower IAE and ITAE with
the smaller control energy consumption ISA. Consequently, by incor­
Case2. The trajectory of the leader is defined as curve line using the porating finite-time velocity observer, prescribed transient perfor­
following velocities: (i) 0 ≤ t < 10,ul = 1.2m/s,vl = 0m/s,rl = 0rad/s; mance, and filter-backstepping technique in output-feedback controller
(ii) 10 ≤ t < 20, ul = sin(πt /20)m /s, vl = 0m/s, rl = 0rad/s; (iii) 20 ≤ design, we conclude that the proposed scheme provides the better per­
t < 50, ul = 1m/s, vl = 0m/s,rl = − 0.1cos(πt /20)rad /s, and (iv)50 ≤ formances in the presence of model uncertainties, unmeasured velocity,
t < 100,ul = 1m/s,vl = 0m/s,rl = 0rad/s. The initial condition of leader external disturbance, and prescribed performance.
is defined as ηl (0) = [0m, 0m, 0rad]T . The initial conditions of two fol­
Remark 10. The control parameters have some effects on stability and
lowers are defined as η1 = [10m, − 10m, 0rad]T , ν1 =
T
performance. (i)li1 andli2 are associated with the convergence of trans­
[0m/s, 0m/s, 0rad/s] T
and η2 = [10m, − 10m, 0rad] , ν2 =
formed errorsEi = [exi , eyi , eψ i ]T , li3 is related to estimating a time-
[0m/s, 0m/s, 0rad/s]T , respectively. varying approach angle for tracking straight-path and curve-path,
The simulation results are illustrated in Figs. 4–18. Figs. 4–6 show andli4 should be selected sufficient large to ensure the close-loop sys­
that the leader trajectory can be tracked under the proposed output- tem. (ii)πi1 andπi3 are related to compensated effects of saturated inputs,
feedback controller, which indicates that the leader-follower forma­ andπi2 influences the underactuated problem. (iii) ϖ 1,i ,ϖ 2,i ,lui ,
tion is well established in both straight line and curve line. As depitcted lvi andlri are parameters for the finite-time velocity observer, where­
in Figs. 6–12, output tracking errors can not only satisfy the prescribed ϖ 1,i andϖ 2,i should be selected as small as possible, whilelui , lvi andlri are
performance bounds, but also converge to a small value close to zero chosen larger enough, thereby achieving a small estimation error.
under model uncertainties and external disturbances. It is observed that
the proposed scheme is robust enough against uncertainties and dis­ 6. Conclusion
turbances. Besides, Figs. 12 and 13 exhibit velocity observation errors,
which converge to the neighborhood of origin within a short time This paper presented an adaptive output-feedback trajectory
period, thereby verifying the fact that the designed observer is effective. tracking formation control scheme for underactuated UUVs with pre­
Figs. 14–16 show the filter errors for virtual velocity design of back­ scribed transient performance in the presence of model uncertainties
stepping, all of which remain bounded with good convergence. and external disturbances. To reflect realistic situation, we considered
Furthermore, Fig. 16 shows norm of neural weights provided by adap­ that mass and damping matrices were off-diagonal and only partial in­
tive update law (47) for two cases, all of which are bounded. The fast formation about hydrodynamics was known. To design controller with
responses during the initial phase wtih transitory oscillations and finite- unknown velocites, a finite-time velocity observer was exactly con­
time convergent NNs are accounted for by its stronger capability to structed by aiding of only position information. RBF-NNs were used to

15
H. Liang et al. Ocean Engineering 233 (2021) 109071

idenfity the lumped nonlinear terms composed of model uncertainies Li, J., Du, J.L., Hu, X., 2020. Robust adaptive prescribed performance control for
dynamic positioning of ships under unknown disturbances and input constraints.
and external disturbances. The prescribed transient performance func­
Ocean. Eng. 206, 107254.
tions were introduced to guarantee that all tracking errors were bounded Liang, H.T., Fu, Y.F., Kang, F.J., Gao, J., Qiang, N., 2020a. A behavior-driven
within prescribed constraints in transformed coordinate. Stability coordination control framework for target hunting by UUV intelligent swarm. IEEE
analysis showed that uniformly ultimately boundness of all close-loop Access 8, 4838–4859.
