You are on page 1of 11

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS 1

Neural Network-Based Adaptive Asymptotic


Prescribed Performance Tracking Control of
Hydraulic Manipulators
Wenxiang Deng , Hang Zhou, Jin Zhou, and Jianyong Yao , Member, IEEE

Abstract—Hydraulic manipulators are extremely complicated strong flexibility, rapid response, high power-to-weight ratios,
systems due to the highly nonlinear characteristics, strong and large torque/force outputs. During the past decades,
coupling among multiple joints, and heavy modeling uncertain- high-performance tracking control of robot manipulators has
ties, which greatly complicate the high-performance tracking
controller development compared with conventional manipu- attracted wide attention and great efforts have been made to
lators driven by electrical motors. To overcome the above deal with this challenging issue with many advanced control
obstacles, this article proposes a novel radial basis function methods proposed, such as the sliding-mode control [1], [2],
neural network (RBFNN)-based adaptive asymptotic prescribed fuzzy control [3], adaptive control [4], [5], [31], active dis-
performance controller for n-degrees of freedom (DOF) hydraulic turbance rejection control [6], [7], and neural network (NN)
manipulators. First, the entire manipulator system model con-
taining actuator dynamics is derived. Then, the adaptive asymp- control [8]–[10].
totic prescribed performance controller is synthesized based on However, most of the existing works ignore the actuator
the backstepping framework, in which the RBFNN is employed dynamics in the modeling of robot manipulator systems so
to estimate the unknown joint coupling dynamics, while the as to simplify the entire system model into a simple second-
RBFNN reconstruction error and uncertainties of the actuator order system. Whereas, the actuator dynamics, especially the
dynamics are handled by the robust integral of the sign of the
error (RISE) feedbacks. Meanwhile, a prescribed performance highly nonlinear hydraulic actuator dynamics, have signifi-
function (PPF), which characterizes both the certain transient cant impact on the tracking control performance. Therefore,
and steady-state performance, is innovatively incorporated into it is necessary to consider them in the controller develop-
the control design to restrict the joint tracking errors. The the- ment for tracking performance improvement. In comparison
oretical analysis proves that the proposed control strategy can to robot manipulators driven by electrical motors, the accu-
achieve a prescribed tracking performance and the asymptotic
stability of the whole closed-loop system can also be ensured. rate motion control of hydraulic manipulators is much more
Finally, comparative simulations are conducted to verify the challenging owing to the following reasons. First, the use of
validity of the proposed controller. hydraulic cylinders to drive the joints forms a complicated
Index Terms—Hydraulic manipulators, neural networks (NNs), mechanical configuration and introduces extra nonlinearities.
prescribed performance function (PPF), robust integral of the Second, considering the hydraulic actuator dynamics increases
sign of the error (RISE), uncertainties. the order of the whole system and the unknown joint coupling
dynamics become mismatched to the final control input, which
greatly complicates the control design. Third, the hydraulic
actuators present highly nonlinear dynamic behaviors (e.g.,
I. I NTRODUCTION nonlinear flow feature of servovalve) and are subjected to var-
YDRAULIC robot manipulators are irreplaceable in ious uncertainties (e.g., parametric uncertainties and uncertain
H heavy-duty operations, such as assembly and mainte-
nance of heavy equipment, energy exploitation, and disaster
nonlinearities) [11]. As a result, few existing results are avail-
able for precision motion control of hydraulic manipulators
rescue in virtue of their outstanding advantages, including and it deserves further study.
Actually, there are plenty of studies devoted to solving the
Manuscript received December 7, 2021; accepted May 22, 2022. This tracking control problem of hydraulic actuators, which could
work was supported in part by the National Natural Science Foundation of
China under Grant 51905271 and Grant 52075262, and in part by the Natural provide some reference for the joint tracking control of com-
Science Foundation of Jiangsu Province under Grant BK20190459. This arti- plex hydraulic manipulators. Specifically, in [11], Yao et al.
cle was recommended by Associate Editor W. Sun. (Corresponding author: proposed an adaptive robust control (ARC) method, which
Jianyong Yao.)
Wenxiang Deng and Jianyong Yao are with the School of Mechanical deals with uncertain system parameters and unmodeled dis-
Engineering, Nanjing University of Science and Technology, Nanjing 210094, turbances simultaneously for single-rod hydraulic actuators.
China (e-mail: wxdeng_njust@163.com; jerryyao.buaa@gmail.com). Meanwhile, lots of disturbance observation-based control strate-
Hang Zhou is with the school of Aeronautics and Astronautics, Shanghai
Jiao Tong University, Shanghai 200240, China gies have also been studied for hydraulic actuation systems
Jin Zhou is with the Nanjing Engineering Institute of Aircraft Systems, by adopting various disturbance observers, such as extended
Jincheng, AVIC, Nanjing 211106, China. state observer [12], [13], high-gain disturbance observer [14],
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSMC.2022.3178626. and finite-time disturbance observer [15]. In addition, with
Digital Object Identifier 10.1109/TSMC.2022.3178626 the powerful approximation stability of NN, some NN-based
2168-2216 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

