IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO.

2, FEBRUARY 2019 755

Neural Network Control of a Two-Link Flexible

Robotic Manipulator Using Assumed
Mode Method
Hejia Gao, Student Member, IEEE, Wei He , Senior Member, IEEE, Chen Zhou, and Changyin Sun

Abstract—In this paper, the n-dimensional discretized control of flexible manipulator [2]. Various types of discretized
model of the two-link flexible manipulator is developed by models have been proposed via finite element method [3], as-
the assumed mode method (AMM). Subsequently, based sumed mode method (AMM) [4], finite difference method, and
on the discretized dynamic model, both full-state feedback
control and output feedback control are investigated to lumped parameter method [5]. In [6], Celentano and Coppola
achieve the trajectory tracking and vibration suppression. proposed a computationally efficient method for modeling flex-
In order to guarantee the stability strictly, uniform ultimate ible robots based on the assumed modes method. In [7], Celen-
boundedness (UUB) of the closed-loop system is realized tano presented a numerically stable and computationally effi-
by the Lyapunov’s stability. Furthermore, through appropri-
cient method, which allows the modeling and simulation of a
ately choosing control parameters, the states of the system
will converge to zero within a small neighborhood. Eventu- flexible robot with high accuracy. The appropriate discretized
ally, extensive simulations and experiments on the Quanser model can ensure the accurate positioning and tracking of the
platform for a two-link robotic manipulator are carried out to flexible manipulator end during the movement.
demonstrate the feasibility of the proposed neural network In this paper, the two-link flexible manipulator is modeled
controller. by the AMM. We assumed that the elastic deflection is small,
Index Terms—Assumed mode method (AMM), dynamic then, the two-link flexible manipulator can be expressed as a
modeling, flexible robotic manipulator, flexible structure, summation of finite number of modes [8]. The proposed discrete
neural networks (NN), vibration control, two-link. model can effectively overcome the hidden defects in the data
and make the model result more stable. It is helpful to describe
I. INTRODUCTION the nonlinear relationship.
According to the proposed dynamic model, the control de-
N RECENT years, the flexible robotic manipulator as the
I most direct application object of a flexible multibody system
dynamics analysis and control theory [1], because of its obvious
sign can be carried out to achieve the control objectives that the
joint motion can converge to the desired position [9] and the
elastic vibrations can be effectively suppressed [10]. In litera-
physical model and the easily tested features, has developed into tures, many control strategies have been proposed and applied
a key issue for the robot, aviation, and aerospace technology. to various systems [11]–[17]. In [18], the boundary control was
The research on the flexible manipulator content is divided into proposed to solve the boundary external disturbance and input
two aspects: one is dynamic modeling of flexible manipulators; dead-zone nonlinearity. In [19], an effective all-purpose adaptive
the other one is the control strategy design. fuzzy controller was proposed for the overhead cranes with non-
The flexible manipulator is a system with complicated dy- linear disturbance. In [20], robust tracking control was applied
namics. The dynamic model is the foundation of realizing the to electrically driven flexible-joint robot by voltage control. In
[21], an adaptive neural control method was proposed for robotic
Manuscript received September 27, 2017; revised December 27, manipulators with output constraints and uncertainties. In [22]
2017; accepted March 1, 2018. Date of publication March 22, 2018; and [23], a novel adaptive fuzzy control approach was devel-
date of current version February 1, 2019.This work was supported by the oped for nonlinear systems. In [24], experiments on the Baxter
National Natural Science Foundation of China under Grant 61522302,
Grant 61761130080, and Grant U1713209, and the Fundamental Re- Robot have been carried out to demonstrate the performance of
search Funds for the China Central Universities of USTB under Grant the proposed neural control. In [25], Dai et al. proposed a neu-
FRF-BD-17-002A and Grant FRF-TP-15-005C1. Paper no. TII-17-2232. ral network (NN) control strategy to solve the learning control
(Corresponding author: Wei He.)
H. Gao, W. He, and C. Zhou are with the School of Automation and problem for ocean surface ship.
Electrical Engineering, and Key Laboratory of Knowledge Automation for However, compared with these conventional control strate-
Industrial Processes, Ministry of Education, University of Science and gies, NN requires relatively less dynamic information. Also, in
Technology Beijing, Beijing 100083, China (e-mail: [email protected]).
C. Sun is with the School of Automation, Southeast University, Nanjing dealing with the uncertain nonlinear system [26], [27], it has al-
210096, China. ready been proved to be effective [28]. Under certain conditions,
Color versions of one or more of the figures in this paper are available artificial neural networks (ANNs) have been widely used in the
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TII.2018.2818120 control system because of its inherent nonlinear simulation abil-