Liang, X., Qu, X., Wang, N., Zhang, R., Li, Y., 2020b. Three-dimensional trajectory
signals were guraranteed. Simulation results demonstrated out­ tracking of an underactuated AUV based on fuzzy dynamic surface control. IET
performing performance of the proposed control scheme. For future Intell. Transp. Syst. 14, 364–370.
work, it is of interest to investigate avoiding-obstacle and preserving- Lionel, L., Didik, S., 2007. Nonlinear path-following control of an AUV. Ocean. Eng. 34,
1734–1744.
connection with prescribed performance in leader-following formation Liu, H., Chen, G., 2020. Robust trajectory tracking control of marine surface vessels with
control of underactuated UUVs. uncertain disturbances and input saturation. Nonlinear Dynam. 100, 3513–3528.
Liu, H., Lyu, Y.F., Lewis, F.L., Wan, Y., 2019. Robust time-varying formation control for
multiple underwater vehicles subject to nonlinearities and uncertainties. Int. J.
Declaration of competing interest Robust Nonlinear Control 29, 2712–2724.
Liu, X., Zhang, M.J., Wang, S.M., 2020. Adaptive region tracking control with prescribed
transient performance for autonomous underwater vehicle with thruster fault.
The authors declare that they have no known competing financial Ocean. Eng. 196, 106804.
interests or personal relationships that could have appeared to influence Lu, Y., Zhang, G.Q., Sun, Z.J., Zhang, W.D., 2018. Adaptive cooperative formation
the work reported in this paper. control of autonomous surface vessels with uncertain dynamics and external
disturbances. Ocean. Eng. 167, 36–44.
Makavita, C., Jayasinghe, S., Nguyen, H., Ranmuthugala, D., 2019. Experimental study
Acknowledgments of command governor adaptive control for unmanned underwater vehicles. IEEE
Trans. Contr. Syst. Technol. 27, 332–345.
Omid, E., Khoshnam, S., 2018. Neural network-based target tracking control of
The authors acknowledge the financial support from the National underactuated autonomous underwater vehicles with a prescribed performance.
Natural Science Foundation of China under Grant 11404205, Natural Ocean. Eng. 167, 239–256.
ScienceFoundation of Shaanxi under Grant 2019JQ-026 and Funda­ Omid, E., Khoshnam, S., 2020. A robust neural network approximation-based prescribed
performance output-feedback controller for autonomous underwater vehicles with
mental Research Funds for Central Universities under Grant actuators saturation. Eng. Appl. Artif. Intell. 88, 103382.
GK202103014 and GK201803023. And authors would like to thank all Park, B.S., 2015. Adaptive formation control of underactuated autonomous underwater
reviewers and editors who provided extensive invaluable comments and vehicles. Ocean. Eng. 96, 1–7.
Park, B.S., Kwon, J.W., Kim, H.k., 2017. Neural network-based output feedback control
suggestions.
for reference tracking of underactuated surface vessels. Automatica 77, 353–359.
Peng, Z.H., Wang, D., Hu, X.J., 2011. Robust adaptive formation control of autonomous
References surface vehicles with uncertain dynamics. IET Control 5, 1378–1387. Theory A.
Peng, Z., Wang, D., Wang, J., 2017a. Predictor-based neural dynamic surface control for
uncertain nonlinear systems in strict-feedback form. IEEE Trans. Neur. Net. Lear.
Antonelli, G., Caccavale, F., Chiaverini, S., Fusco, G., 2003. A novel adaptive control law
Syst. 28, 2156–2167.
for underwater vehicles. IEEE Trans. Contr. Syst. Technol. 11, 221–232.
Peng, Z., Wang, J., Wang, D., 2017b. Distributed containment maneuvering of multiple
Balch, T., Arkin, R.C., 1998. Behavior-based formation control for multirobot teams.
marine vessels via neurodynamics-based output feedback. IEEE Trans. Ind. Electron.
IEEE Trans. Robot. Autom. 14, 926–939.
64, 3831–3839.