adaptive tracking controllers have been developed for hydraulic the PPF-based method into the adaptive NN control for a
actuators (see [16]). However, all above-mentioned works can second-order motor servo mechanism without mismatched
only ensure a uniformly bounded tracking performance in the uncertainties in [28]. The PPF was also incorporated into the
presence of both matched and mismatched uncertainties. Using neural adaptive backstepping control of a hydraulic manip-
the strong robustness and continuity of the integral of the sign ulator for the first time [29]. Unfortunately, although the
of the error (RISE) control law [17], it was combined with prescribed transient tracking performance is ensured via the
adaptive control based on the backstepping method for motion PPF, the excellent asymptotic stability with the steady-state
control of hydraulic rotary actuators with mismatched uncer- tracking errors converging to zero is lost in [28] and [29].
tainties in [18]. Compared to ARC, an appealing feature of In this article, the above-mentioned issues are addressed.
the adaptive RISE controller in [18] is that the output tracking An NN-based adaptive asymptotic prescribed performance
error can be guaranteed to converge to zero asymptotically. tracking control scheme is presented for multi-DOF hydraulic
However, the uncertain nonlinearities in the actuator dynamics manipulators. To facilitate the control design, the nonlinear
that is matched to the control input were neglected in [18], model of an n-DOF hydraulic manipulator containing the
which is often inconsistency with real systems. To cope with actuator dynamics is first established. In the controller devel-
this issue, Deng and Yao further designed a recursive adap- opment, the PPF is introduced to restrict the joint tracking
tive RISE control strategy ensuring remarkable asymptotic errors in a prescribed performance constraint. In addition,
stability in [19], and applied it to hydraulic servosystems with the radial basis function NN (RBFNN) is integrated with
both matched (e.g., modeling errors of pressure dynamics) and the recursive RISE feedback control via the backstepping
mismatched uncertainties (e.g., unmodeled friction effects). control framework, in which the RBFNN is synthesized to
However, the applicability of the method in [19] to complex approximate the uncertain joint coupling dynamics for feed-
hydraulic manipulators is not clear. forward compensation and the RISE feedbacks are constructed
In comparison to single degrees of freedom (DOF) hydraulic to attenuate the RBFNN reconstruction error and uncertain-
actuation systems, fewer results are available for tracking ties in the actuator dynamics. Theoretical analysis indicates
control of multi-DOF hydraulic manipulators. In [20], an inte- that the proposed control law ensures a prescribed tracking
grated direct/indirect ARC scheme with valve dead-zone com- performance, meanwhile, ensuring an excellent asymptotic sta-
pensation was developed for a simple single-joint hydraulic bility result. Comparative simulation results are provided to
manipulator without joint coupling dynamics. The strong cou- verify the validity and superiority of the proposed control
pling joint dynamics were considered in [21] and the time strategy.
delay estimation (TDE) method was employed to estimate The main contributions of this article are summarized as
the uncertain joint dynamics; however, the hydraulic actuator follows.
dynamics were neglected to simplify the hydraulic manipula- 1) A novel PPF-based adaptive NN asymptotic prescribed
tor system as a second-order Euler–Lagrange system. In [22], performance controller is proposed by combining the
the iterative learning control was combined with a disturbance RBFNN with the recursive RISE feedback control for
observer, and the disturbance rejection control of a hydraulic multi-DOF hydraulic manipulators with both matched
excavator was realized. It should be pointed out that the con- and mismatched uncertainties.
trol methods in [22] can only ensure the joint tracking errors 2) With the proposed controller, the prescribed tracking
to converge to a bounded range. A sliding mode control algo- performance and asymptotic stability can be simultane-
rithm using a finite-time disturbance observer was synthesized ously ensured for hydraulic manipulators for the first
in [23] for a 3-DOF hydraulic manipulator considering the time.
actuator dynamics and the joint tracking errors were guar- 3) Since the major uncertain system dynamics can be
anteed to converge to zero in finite time under the strong estimated via NN for feedforward compensation, the
assumption that the uncertain joint dynamics were time invari- residual uncertainties that need to be attenuated by the
ant, which is unacceptable in practical applications. Hence, nonlinear RISE feedback are much reduced and the
asymptotic tracking control of complex multi-DOF hydraulic system robustness will be improved.
manipulators still remains an issue and is expected to be The residual part of this article is arranged as follows.
addressed in this article. Section II gives the system models and problem formulation
The other motivation of this article is that all afore- for n-DOF hydraulic manipulators. The development of the
mentioned control approaches mainly focus on the steady- proposed controller is provided in Section III. The closed-loop
state tracking performance improvement, while the transient system stability analysis is shown in Section IV. Section V
performance is often unclear. With the purpose of restricting gives the comparative simulation results for a 2-DOF hydraulic
the joint tracking errors in a prescribed performance con- manipulator. Finally, Section VI draws the conclusion and
straint, a barrier Lyapunov function (BLF) was employed expresses the intended future works.
in the fuzzy control [24] and NN control [25] design for
rigid manipulators. However, the actuator dynamics were
ignored and only the uniformly ultimately bounded stability II. S YSTEM M ODELS AND P ROBLEM F ORMULATION
was achieved. In addition, the prescribed performance func- The structure of an n-DOF hydraulic manipulator studied in
tion (PPF)-based methods [26], [27] can also be used to guar- this article is shown in Fig. 1. The links are driven by single-
antee the prescribed tracking performance. Na et al. introduced rod hydraulic cylinders and/or hydraulic rotary actuators. The

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: NN-BASED ADAPTIVE ASYMPTOTIC PRESCRIBED PERFORMANCE TRACKING CONTROL 3

Since it is difficult to obtain the exact joint dynamics of


hydraulic manipulator, we define the following matrices:


⎪ M0 (q) = M(q) − M(q)

C0 (q, q̇) = C(q, q̇) − C(q, q̇)
(3)

⎪ G0 (q) = G(q) − G(q)

Tf 0 (q̇) = Tf (q̇) − Tf (q̇)
which can be obtained via experimental identification and uti-
lized in the control design process and M(q), C(q, q̇),
G(q), and Tf (q̇) represent the unmodeled parts of joint
coupling dynamics.
Property 1: For any vector θ ∈ Rn , the following inequality
holds:
mθ 2 ≤ θ T M(q)θ ≤ m̄θ 2 (4)
Fig. 1. Structure of an n-DOF hydraulic manipulator. where m and m̄ are both positive constants, and ||•|| represents
the standard Euclidean norm.
Property 2: The matrix Ṁ(q) − 2C(q, q̇) is skew symmetric,
detailed schematic structure of the hydraulic servo system is i.e.,
shown in the right part of Fig. 1.  
In this section, to facilitate the controller development, θ T Ṁ(q) − 2C(q, q̇) θ = 0. (5)
the mathematical model of the n-DOF hydraulic manipulator
including actuator dynamics is established. B. Dynamics of Hydraulic Actuators
As depicted in Fig. 1, the manipulator joint is driven by a
A. Dynamics of n-DOF Manipulators hydraulic cylinder and/or hydraulic rotary actuator controlled by
a servo valve. Let xa = [xa1 , xa2 , . . . , xan ]T be a displacement
The joint dynamics of the considered manipulator can be
vector in the actuator space. The displacement xa and velocity
described as follows:
ẋa in the actuator space can be calculated from the joint vectors
M(q)q̈ + C(q, q̇)q̇ + G(q) + Tf + t1 = T (1) q and q̇ by (6) and (7), as shown at the bottom of the page.
Without losing generality, the pressure dynamics in the ith
where q, q̇, q̈ ∈ Rn represent the vectors of joint angular dis- hydraulic actuator can be written as
placement, velocity, and acceleration, respectively, M(q) ∈  −1
Rn×n is the symmetric and positive definite inertia matrix, Ṗi1 = βe Vi1 (Qi1 − Ai1 ẋai − Cti (P1 − P2 )) + di1 (t)
−1 (8)
C(q, q̇) q˙ ∈ Rn denotes the Centripetal and Coriolis torque, Ṗi2 = βe Vi2 (−Qi2 + Ai2 ẋai + Cti (P1 − P2 )) − di2 (t)
G(q) ∈ Rn represents the vector of gravity, Tf ∈ Rn is where βe is the effective bulk modulus of the oil; Vi1 =
the nonlinear friction torque vector, which will be described Voi1 + Ai1 · xai and Vi2 = Voi2 + Ai2 · xai represent the control
later in detail, t1 ∈ Rn stands for the modeling error, includ- volumes of the actuator chambers, and Voi1 , Voi2 are the initial
ing unmodeled dynamics and disturbances, and T ∈ Rn is the control volumes; Ai1 and Ai2 are the piston area of two cham-
control torque. bers of the cylinder, but stand for the radian displacement of
In this given system models and problem formulation, a the actuator of the rotary motor; Cit denotes the coefficient of
continuously differentiable friction model [30] suitable for the internal leakage; Qi1 and Qi2 are the supplied and return
backstepping design is adopted, which has the following form: flow rate, respectively; di1 (t) and di2 (t) are the lumped distur-
Tf = α1 tanh(β1 q̇) − tanh(β2 q̇) + α2 (tanh(β3 q̇)) + α3 q̇ (2) bance of two chambers, including unmodeled internal leakage
characteristics and other complicated effects.
where α1 , α2 , and α3 are friction coefficients representing dif- Consider that high-response servo valves are used in
ferent friction levels, and β1 , β2 , and β3 are shape coefficients the considered hydraulic manipulators, their dynamics are
utilized to approximate friction effects. neglected due to much faster response than the remaining


T
xa = n(q) = n1 (q1 ), n2 (q2 ), . . . , nn (qn ) ∈ Rn

2 + d 2 − 2d d cos(q ) − d , if the ith link is
di1 driven by a linear actuator
ni (qi ) = i2 i1 i2 i i0 (6)
qi , if the ith link is driven by a rotary actuator
ẋa = Ja (q)q̇ = diag{Ja1 (q1 ), Ja2 (q2 ), . . . , Jan (qn )}q̇