1551-3203 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 756 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 2, FEBRUARY 2019

ity, high adaptability, and fault tolerance [29]. Moreover, NNs

have learning capability which makes NNs more widely em-
ployed for the controller design of various nonlinear systems,
such as [30], [31], and [32].
In this paper, aiming at handling the system uncertainties and
improving the robustness, an NN controller is developed for
a two-link flexible robotic manipulator based on radial basis
function neural networks (RBFNNs). Based on RBFNNs, the
flexible robotic manipulator system has strong robustness, mem-
ory ability, nonlinear mapping ability, and self-learning ability.
Compared with the existing works, the main contributions of
this paper are as follows.
1) Based on the n-dimensional ordinary differential equation
(ODE) model, NN control torques by full state feedback Fig. 1. Diagram of the two-link flexible manipulator.
are designed and implemented on double motors mounted
at the joint of the flexible link to achieve trajectory track-
ing and vibration suppression.
2) Output feedback control method and a high-gain observer dinate along the axial center of the ith link. We use derivative
˙ = ∂∂ ∗t , (¨∗) = ∂∂ (∗)
2

are adopted to estimate the unmeasured signals directly. notions throughout this paper as (∗) t 2 , (∗) =
, (∗) = ∂∂ x(∗) = ∂∂ x(∗) = ∂∂ x(∗)
2 3 4
Besides, neural network controller is designed to ensure ∂∗  
∂ xi 2 , (∗) 3 , and (∗) 4 .
the stability of the flexible system theoretically. i i i
The dynamic modeling needs to consider the coupling be-
3) Extensive experiments on Quanser Flexible Link Sys-
tween the first link and the second link. Based on the energy
tem platform for a two-link robotic manipulator are im-
equation and the Hamilton’s principle [34], we can get the cor-
plemented to verify the feasibility of the proposed NN
responding governing equations as follows [35]:
control in engineering practice.
The rest of the paper is organized as follows. The dynamic
modeling is derived in Section II. Adaptive neural network con- ρ1 Y¨1 (x1 ) = −EI1 ω1 (x1 ), (1)
trol of a two-link flexible manipulator by full-state and output
feedback is proposed in Sections III and IV. Simulations and ρ2 Y¨2 (x2 ) = −EI2 ω2 (x2 ). (2)
experiments based on the Quanser platform are implemented
to illustration the performance of the proposed controllers in The motion of the angular position are as follows:
Sections V and VI, respectively. Conclusions are given in
Section VII.
τ1 = Ih1 θ¨1 (t) − EI1 ω1 (0),

II. DYNAMIC MODELING (It1 + Ih2 )τ2 = It1 Ih2 θ¨2 (t) − Ih2 EI1 ω1 (L1 ) (3)
In this section, we make some assumptions before the dynam- − It1 EI2 ω2 (L2 ). (4)
ics modeling. In the premise of small deformation, the influence
of shear deformation, the moment inertia of the neutral axis, and
The boundary conditions are as follows:
the manipulator’s gravity is ignored. Considering the motion of
two-degree-of-freedom (2 DOF) manipulator in the horizontal
plane, we regard it as an Euler-Bernoulli beam in the modeling g(t)Y¨1 (L1 ) + ġ(t)Y˙1 (L1 ) + EI2 v˙1 (t) = EI1 ω1 (L1 ), (5)
process [33]. Then, the manipulator system will be discretized
via AMM in the following part. ρ2 L2 [Y˙1 (L1 )]2 sin θ2 (t) cos θ2 (t) − EI1 ω1 (L1 )
Fig. 1 shows the structure of the two-link manipulator sys- + EI2 Y˙1 (L1 )v2 (t) sin θ2 (t) + EI2 ω2 (0)
tem. The frame XOY and xi Oyi represent the inertia frame
and the local rotating reference frame with the hub, respec- − Ih2 [θ¨1 (t) + θ¨2 (t)] − It1 [θ¨1 (t) + ω¨1  (L1 )]
tively. Li represents the length of ith link. ρi is the mass per
unit length. Ihi denotes the inertia of the ith hub. Iti is the tip − m2 Y˙1 (L1 ) sin θ2 (t)[2L2 θ˙2 (t) + ω˙2 (L2 )] = 0, (6)
inertia of the ith link. Ioi is the inertia of ith link. Mi is the m2 Y¨2 (L2 ) = EI2 ω2 (L2 ), (7)
mass of the ith link. mi represents the tip mass. θi (t) denotes
the angular position of the ith hub. EIi is the bending stiff- It2 [θ¨1 (t) + θ¨2 (t) + ω¨2  (L2 )] + EI2 ω2 (L2 ) = 0, (8)
ness of the ith flexible link. wi (xi , t) represents the transverse
ωi (0) = ωi (0) = 0, i = 1, 2 (9)
movement of point i at the position xi for time t of the ith
link. Yi (xi , t) represents the position of the ith flexible link
with respect to frame XOY at the position xi for time t. τi (t) where, in order to facilitate writing, we define the correspond-
donates the input torque at the ith hub. xi ∈ [0, Li ] is the coor- ing coefficients in the boundary conditions g(t), v1 (t), v2 (t) as