Beard, R.W., Lawton, J., Hadaegh, F.Y., 2001. A coordination architecture for spacecraft
Peng, Z., Wang, J., Han, Q., 2019a. Path-following control of autonomous underwater
formation control. IEEE Trans. Contr. Syst. Technol. 9, 777–790.
vehicles subject to velocity and input constraints via neurodynamic optimization.
Bechlioulis, C.P., Rovithakis, G.A., 2008. Robust adaptive control of feedback
IEEE Trans. Ind. Electron. 66, 8724–8732.
linearizable MIMO nonlinear systems with prescribed performance. IEEE Trans.
Peng, Z., Wang, J., Wang, J., 2019b. Constrained control of autonomous underwater
Automat. Contr. 53, 2090–2099.
vehicles based on command optimization and disturbance estimation. IEEE Trans.
Bechlioulis, C.P., Karras, G.C., Heshmati-Alamdari, S., Kyriakopoulos, K.J., 2017.
Ind. Electron. 66, 3627–3635.
Trajectory tracking with prescribed performance for underactuated underwater
Qi, X., Cai, Z., 2018. Three-dimensional formation control based on nonlinear small gain
vehicles under model uncertainties and external disturbances. IEEE Trans. Contr.
method for multiple underactuated underwater vehicles. Ocean. Eng. 151, 105–114.
Syst. Technol. 25, 429–440.
Qiao, L., Zhang, W., 2019. Double-loop integral terminal sliding mode tracking control
Chen, L., Cui, R., Yang, C., Yan, W., 2020. Adaptive neural network control of
for UUVs with adaptive dynamic compensation of uncertainties and disturbances.
underactuated surface vessels with guaranteed transient performance: theory and
IEEE J. Ocean. Eng. 44, 29–53.
experimental results. IEEE Trans. Ind. Electron. 67, 4024–4035.
Qiao, L., Zhang, W., 2020. Trajectory tracking control of AUVs via adaptive fast
Cui, R.X., Ge, S.Z.S., How, B.V.E., Choo, Y.S., 2010. Leader-follower formation control of
nonsingular integral terminal sliding mode control. IEEE Trans. Ind. Electron. 16,
underactuated autonomous underwater vehicles. Ocean. Eng. 37, 1491–1502.
1248–1258.
Dai, S., He, S., Lin, H., Wang, C., 2018. platoon formation control with prescribed
Qu, X.R., Liang, X., Hou, Y.H., Li, Y., Zhang, R.B., 2020. Path-following control of
performance guarantees for USVs. IEEE Trans. Ind. Electron. 65, 4237–4246.
unmanned surface vehicles with unknown dynamics and unmeasured velocities.
Dennis, B., Mohamed, A.M., Claudio, P., Kristin, Y.P., 2019. Observer based path
J. Mar. Sci. Technol. https://doi.org/10.1007/s00773-020-00744-3.
following for underactuated marine vessels in the presence of ocean currents: a
Ryan, E.P., 1991. A universal adaptive stabilizer for a class of nonlinear systems. Syst.
global approach. Automatica 100, 123–134.
Contr. Lett. 16, 209–218.
Do, K.D., Pan, J., 2005. Global tracking control of underactuated ships with nonzero off-
Saback, R., Conceicao, A., Santos, T., Albiez, J., Reis, M., 2020. Nonlinear model
diagonal terms in their system matrices. Automatica 41, 87–95.
predictive control applied to an autonomous underwater vehicle. IEEE J. Ocean. Eng.
Do, K.D., Pan, J., 2009. Control of Ships and Underwater Vehicles: Design for
45, 799–812.
Underactuated and Nonlinear Marine Systems. Springer Science & Business Media,
Sahoo, A., Dwivedy, S., Robi, P., 2019. Advancements in the field of autonomous
London.
underwater vehicle. Ocean. Eng. 181, 145–160.
Fiorelli, E., Leonard, N.E., Bhatta, P., Paley, D.A., Bachmayer, R., Fratantoni, D.M., 2006.
Sahu, B.K., Subudhi, B., 2018. Flocking control of multiple AUVs based on fuzzy
Multi-AUV control and adaptive sampling in Monterey Bay. IEEE J. Ocean. Eng. 31,
potential functions. IEEE Trans. Fuzzy Syst. 26, 2539–2551.