⎨ di1 di2 sin(qi ) , if the ith link is driven by a linear actuator
Jai (qi ) = 2 +d 2 −2d d cos(q )
di1 i2 i1 i2 i (7)
⎩ 1, if the ith link is driven by a rotary actuator

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

parts. Therefore, the spool position xiu is proportional to its Assumption 1 [19], [36]: The modeling error terms t1
control input voltage ui , i.e., xiu = kiu · ui . Then, Qi1 and Qi2 and t2 are smooth enough such that their first and second
can be described by time derivatives are bounded, i.e., ˙ t1  ≤ δ1 ,  ¨ t1  ≤ δ2
  √ √  ˙ ¨
,t2  ≤ δ3 , and t2  ≤ δ4 , where δ1 , δ2 , δ3 , and δ4 are
Qi1 = kiq kiu ui s(ui )√Pis − Pi1 + s(−ui )√Pi1 − Pir 
(9) known positive constants.
Qi2 = kiq kiu ui s(ui ) Pi2 − Pir + s(−ui ) Pis − Pi2
Assumption 2: The desired position trajectories of all joints

where kiq =Cid ωi 2/ρ and Cid is the discharge coefficient; ωi are third-order differentiable and bounded, i.e., x1d (t) ∈ C3 .
represents the spool valve area gradient; ρ denotes the density
of oil; Pis and Pir stand for the supply and return pressure, A. Prescribed Performance Function
respectively; and s(ui )is defined as First, we define the joint tracking error vector e(t) =
 [e1 (t), e2 (t), . . . , en (t)] ∈ Rn as
1, ui ≥ 0
s(ui ) = (10)
0, ui < 0. 
e(t) = x1 (t) − x1d (t). (15)
The torque acting on the manipulator joint can be written as For the purpose of constraining the joint tracking errors
T = Ja (q)(A1 P1 − A2 P2 ) (11) to guarantee prescribed transient and steady-state tracking
performance, a positive and exponentially decreasing function
where A1 = diag{A11 , A21 , . . . , An1 } ∈ Rn×n , A2 = μ(t) ∈ R+ is introduced [26], [27]
diag{A12 , A22 , . . . , An2 } ∈ Rn×n , P1 = [P11 , P21 , . . . , Pn1 ]T ∈
μ(t) = (μ0 − μ∞ )e−κt + μ∞ (16)
Rn , and P2 = [P12 , P22 , . . . , Pn2 ]T ∈ Rn .
in which μ0 > μ∞ > 0 and κ > 0 are parameters to be
C. State Space Model selected.
Apparently, we are able to guarantee the control objective
Define the state variable vector as x = [x1T , x2T , x3T ]T , in
by satisfying
which x1 = q, x2 = q̇, x3 = A1 P1 − A2 P2 . From (1), (8), and
(11), the whole hydraulic manipulator model can be expressed − δ i μ(t) < ei (t) < δ̄i μ(t) ∀t > 0 (17)
as the following state-space form:
⎧ where δ i , δ̄i > 0 for i = 1, 2, . . . , n. Moreover, it is obvious
⎨ ẋ1 = x2
that −δ i μ0 and δ̄i μ0 denote the lower bound of the undershoot
ẋ = M −1 (x1 ) Ja (x1 )x3 − C(x1 , x2 )x2 − G(x1 ) − Tf − t1 and the maximum value of the overshoot of ei (t); κ determines
⎩ 2
ẋ3 = ϕ1 u − ϕ2 − ϕ3 + t2 the convergence rate of ei (t); and −δ i μ∞ and δ̄i μ∞ signify
(12) the allowable steady-state value of ei (t).
In order to introduce the performance function μ(t) into
in which the process of the controller design, we need to transform the
⎧ 
⎪  constrained error signal (17) into an equivalent unconstrained

⎪ ϕ1 = βe kt AV11 R1 + AV22 R2

⎨   one. To this end, we utilize a smooth and strictly increasing
 A21 A22
ϕ2 = βe Ja (x1 ) V1 + V2 x2 (13) function ψi (z1i ) with respect to z1i (t) ∈ R, which has the

⎪ 

⎪ 
following properties.
⎩ ϕ3 = βe Ct AV11 + AV22 (P1 − P2 ) 1) −δ i < ψi (z1i ) < δ̄i ∀z1i ∈ L∞ .
2) limz1 →+∞ ψi (z1i ) = δ̄i , limz1i →−∞ ψi (z1i ) = −δ i .
where kt = diag{k1t , k2t , . . . , knt } ∈ Rn×n , Based on the above properties, we have
kit kiu · kiq , V1 = diag{V11 , V21 , . . . , Vn1 } ∈
Rn×n , V2 = diag{V12 , V22 , . . . , Vn2 } ∈ Rn×n , ei (t) = μ(t)ψi (z1i ). (18)
Ct = diag{C1t , C2t , . . . , Cnt } ∈ Rn×n , t2 = As in [28], we define ψi (z1i ) as
A1 d1 (t) − A2 d2 (t), d1 = [d11 (t), d21 (t), . . . , dn1 (t)]T ∈ Rn ,
d2 = [d12 (t), d22 (t), . . . , dn2 (t)]T ∈ Rn , and R1 and R2 have δ̄i ez1i − δ i e−z1i
ψi (z1i ) = (19)
the following definitions: ez1i + e−z1i
√ √ Then, the transformed error signal z1i (t) can be described

R1 = s(u) Ps − P1 + s(−u) P1 − Pr as follows:
 √ √ (14)  
R2 = s(u) P2 − Pr + s(−u) Ps − P2 . −1 ei (t) 1 λi (t) + δ i
z1i (t) = ψi = ln (20)
μ(t) 2 δ̄i − λi (t)
III. C ONTROLLER D ESIGN where λi (t) = ei (t)/μ(t).
The objective of this gives the system models and Remark 1: The parameters μ0 , μ∞ , κ, δ i , and δ̄i should
problem formulation is to have all joints of the manip- be selected to guarantee that for any initial condition
ulator track the desired trajectory x1d (t) = qd (t) = ei (0), −δ i μ(0) < ei (0) < δ̄i μ(0) should be satisfied. In addi-
[q1d (t), q2d (t), . . . , qnd (t)]T ∈ Rn as accurately as possible. tion, as long as z1i (t) can be bounded under the controller
Furthermore, the joint tracking errors must be restricted in the action, the condition (17) can be guaranteed according to the
prescribed performance constraint. properties of ψi (z1i ). It indicates that if z1i (t) can be made
The following assumptions that will be utilized in the asymptotically converge to zero, the joint tracking error ei (t)
proposed control design are presented. will also converge to zero asymptotically.