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 GAO et al.: NEURAL NETWORK CONTROL OF A TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR USING ASSUMED MODE METHOD 757

follows:

Y1 (x1 ) = x1 θ1 (t) + ω1 (x1 ),

 t
Y2 (x2 ) = x2 θ2 (t) + ω2 (x2 ) + [Y˙1 (L1 , ξ) cos θ2 (ξ)]dξ,
0

g(t) = (M2 + m2 ) sin2 θ2 (t) + m1 + M2 ,

 t
v1 (t) = cos θ2 (t) ω2 (0, ξ)dξ,
0 Fig. 2. Diagram of full-state feedback.
 t
v2 (t) = [ω2 (L2 , ξ) − ω2 (0, ξ)]dξ.
0 The gain matrix K1 = K1T > 0 ∈ R(φ+2)×(φ+2) , then
Due to the flexible characteristics of the manipulator, elastic ż1 = z2 − K1 z1 , (14)
deformation will occur during the movement of the manipulator.
And it requires satisfying the appropriate boundary conditions ż2 = B −1 (q)[τ − H(q, q̇)q̇ − K(q)] − α̇1 (t). (15)
and governing equations. ωi is the elastic deformation of link i For achieving the stability of the manipulator system, we con-
and can be described via AMM: struct a Lyapunov function. Consider a Lyapunov function can-

μi didate as
ωi (xi , t) = Fij (xi )pij (t), i = 1, 2 (10) 1
j =1 V1 = z1T z1 . (16)
2
where Fij (xi ) denotes the modes shape function, and pij (t) is Further, the time derivative of (16)
the generalized coordinate. V˙1 = z1T z2 − z1T K1 z1 . (17)
On the premise of small deflection, the dynamic model de-
rived by Lagrange’s equation and AMM is as follows [36]: Because we have obtained the manipulator system ż1 and ż2 ,
we can guarantee the system stability by removing the coupling
B(q)q̈ + H(q, q̇)q̇ + K(q) = τ (t) (11) term z1T z2 in (17). Therefore, the Lyapunov function candidate
is introduced as
where q = [θ, p]T , θ = [θ1 , θ2 ]T is the vector of deflection 1 1
angle, p = [p11 . . . p1n 1 , p21 . . . p2n 2 ]T represents the flexible V2 = z1T z1 + z2T B(q)z2 . (18)
2 2
generalized coordinate vector. φ = n1 + n2 is the total num-
ber of flexible variables. τ ∈ R(φ+2) = [τ1 , τ2 , 0, . . . , 0]T rep- According to the definition of the skew-symmetric matrix,
resents the generalized control torque vector applied at each let A be a square matrix, if its transpose AT plus itself equals
joint. B(q) ∈ R(φ+2)×(φ+2) is the inertia matrix. H(q, q̇) ∈ exactly zero matrix, then A is called skew symmetric matrix.
R(φ+2)×(φ+2) represents the matrix of coriolis and centripetal Therefore, we obtain
forces. K(q) ∈ R(φ+2) is the matrix of gravity and elasticity −z2T Az2 = z2T AT z2 = z2T Az2 . (19)
forces.
According to (19), we know that = 0. Because
z2T Az2
2 (Ḃ(q) − 2H(q, q̇)) is a skew-symmetric matrix, we can state
1
III. CONTROL DESIGN FOR ADAPTIVE NN CONTROL WITH that
FULL-STATE FEEDBACK
1 T
In this section, considering full state information, q and q̇ z (Ḃ(q) − 2H(q, q̇))z2 = 0. (20)
2 2
are available, we introduce an ANN controller with full-state Differentiating (18) with respect to time, and substituting (20)
feedback to effectively track the state vector q. Lyapunov’s di- into this equation, we can obtain
rect method and back-stepping method are adopted to propose
the desired control and prove the stability of the system un- V˙2 = z1T z2 − z1T K1 z1
der the controller. The diagram of full-state feedback control is
+ z2T [τ − K(q) − H(q, q̇)α1 (t) − B(q)α̇1 (t)]. (21)
displayed in Fig. 2.
Considering the dynamical equation (11), we have the de- We need to make V̇2 < 0 by designing control torque τ0 when
scription of the system dynamics as B(q), H(q, q̇), and K(q) are available. Hence, the control input
is as follows:
q̈ = B −1 (q)[τ − H(q, q̇)q̇ − K(q)]. (12)
τ = −z1 − K2 z2 + H(q, q̇)α1 (t) + K(q) + B(q)α˙1 (t) (22)
We introduce two error variables
where the gain matrix K2 = K2T > 0.
z1 (t) = q(t) − qd (t), z2 (t) = q̇(t) − α1 (t) (13) Substituting (22) into (21), we have

where α1 = −K1 z1 + q̇d is a virtual back-stepping control. V˙2 = −z1T K1 z1 − z2T K2 z2 (23)