935–948.
Sarda, E.I., Dhanak, M.R., 2019. Launch and recovery of an autonomous underwater
Fossen, T.I., 1994. Guidance and Control of Ocean Vehicles. Wiley, Chichester, U.K.
vehicle from a station-keeping unmanned surface vehicle. IEEE J. Ocean. Eng. 44,
Fossen, T.I., 2011. Handbook of Marine Craft Hydrodynamics and Motion Control. Wiley,
290–299.
New York.
Shtessel, Y., Shkolnikov, I., Levant, A., 2007. Smooth second-order sliding modes: missile
Ge, S.S., Wang, C., 2002. Direct adaptive NN control of a class of nonlinear systems. IEEE
guidance application. Automatica 43, 1470–1476.
Trans. Neural Network. 13, 214–221.
Sung, J.Y., Bong, S.P., 2015. Guaranteed performance design for distributed bounded
Jia, Z.H., Hu, Z.H., Zhang, W.D., 2019. Adaptive output-feedback control with prescribed
containment control of networked uncertain underactuated surface vessels. Ocean.
performance for trajectory tracking of underactuated surface vessels. ISA Trans. 95,
Eng. 354, 1584–1602.
18–26.
Swaroop, D., Hedrick, J.K., Yip, P.P., Gerdes, J.C., 2000. Dynamic surface control for a
Khalil, H.K., 2002. Nonlinear Systems, 3rd. Prentice Hall, New Jewsey.
class of nonlinear systems. IEEE Trans. Automat. Contr. 45, 1893–1899.
Li, S., Wang, X., Zhang, L., 2015. Finite-time output feedback tracking control for
Wang, Z., Lu, R., Gao, F., Liu, D., 2017. An indirect data-driven method for trajectory
autonomous underwater vehicles. IEEE J. Ocean. Eng. 40, 727–751.
tracking control of a class of nonlinear discrete-time systems. IEEE Trans. Ind.
Li, H., Xie, P., Yan, W., 2017. Receding horizon formation tracking control of constrained
Electron. 64, 4121–4129.
underactuated autonomous underwater vehicles. IEEE Trans. Ind. Electron. 64,
Wang, J.Q., Wang, C., Wei, Y.J., Zhang, C.J., 2019a. Command filter based adaptive
5004–5013.
neural trajectory tracking control of an underactuated underwater vehicle in three
Li, J., Du, J.L., Chang, W.J., 2019. Robust time-varying formation control for
dimensional space. Ocean. Eng. 180, 175–186.
underactuated autonomous underwater vehicles with disturbances under input
saturation. Ocean. Eng. 179, 180–188.

16
H. Liang et al. Ocean Engineering 233 (2021) 109071

Wang, H.M., Dai, S., Luo, F., 2019b. Leader-follower formation control of USVs with Wang, Y.H., Wang, H.B., Li, M.Y., Wang, D.S., Fu, M.Y., 2021. Adaptive fuzzy controller
prescribed performance and collision avoidance. IEEE Trans. Ind. Electron. 15, design for dynamic positioning ship integrating prescribed performance. Ocean. Eng.
572–581. 219, 107956.
Wang, J.Q., Wang, C., Wei, Y.J., Zhang, C.J., 2020a. Sliding mode based neural adaptive Wen, G., Chen, C.L.P., Feng, J., Zhou, N., 2018. Optimized multi-agent formation control
formation control of underactuated AUVs with leader-follower strategy. Appl. Ocean based on an identifier-actor-critic reinforcement learning algorithm. IEEE Trans.
Res. 94, 101971. Fuzzy Syst. 26, 2719–2731.
Wang, J.Q., Wang, C., Wei, Y.J., Zhang, C.J., 2020b. Filter-backstepping based neural Wu, Y., Low, K.H., Lv, C., 2020. Cooperative path planning for heterogeneous unmanned
adaptive formation control of leader-following multiple AUVs in three dimensional vehicles in a search-and-track mission aiming at an underwater target. IEEE Trans.
space. Ocean. Eng. 201, 107150. Veh. Technol. 69, 6782–6787.

17