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: NN-BASED ADAPTIVE ASYMPTOTIC PRESCRIBED PERFORMANCE TRACKING CONTROL 5

B. Transformed Filtered Tracking Error Dynamics Furthermore, the unknown term d , including the uncer-
Define the transformed error vector z1 (t) = [z11 (t), z12 (t)]T tain joint coupling dynamics and unmodeled disturbances, is
and differentiate z1 (t) with respect to time, we obtain defined as follows:
    
ż1 = 12 λ+δ
1
− λ−1 ė
− eμ̇
= ρ − eμ̇ d = −(M(x1d )ẍ1d + C(x1d , ẋ1d )ẋ1d
δ̄ μ μ 2 ė μ (21) 
+ G(x1d ) + Tf (ẋ1d ) + t1 . (28)
where λ= diag{λ1 , λ2 , . . . , λn } ∈ Rn×n , e = [e1 , e2 , . . . , en ]T ∈
Rn , δ = diag{δ 1 , δ 2 , . . . , δ n }∈ Rn×n , δ̄ =diag{δ̄1 , δ̄2 , . . . , δ̄n }∈ It can be inferred that d is a smooth function according to
Rn×n , and ρ = (1/2μ)[1/(λ + δ) − 1/(λ − δ̄)] ∈ Rn×n . Assumptions 1 and 2, and can be approximated by RBFNNs.
Property 3: Not hard to see that variable ρ is bounded, i.e., Therefore, d can also be expressed as
   
T
0 < ρ < ρM = δ + δ̄ / μ∞ δ δ̄ . (22) d = W1T h1 (χ1 ), W2T h2 (χ2 ), . . . , WnT hn (χn ) + ε (29)

Furthermore, the second-order differential of z1 is available where Wi ∈ RHi (i = 1, 2, . . . , n) is the bounded constant ideal
by combining (12) and weight matrix, and Hi denotes the quantity of the nodes of the
    hidden layer, hi (χi ) ∈ RHi (i = 1, 2, . . . , n) signify the output
eμ̇ ėμ̇ eμ̈ eμ̇2 of network’s Gauss radial function and the detailed expression
z̈1 = ρ̇ ė − −ρ + − 2
μ μ μ μ can be found in [29], χi = [1, qid , q̇id , q̈id ]T (i = 1, 2, . . . , n)
−1 denotes the NN’s input, and ε = [ε1 , ε2 , . . . , εn ]T ∈ Rn repre-
+ M (x1 )ρJa (x1 )x3
 sents the functional reconstruction error and has the following
− M −1 (x1 )ρ M(x1 )ẍ1d + C(x1 , x2 )x2
 property.
+ G(x1 ) + Tf + t1 . (23) Property 4: Referring to [34] and Assumption 1, the approx-
Define a set of filtered error signal z2 (t), r1 (t), z3 (t), r2 (t) imation error and its time derivatives up to second order are
∈ Rn as bounded

z2 = ż1 + k1 z1 , r1 = ż2 + k2 z2 , ε ≤ εd1 , ε̇ ≤ εd2 , ε̈ ≤ εd3 (30)


z3 = x3 − α2 , r2 = ż3 + k3 z3 (24) where εd1 , εd2 , and εd3 are both positive constants.
Based on (29), the approximation  ˆ d for d is given by
where k1 , k2 , k3 ∈ Rn×n are positive diagonal constant matri-
 T
ces, and α 2 is the virtual control law that will be given in the ˆ d = Ŵ1T h1 (χ1 ), Ŵ2T h2 (χ2 ), · · · , ŴnT hn (χn ) (31)
subsequent design process.
where Ŵi ∈ RHi (i = 1, 2, . . . , n) is the estimation of the ideal
C. Neural Network Approximation weight matrix Wi .
The complex joint coupling dynamics and unmodeled dis- The definition of the weight matrix estimation error is
turbances are the major uncertainties of the hydraulic manip- given as
ulator, which are the main obstacles to achieve high-precision
W̃i = Wi − Ŵi . (32)
motion tracking. Hence, in this section, the RBF NN is
employed to estimate the uncertain joint coupling dynamics The update law of Ŵi will be given subsequently. The
and unmodeled disturbances for feedforward compensation in following variables are defined to simplify the deduction:
the controller design.  T
˜d =

T
First, based on (1), (23), and (24), we have  W̃1 h1 (χ1 ), W̃2T h2 (χ2 ), . . . , W̃nT hn (χn )
 
˜ dW = ˙T ˙ T h (χ ), . . . , W̃ ˙ T h (χ ) T
M(x1 )r1 = E + Fn + ρd + ρJa (z3 + α2 ) (25)  W̃1 h1 (χ1 ), W̃ 2 2 2 n n n

˜  T
T
where E is defined as dh = W̃1 ḣ1 (χ1 ), W̃2 ḣ2 (χ2 ), . . . , W̃n ḣn (χn ) .
T T


E = M(x1 )[(k1 ż1 + k2 z2 )]
     D. RISE-Based Controller Design
eμ̇ ėμ̇ eμ̈ eμ̇2
+ M(x1 ) ρ̇ ė − −ρ + − 2 The control scheme is developed via the backstepping con-
μ μ μ μ trol framework since the uncertain joint coupling dynamics and
+ ρ{[M0 (x1d ) + M(x1d ) − M0 (x1 ) + M(x1 )]ẍ1d } unmodeled disturbances are mismatched to the control input.
+ ρ[C0 (x1d , ẋ1d )ẋ1d − C0 (x1 , x2 )x2 ] The entire diagram of the proposed controller is shown in
+ ρ[C(x1d , ẋ1d )ẋ1d − C(x1 , x2 )x2 ] Fig. 2.
+ ρ{[G0 (x1d ) + G(x1d )] − [G0 (x1 ) + G(x1 )]} Step 1: Based on (24) and (25), the virtual control law α2


 is design as
+ ρ Tf 0 (ẋ1d ) + Tf (ẋ1d ) − Tf 0 (x2 ) + Tf (x2 ) .
(26) α2 = ρ −1 Ja−1 (α2a + α2s1 + α2s2 )
α2a = −Fn − ρ  ˆd
The desired trajectory-based nominal compensation term Fn α2s1 = −(ks + kr1 )z2 + (ks + kr1 )z2 (0)
can be calculated by
α2s2 = − ∫t0 ((ks + kr1 )k2 z2 + β1 sign(z2 ))dτ. (33)

Fn = −ρ[M0 (x1d )ẍ1d + C0 (x1d , ẋ1d )ẋ1d where ks , kr1 , β 1 ∈ Rn×n
are diagonal positive matrices, and α 2a

+ G0 (x1d ) + Tf 0 (ẋ1d ) . (27) is designed so as to compensate the parameter uncertainties, and

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

1
M(x1 )ṙ1 = Ñ + N − Ṁ(x1 )r1
2
− (ks + kr1 )r1 − β1 sign(z2 ) + ρJa ż3 (41)
 
˜ dW + (ρ̇Ja + ρ J̇a )z3 , N =
where Ñ = −1/2Ṁ(x1 )r1 + Ė + ρ 
   ˜
Nd + NB1 + NB2 , NB = NB1 + NB2 , Nd = ρ ε̇ + ρ̇ε, NB1 =  d,
 ˜ dh + ρ̇ 
˜d − ˜ d.
and NB2 = ρ 
Lemma 1 [36]: Utilizing the mean value theorem, Ñ has
can be upper bounded by
 
Fig. 2. Entire diagram of the proposed controller. Ñ  ≤ η(z)z (42)
where z = [zT1 zT2 r1T zT3 r2T ]T ∈ R5n , and the bounding function
α 2s1 is the linear robust feedback term to stabilize the nominal η (z) is positive globally invertible nondecreasing function.
model of the manipulator and α 2s2 is the RISE feedback Property 8: Based on Assumption 1, and Properties 3–5 and
control law to dominate the approximation error. 7, the following inequalities hold:
In the meantime, the weight update law of the RBFNN is
N  ≤ ς1 , NB1  + NB2  ≤ ς2
designed as  d  
 Ṅd  ≤ ς3 , ṄB  ≤ ς4 + ς5 z2  (43)
˙ = sat
Ŵi ẆiM ProjŴi (i τi )
where ςi , i = 1, 2, . . . , 5 are known positive constants.
τi = hi (χi )z2i (34) Combining (12) and (24), the filtered error r2 can be
where i ∈ RHi ×Hi (i = 1, 2, . . . , n) is a positive definite diag- expressed as
onal matrix to determine the weight adaptation rate, and τi is r2 = ϕ1 u − ϕ2 − ϕ3 + t2 − α̇2 + k3 z3 . (44)
the designed adaptation function for the RBF NN used on the
ith joint. As in [35], we use the projection mapping ProjŴ (•) Therefore, the actual control input u is designed as follows:
to keep the estimated weight within the preset bounded set u = ua + urise + us , us = −ϕ1−1 k3 z3
W = {W|Wmin ≤ W ≤ Wmax }
⎧ ua = ϕ1−1 (α̇2 + ϕ2 + ϕ3 )
⎨ 0 Ŵ = Wmax and ζ > 0 
urise = ϕ1−1 − kr2 z3 + kr2 z3 (0)
ProjŴ (ζ ) = 0 Ŵ = Wmin and ζ < 0 (35) 
⎩ − ∫t0 (kr2 k3 z3 + β2 signa(z3 ))dτ
ζ, otherwise.
(45)
Furthermore, to facilitate the stability analysis, the satura-
tion function satẆM (•) is used to guarantee the adaptation rate where ua is utilized to compensate the nominal model of
always within the limit ẆM and is defined as follows: hydraulic actuators, urise is the RISE feedback term to deal
with the matched model uncertainties, and kr2 , β 2 ∈ Rn×n
satẆM (ζ ) = s0 ζ are positive diagonal constant matrices, us is the linear robust

1 ζ  ≤ ẆM feedback to stabilize the system. Applying the designed con-
s0 = (36)
ẆM /ζ  ζ  > ẆM trol input u, the dynamics of filtered error signal r2 can be
expressed by
Using the update law given by (34) with Wmin ≤Ŵ(0)≤
Wmax , for any adaptation function τ i , we have the following ˙ t2 − kr2 r2 − β2 sign(z3 ).
ṙ2 =  (46)
properties.
Property 5: The weight estimates can be guaranteed to be IV. S TABILITY A NALYSIS
within the preset bounded set W , i.e., Lemma 2: Let us define two auxiliary functions L1 (t), P1 (t),
Wimin ≤ Ŵi (t) ≤ Wimax , i = 1, 2, . . . , n ∀t. (37) L2 (t), P2 (t) ∈ R

Property 6: L1 (t) = r1T (NB2 (t) + Nd (t) − β1 sign(z2 ))
 +żT2 (t)NB1 (t) −  z2 2 (47)
WiT i−1 ProjŴ (i τi ) − i ≤ 0, i = 1, 2, . . . , n ∀t. (38)   
L2 (t) = r2T  ˙ t2 (t) − β2 sign(z3 ) (48)
Property 7: The weight update rate is bounded by
 n
  P1 (t) = β1 |z2i (0)| − zT2 (0)N(0) − ∫t0 L1 (τ )dτ (49)
˙ 
Ŵi (t) ≤ ẆiM , i = 1, 2, . . . , n ∀t. (39) i=1

Step 2: By integrating (24), (29), (32), (33), and (25), we  n


P2 (t) = β2 ˙ t2 (0) − ∫t0 L2 (τ )dτ (50)
|z3i (0)| − zT3 (0)
can rewrite (25) as
i=1
˜ d + ρε.
M(x1 )r1 = E + ρJa z3 + α2s1 + α2s2 + ρ  (40) as long as the subsequent conditions are satisfied, and P1 (t)
and P2 (t) are always positive
Taking the time derivative of (40) on the both sides, we
obtain β1 > In ς1 + In ς2 + ς3 k2−1 + ς4 k2−1 ,  > ς5 (51)

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: NN-BASED ADAPTIVE ASYMPTOTIC PRESCRIBED PERFORMANCE TRACKING CONTROL 7

β2 > In δ3 + δ4 k3−1 (52) 


n
˙
˜d−
+ zT2 k2  k2i W̃iT i−1 Ŵ i
where In represents an identity matrix. The detailed proof of i=1
Lemma 2 can be found in [19], [32], and [36]. = −zT z − r1T ks r1 + r1T Ñ (57)
Theorem: Using the synthesized weight update law of the
RBFNN and selecting large enough control gains k1 , k2 , k3 , where the matrix  is defined as
⎡ ⎤
kr1 , kr2 , β 1 , and β 2 relative to the system initial conditions k1 − I2n 0 0 0
to make the matrix  in (58) positive definite and the condi- ⎢ − In k −  − In 0 ⎥
⎢ 2 2 0 ⎥
⎢ 2
ρJa k3 ⎥
tions (51) and (52) satisfied, the proposed adaptive asymptotic =⎢ 0 − I2n kr1 − ρJ2 a ⎥ (58)
prescribed performance control law (45) guarantees that all ⎢ ρJa k3
2
In ⎥
⎣ 0 0 2 k3 −2 ⎦
the closed-loop system signals are bounded and the output
0 0 − ρJ2 a − I2n kr2
tracking error will be restricted in the prescribed constraint
(17). Moreover, the excellent asymptotic tracking with zero in which 0 represents the zero matrix with proper dimensions.
steady-state error can also be obtained. By choosing large enough feedback gain matrices k1 , k2 , k3 ,
Proof: Define an auxiliary function Q(t) ∈ R as kr1 , and kr2 such that the matrix  is positive definite and
utilizing (42), V̇(γ , t) satisfies the inequality as follows:
1
n

Q(t) = k2i W̃iT i−1 W̃i (53)
2 V̇(γ , t) ≤ −κmin ()z2 − κmin (ks )r1 2 + η(z)zr1 
i=1  
η(z)z 2
and it is obvious that Q(t) ≥ 0. ≤ −κmin ()z − κmin (ks ) r1  −
2
Let D ⊂ R5n+3 be a domain including γ (t) = 0, in which 2κmin (ks )
γ (t)∈ R5n+3 is given as η (z)z
2 2
+
    T 4κmin (ks )
  
γ (t) = zT (t), P1 (t), P2 (t), Q(t) . (54) η2 (z)
≤ − κmin () − z2
Choosing the positive definite Lyapunov function V(γ , t) as 4κmin (ks )

1 T
 1 1 1 ≤ −ϑz2 = −U(γ ) (59)
V(γ , t) = z1 z1 + zT2 z2 + zT3 z3 + r1T M(x1 )r1
2 2 2 2
1 T where ϑ is some positive constant under the following condi-
+ r2 r2 + P1 (t) + P2 (t) + Q(t). (55) tion:
2  
Using Property 1, we can bound (55) by z ≤ η−1 2 κmin () · κmin (ks ) . (60)
U1 (γ ) ≤ V(γ , t) ≤ U2 (γ ) (56) U(γ ) is a positive semidefinite function defined in the domain
  2  D as follows:
where U1 (γ ) = σ1 γ 2 , U2 (γ ) = σ2 γ  , σ1 =  !


min{1, m}/2, σ2 = max{2, m̄}/2, and κmin (•)and κmax (•) D = γ ∈ R5n |γ  ≤ η−1 2 κmin () · κmin (ks ) . (61)
signify the minimum and maximum eigenvalue of the
From (59), it can be inferred that V(γ , t) ∈ L∞ and U(γ )
argument, respectively.
∈ L2 . Based on (55), we can obtain that the error signals
Combining (24), (32), (34), (46), and (50), we have
z1 , z2 , z3 , r1 , r2 , and W̃i are all bounded. Using Assumption
V̇(γ , t) = zT1 ż1 + zT2 ż2 + zT3 ż3 + r1T M(x1 )ṙ1 2, the system states x1 , x2 , and x3 can be concluded to be
1 bounded. By checking the structure of the control input u, it
+ r1T Ṁ(x1 )r1 + r2T ṙ2 + Ṗ1 + Ṗ2 + Q̇ is also bounded. Therefore, all the closed-loop system signals
2
= zT1 (z2 − k1 z1 ) + zT2 (r1 − k2 z2 ) + zT3 (r2 − k3 z3 ) are bounded in the domain D. In addition, based on (24), (41),
1   and (46), we know that the time derivatives of z1 , z2 , z3 , r1 ,
+ r1T Ṁ(x1 )r1 + r2T  ˙ t2 − kr2 r2 − β2 sign(z3 ) and r2 are bounded, which means that U̇(γ ) is an uniformly
2
continuous function in the domain D.
− r1T (NB2 (t) + Nd (t) − β1 sign(z2 ))
  Define a set ϒ ∈ D as
− żT2 (t)NB1 (t) +  z2 2 − r2T  ˙ t2 (t) − β2 sign(z3 )  "  2 #
  "
1 "
ϒ = γ ⊂ D"U2 (γ ) < σ1 η −1
2 κmin () · κmin (ks ) .
+ r1T Ñ + N − Ṁ(x1 )r1 − (ks + kr1 )r1
2 (62)

− β1 sign(z2 ) + ρJa (r2 − k3 z3 ) Using [33, Th. 8.4], it can be inferred that

n
˙ ϑz(t)2 → 0 as t → ∞ ∀γ (0) ∈ ϒ. (63)
− k2i W̃iT i−1 Ŵ i
i=1 Based on the definition of z(t) and Remark 1, we can draw
= −z1 k1 z1 − zT2 k2 z2 − zT3 k3 z3 − r1T kr1 r1 − r2T kr2 r2
T the conclusion that the prescribed constraint (17) is satisfied
+ zT1 z2 + zT2 r1 + zT3 r2 + r1T ρJa r2 − r1T ρJa k3 z3 and the asymptotic tracking is also achieved, i.e.,
+ r1T Ñ − r1T ks r1 +  z2 2 e(t) → 0 as t → ∞ ∀γ (0) ∈ ϒ. (64)

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

TABLE I
Remark 2: Noting that the control gains k1 , k2 , k3 , kr1 , and M ODEL PARAMETERS
kr2 are on the main diagonal of the matrix  in (58), we only
need to select these gains large enough to make the matrix 
positive definite. By doing so, the positive definiteness of 
is easy to satisfy and offline calculation work will be much
reduced.

V. S IMULATION R ESULTS
In this section, simulation results are provided to illus-
trate the superiority of the proposed controller, and a 2-DOF
hydraulic manipulator is chosen to perform the simulation,
whose one link is driven by a hydraulic cylinder and the other
one is driven by a hydraulic rotary actuator.
The system matrix M(q), C(q, q̇), and G(q) utilized in the
simulation has the following structure [37], [38]:
 
m11 m12
M(q) = (65)
m21 m22
   
c c 2 × 104 sin(0.5t)Pa · m2 /s
C(q, q̇) = 11 12 (66) t2 = . (73)
c21 c22 6 × 103 sin(1.8t)Pa · m3 /rad/s
 
g
G(q) = 1 (67) To assess the validity of the proposed control scheme, the
g2 following three controllers are chosen to carry out comparative
in which mij , cij , gi (i, j = 1, 2) can be calculated as (68)–(70) simulation.
⎧ 1) PPF-RISE-NN: This is the proposed controller.
⎪ m11 = I1 + I2 + m1 Lc1 2 + m (L2 + L2 + 2L L cos(q ))
⎪ 2 1 1 c2 2
=


c2 The controller parameters are set as: k1

⎪ +m (L 2 + L 2 + 2L L cos(q ))

L 1 2 1 2 2 diag{5, 5}, k2 = diag{10, 10}, k3 = diag{20, 20}, ks =
m12 = I2 + m2 (Lc2 2 + L L cos(q ))
1 c2 2 I2, kr1 = diag{100, 100}, kr2 = diag{8, 55}, β1 =
(68)
⎪ +mL (L2 + L1 L2 cos(q2 ))

2
diag{0.05, 0.01}, and β2 = diag{0.1, 0.1}. The nodes



⎪ m = m12
⎩ 21 H1 and H2 of the hidden layer used in NN are
m22 = I2 + m2 Lc2 2 + m L2
L 2 27 and 19, respectively. The weight adaption rates


⎪ c11 = −(m2 L1 Lc2 + mL L1 L2 )sin(q2 )q̇2 are set as: 1i = 30, 2i = 25, i = 1, 2, . . . , Hi .

c12 = −(m2 L1 Lc2 + mL L1 L2 )sin(q2 )(q̇2 − q̇1 ) As for the PPF parameters, δ̄i = δ i = 0.2,
(69)
⎪ c21 = m2 L1 Lc2 sin(q2 )q̇1 − mL L1 L2 sin(q2 )q̇2
⎪ μ0 = 0.8, μ∞ = 0.05, κ = 2, and i = 1, 2.

c22 = 0 2) RISE-NN: The RISE controller based on the NN is com-

⎨ g1 = (m1 Lc1 + m2 L1 + mL L1 )g cos(q1 ) pared here, whose virtual and actual control law are
+(m2 Lc2 + mL L2 )g cos(q1 + q2 ) (70) designed as follows:

g2 = (m2 Lc2 + mL L2 )g cos(q1 + q2 ) α2 = Ja−1 (−Fd − ˆ d − (kr1 + ks )z2 + (kr1 + ks )z2 (0)
where mi denotes the mass of the ith link, mL represents the − ∫0 ((kr1 + ks )k2 z2 + βsign(z2 ))dτ )
t
(74)
mass of load, g is the gravitation acceleration, Li stands for 
˙
Ŵ = sat Proj ( h (χ )z ) (75)
the length of the ith link, Lci represents the length from center- i ẆiM Ŵi i i i 2i
of-gravity position to joint axis of the ith link, and Ii is the u= ϕ1−1 (α̇2 + ϕ2 + ϕ3 − k3 z3 − kr2 z3 + kr2 z3 (0)
inertial mass of the ith link around its joint axis. The model 
− ∫t0 (kr2 k3 z3 + β2 sign(z3 ))dτ . (76)
parameters are listed in Table I.
We set the initial value of the joint angle position and The corresponding control gains are selected to be the
angle velocity as [q1 (0), q2 (0)]T = [π/3, 0.1]T (rad) and same with those of PPF-RISE-NN to ensure the fair
[q̇1 (0), q̇2 (0)]T = [0, 0]T (rad/ s), respectively. The desired comparison.
trajectories for joint angles are selected as 3) RISE: The RISE-NN controller degrades into the RISE
  controller by taking out NN approximation. Its virtual
0.4 sin(1.2t) + 1
x1d = (rad). (71) control law and actual control input are expressed as
0.2 sin(2t) + 0.2
In order to compare different controllers impartially, we use α2 = Ja−1 (−Fd − (kr1 + ks )z2 + (kr1 + ks )z2 (0)

60% of the real values of system matrices in actual control − ∫t0 ((kr1 + ks )k2 z2 + βsign(z2 ))dτ (77)
law for feedforward compensation and the residual parts are 
= ϕ1−1 α̇2 + ϕ2 + ϕ3 − k3 z3 − kr2 z3 + kr2 z3 (0)
treated as the uncertain joint coupling dynamics. Moreover, 
the disturbances of the system are set as − ∫t0 (kr2 k3 z3 + β2 sign(z3 ))dτ . (78)
 
150 sin(t) The tracking performance for joint 1 and joint 2 with
t1 = (N.m) (72)
80 sin(2t) the proposed control method is given in Fig. 3. As shown,

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: NN-BASED ADAPTIVE ASYMPTOTIC PRESCRIBED PERFORMANCE TRACKING CONTROL 9

Fig. 6. NN output with PPF-RISE-NN.


Fig. 3. Tracking performance with the proposed controller.

Fig. 4. Tracking errors for joint 1 with different controllers.


Fig. 7. Control inputs of two joints with the compared controllers.

compared with the proposed one. Meanwhile, consistent with


the theoretical result, the remarkable asymptotic tracking is
achieved by the proposed controller. The comparison between
RISE and RISE-NN indicates that the NNs contribute to the
model feedforward compensation and improving the tracking
performance. In addition, the NN output with PPF-RISE-NN
is provided in Fig. 6, which shows that the NN approximation
has an excellent estimation performance. The control inputs of
the compared controllers are presented in Fig. 7, which shows
they are all regular and bounded. The filtered error z1 and ρ
in PPF are presented in Figs. 8 and 9, respectively.

Fig. 5. Tracking errors for joint 2 with different controllers. VI. C ONCLUSION
This gives the system models and problem formulation
focuses on dealing with the tracking control for multi-DOF
the actual positions of the two joints can well track the hydraulic manipulators with both matched and mismatched
desired trajectories. The tracking errors of the three con- uncertainties. The whole nonlinear model of n-DOF hydraulic
trollers for different joints are gathered in Figs. 4 and 5, manipulators, including rigid joint dynamics and hydraulic
from which we can see that both the transient and steady- actuator dynamics, is given. In the control design, the uncer-
state tracking errors are restricted in prescribed performance tain joint coupling dynamics and unmodeled disturbances,
constraints with the proposed PPF-RISE-NN controller owing i.e., mismatched uncertainties, are estimated by RBF NNs.
to the use of PPF. However, the preset constraints are vio- In addition, two RISE feedback terms are integrated to cope
lated for the RISE and RISE-NN controllers, which means with the matched uncertainties and the NN reconstruction
worse tracking performance is obtained via the two controllers error, respectively. Furthermore, the PPF is introduced to

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

[6] D. Shi, J. Zhang, Z. Sun, G. Shen, and Y. Xia, “Composite trajectory


tracking control for robot manipulator with active disturbance rejection,”
Control Eng. Pract., vol. 106, Jan. 2021, Art. no. 104670.
[7] R. Patelski and P. Dutkiewicz, “On the stability of ADRC for manipu-
lators with modelling uncertainties,” ISA Trans., vol. 102, pp. 295–303,
Jul. 2020.
[8] Q. Zhou, S. Zhao, H. Li, R. Lu, and C. Wu, “Adaptive neural network
tracking control for robotic manipulators with dead zone,” IEEE Trans.
Neural Netw. Learn. Syst., vol. 30, no. 12, pp. 3611–3620, Dec. 2019.
[9] W. He, A. O. David, Z. Yin, and C. Sun, “Neural network control of a
robotic manipulator with input deadzone and output constraint,” IEEE
Trans. Syst., Man, Cybern., Syst., vol. 46, no. 6, pp. 759–770, Jun. 2016.
[10] W. He, Y. Chen, and Z. Yin, “Adaptive neural network control of an
uncertain robot with full-state constraints,” IEEE Trans. Cybern., vol. 46,
no. 3, pp. 620–629, Mar. 2016.
[11] B. Yao, F. Bu, J. Reedy, and G. T.-C. Chiu, “Adaptive robust motion
control of single-rod hydraulic actuators: Theory and experiments,”
IEEE/ASME Trans. Mechatronics, vol. 5, no. 1, pp. 79–91, Mar. 2000.
Fig. 8. Signal ρ of two joints with PPF-RISE-NN controller. [12] W. Deng, J. Yao, Y. Wang, X. Yang, and J. Chen, “Output feed-
back backstepping control of hydraulic actuators with valve dynamics
compensation,” Mech. Syst. Signal Process., vol. 158, Sep. 2021,
Art. no. 107769.
[13] W. Deng and J. Yao, “Extended-state-observer-based adaptive con-
trol of electrohydraulic servomechanisms without velocity measure-
ment,” IEEE/ASME Trans. Mechatronics, vol. 25, no. 3, pp. 1151–1161,
Jun. 2020.
[14] D. Won, W. Kim, D. Shin, and C. C. Chung, “High-gain distur-
bance observer-based backstepping control with output tracking error
constraint for electro-hydraulic systems,” IEEE Trans. Control Syst.
Technol., vol. 23, no. 2, pp. 787–795, Mar. 2015.
[15] Z. Xu, D. Ma, J. Yao, and N. Ullah, “Feedback nonlinear robust control
for hydraulic system with disturbance compensation,” Proc. Inst. Mech.
Eng. I, J. Syst. Control Eng., vol. 230, no. 9, pp. 978–987, Oct. 2016.
[16] Y.-J. Liu, Q. Zeng, L. Liu, and S. Tong, “An adaptive neural network
controller for active suspension systems with hydraulic actuator,” IEEE
Trans. Syst., Man, Cybern., Syst., vol. 50, no. 12, pp. 5351–5360,
Dec. 2020.
[17] B. Xian, D. M. Dawson, M. S. de Queiroz, and J. Chen, “A continuous
asymptotic tracking control strategy for uncertain nonlinear systems,”
Fig. 9. Signal z1 of two joints with PPF-RISE-NN controller. IEEE Trans. Autom. Control, vol. 49, no. 7, pp. 1206–1211, Jul. 2004.
[18] J. Yao, Z. Jiao, D. Ma, and L. Yan, “High-accuracy tracking control
restrict both the transient and steady-state tracking errors to of hydraulic rotary actuators with modeling uncertainties,” IEEE/ASME
Trans. Mechatronics, vol. 19, no. 2, pp. 633–641, Apr. 2014.
prescribed performance constraints. The Lyapunov-based sta- [19] W. Deng and J. Yao, “Asymptotic tracking control of mechani-
bility analysis indicates that the excellent asymptotic stability cal servosystems with mismatched uncertainties,” IEEE/ASME Trans.
can be ensured and both the transient and steady-state track- Mechatronics, vol. 26, no. 4, pp. 2204–2214, Aug. 2021.
[20] A. Mohanty and B. Yao, “Integrated direct/indirect adaptive robust con-
ing errors are constrained in preset regions. Finally, a 2-DOF trol of hydraulic manipulators with valve deadband,” IEEE/ASME Trans.
hydraulic manipulator is taken as an example to carry out the Mechatronics, vol. 16, no. 4, pp. 707–715, Aug. 2011.
comparative simulations, which verifies the effectiveness and [21] Y. Wang, G. Luo, L. Gu, and X. Li, “Fractional-order nonsingular
terminal sliding mode control of hydraulic manipulators using time delay
superiority of the proposed controller. In the future, it is worth estimation,” J. Vib. Control, vol. 22, no. 19, pp. 3998–4011, Nov. 2016.
developing the NN-based output feedback asymptotic tracking [22] G. J. Maeda, I. R. Manchester, and D. C. Rye, “Combined ILC and dis-
controllers for hydraulic manipulators. turbance observer for the rejection of near-repetitive disturbances, with
application to excavation,” IEEE Trans. Control Syst. Technol., vol. 23,
no. 5, pp. 1754–1769, Sep. 2015.
R EFERENCES [23] T. X. Dinh, T. D. Thien, T. H. V. Anh, and K. K. Ahn, “Disturbance
observer based finite time trajectory tracking control for a 3 DOF
[1] Y. Zhu, J. Qiao, and L. Guo, “Adaptive sliding mode disturbance hydraulic manipulator including actuator dynamics,” IEEE Access,
observer-based composite control with prescribed performance of space vol. 6, pp. 36798–36809, 2018.
manipulators for target capturing,” IEEE Trans. Ind. Electron., vol. 66, [24] L. Kong, W. He, W. Yang, Q. Li, and O. Kaynak, “Fuzzy approximation-
no. 3, pp. 1973–1983, Mar. 2019. based finite-time control for a robot with actuator saturation under time-
[2] J. Baek, M. Jin, and S. Han, “A new adaptive sliding-mode con- varying constraints of work space,” IEEE Trans. Cybern., vol. 51, no. 10,
trol scheme for application to robot manipulators,” IEEE Trans. Ind. pp. 4873–4884, Oct. 2021, doi: 10.1109/TCYB.2020.2998837
Electron., vol. 63, no. 6, pp. 3628–3637, Jun. 2016. [25] W. He, H. Huang, and S. S. Ge, “Adaptive neural network control of a
[3] Z. Chen, F. Huang, C. Yang, and B. Yao, “Adaptive fuzzy backstep- robotic manipulator with time-varying output constraints,” IEEE Trans.
ping control for stable nonlinear bilateral teleoperation manipulators Cybern., vol. 47, no. 10, pp. 3136–3147, Oct. 2017.
with enhanced transparency performance,” IEEE Trans. Ind. Electron., [26] C. P. Bechlioulis and G. A. Rovithakis, “Robust adaptive control
vol. 67, no. 1, pp. 746–756, Jan. 2020. of feedback linearizable MIMO nonlinear systems with prescribed
[4] M. Yuan, Z. Chen, B. Yao, and X. Liu, “Fast and accurate motion track- performance,” IEEE Trans. Autom. Control, vol. 53, no. 9,
ing of a linear motor system under kinematic and dynamic constraints: pp. 2090–2099, Oct. 2008.
An integrated planning and control approach,” IEEE Trans. Control Syst. [27] C. P. Bechlioulis and G. A. Rovithakis, “Adaptive control with guaran-
Technol., vol. 29, no. 2, pp. 804–811, Mar. 2021. teed transient and steady state tracking error bounds for strict feedback
[5] J. Na, M. N. Mahyuddin, G. Herrmann, X. Ren, and P. Barber, systems,” Automatica, vol. 45, no. 2, pp. 532–538, Feb. 2009.
“Robust adaptive finite-time parameter estimation and control for [28] J. Na, Q. Chen, X. Ren, and Y. Guo, “Adaptive prescribed performance
robotic systems,” Int. J. Robust Nonlinear Control, vol. 25, no. 16, motion control of servo mechanisms with friction compensation,” IEEE
pp. 3045–3071, Nov. 2015. Trans. Ind. Electron., vol. 61, no. 1, pp. 486–494, Jan. 2014.

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: NN-BASED ADAPTIVE ASYMPTOTIC PRESCRIBED PERFORMANCE TRACKING CONTROL 11

[29] Q. Guo, Y. Zhang, B. G. Celler, and S. W. Su, “Neural adaptive back- Hang Zhou received the B. Tech. degree in mechan-
stepping control of a robotic manipulator with prescribed performance ical engineering from the Nanjing University of
constraint,” IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 12, Science and Technology, Nanjing, China, in 2021.
pp. 3572–3583, Dec. 2019. He is currently pursuing the Ph.D. degree in aero-
[30] C. Makkar, W. E. Dixon, W. G. Sawyer, and H. Gu, “A new continu- nautical and astronautical science and technology
ously differentiable friction model for control systems design,” in Proc. with the School of Aeronautics and Astronautics,
IEEE/ASME Int. Conf. Adv. Intell. Mechatronics, Jul. 2005, pp. 600–605. Shanghai Jiao Tong University, Shanghai, China.
[31] W. Sun, Y. Zhang, Y. Huang, H. Gao, and O. Kaynak, “Transient- His current research interests include servo con-
performance-guaranteed robust adaptive control and its application to trol of mechatronic systems, hydraulic robot control,
precision motion control systems,” IEEE Trans. Ind. Electron., vol. 63, and intelligent control.
no. 10, pp. 6510–6518, Oct. 2016.
[32] Z. Yao, J. Yao, and W. Sun, “Adaptive RISE control of hydraulic systems
with multilayer neural-networks,” IEEE Trans. Ind. Electron., vol. 66,
no. 11, pp. 8638–8647, Nov. 2019.
[33] H. Khalil, Nonlinear Systems. Upper Saddle River, NJ, USA:
Prentice-Hall, 2002.
[34] F. L. Lewis, S. Jagannathan, and A. Yesildirek, Neural Network Control
of Robot Manipulators and Nonlinear Systems. Philadelphia, PA, USA:
Taylor Francis, 1999.
Jin Zhou received the B. Tech. degree in mechan-
[35] A. Mohanty and B. Yao, “Indirect adaptive robust control of hydraulic
ical engineering from the Nanjing University of
manipulators with accurate parameter estimates,” IEEE Trans. Control
Science and Technology, Nanjing, China, in 2018,
Syst. Technol., vol. 19, no. 3, pp. 567–575, May 2011.
and the M.E. degree in mechanical engineering
[36] P. M. Patre, W. MacKunis, K. Kaiser, and W. E. Dixon, “Asymptotic
from the School of Mechanical Engineering, Nanjing
tracking for uncertain dynamic systems via a multilayer neural network
University of Science and Technology, in 2021.
feedforward and RISE feedback control structure,” IEEE Trans. Autom.
He is currently an Engineer with the Nanjing
Control, vol. 53, no. 9, pp. 2180–2185, Oct. 2008.
Engineering Institute of Aircraft Systems, Jincheng,
[37] Y.-S. Hao, Z.-G. Su, and X. Wang, “Finite-time output feedback con-
AVIC, Nanjing. His current research interests
trol for a rigid hydraulic manipulator system,” Math. Problems Eng.,
include servo control of mechatronic systems,
vol. 2018, Jul. 2018, Art. no. 9316562.
hydraulic robot control, and robust control.
[38] Q. Guo, T. Yu, and D. Jiang, “Robust H∞ positional control of 2-DOF
robotic arm driven by electro-hydraulic servo system,” ISA Trans.,
vol. 59, pp. 55–64, Nov. 2015.

Jianyong Yao (Member, IEEE) received the


B. Tech. degree in mechatronics from Tianjin
Wenxiang Deng received the B. Tech. degree University, Tianjin, China, in 2006, and the Ph.D.
in mechanical engineering from Central South degree in mechatronic engineering from Beihang
University, Changsha, China, in 2013, and the Ph.D. University, Beijing, China, in 2012.
degree in mechanical engineering from the Nanjing He was a visiting exchange student with
University of Science and Technology, Nanjing, the School of Mechanical Engineering, Purdue
China, in 2018. University, West Lafayette, IN, USA, from
He is currently a Lecturer with the School October 2010 to October 2011. In 2012, he joined
of Mechanical Engineering, Nanjing University the School of Mechanical Engineering, Nanjing
of Science and Technology. His current research University of Science and Technology, Nanjing,
interests include servo control of mechatronic China, where he is currently a Full Professor. His current research interests
systems, hydraulic robot control, robust adaptive include servo control of mechatronic systems, adaptive and robust control,
control, and nonlinear compensation. and fault detection and accommodation of dynamic systems.

Authorized licensed use limited to: Kyunghee Univ. Downloaded on June 11,2022 at 09:34:16 UTC from IEEE Xplore. Restrictions apply.

You might also like