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 758 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 2, FEBRUARY 2019

where V˙2 is negative definite and τ0 is feasible. Applying (25) into (31)
Due to the uncertainties of B(q), H(q, q̇), and K(q), the pro- V˙3 ≤ z1T z2 − z1T K1 z1 + z2T [τ − W ∗T R(Z) − (Z)]
posed control input cannot be applied to the actual manipulator
system. To address this challenge, we propose the following 
φ+2
T ˙
control to approximate the uncertain dynamics and polish up + W̃i Γ−1
i Ŵi . (32)
the control performance by the online estimation i=1

Applying the following inequalities:

τ = −z1 − K2 z2 + Ŵ T R(Z) (24)

n
T 
n
T
− σi W̃i Ŵi = − σi W̃i (W̃i + Wi∗ )
where Wi∗ , Ŵi , and W̃i are the true value, estimation, and error
i=1 i=1
of neural networks weights, respectively. R(Z) ∈ Rl is the basis
function. 
n
σi
The function approximate W ∗T R(Z) can be defined as fol- ≤− ( W̃i 2 −  Wi∗ 2 ), (33)
2
i=1
lows:
1
W ∗T R(Z) = H(q, q̇)α1 (t) + K(q) + B(q)α˙1 (t) − (Z) −z2T (Z) ≤( ¯(Z) 2 +z2T z2 ). (34)
2
(25) Substituting (24) and (26) into (32):
where Z = [q T , q̇ T , α1T , α˙1 T ] represents the inputs of the neural 1
V˙3 ≤ −z1T K1 z1 − z2T K2 − I(φ+2)×(φ+2) z2
networks and (Z) ∈ R(φ+2) is the approximation error. 2
Moreover, the adaptive law is represented as follows:
1  σi φ+2
˙
Ŵi = −Γi [Si (Z)z 2 ,i + σi Ŵi ] (26) +  ¯(Z) 2 − ( W̃i 2 −  Wi∗ 2 )
2 2
i=1
where Γi is the constant gain matrix, and σi is a small positive ≤ −γ1 V3 + β1 (35)
constant in order to improve robustness.
Remark 1: In this paper, a more challenging problem that where
B(q), H(q, q̇), and K(q) are totally unavailable is studied. If 2λm in (K2 − 12 I(φ+2)×(φ+2) )
the individual term can be obtained in practical engineering, it γ1 = min 2λm in (K1 ), ,
λm ax (B)
can be removed from (25) and rewritten explicitly in (24).
Theorem 1: For the dynamic equation (11), the control σi
min , (36)
law (24) with the adaptive law (26), because full state i=1,...φ+2 Γ−1
i
information is known, for each compact set Ω0 where
(q(0), q̇(0), Ŵ1 (0), . . . , Ŵφ+2 (0)) ∈ Ω0 , the initial conditions 
φ+2
σi 1
are bounded, UUB of the closed-loop system trajectories is β1 =  Wi∗ 2 +  ¯(Z) 2 (37)
2 2
i=1
achieved. z1 , z2 , and W̃ will remain within the compact sets
Ωz 1 , Ωz 2 , ΩW where λm ax (A) and λm in (A) donate the maximum and min-
 √  imum Eigenvalues of matrix A, respectively. To guarantee γ1
Ωz 1 : = z1 ∈ Rφ+2 | z1 ≤ M , (27) more than zero, K1 and K2 are required to satisfy
   1
M λm in (K1 ) > 0, λm in K2 − I > 0. (38)
Ωz 2 : = z2 ∈ R φ+2
| z2 ≤ , (28) 2
λm in (B)
Multiplying (35) by eγ t :
  
M d
ΩW : = W̃ ∈ R l×(φ+2)
| W̃ ≤ (29) (V3 eγ 1 t ) ≤ β1 eγ 1 t . (39)
λm in (Γ−1 ) dt
Integrating both sides of (39)
where M = 2(V (0) + βγ ), β > 0 and γ > 0. β1 t
Proof: Construct a Lyapunov function as V3 eρ 1 t |t0 ≤ |, (40)
γ1 0
1  T −1
φ+2
1 1 β1 β1 β1
V3 = z1T z1 + z2T Bz2 + W̃i Γi W̃i . (30) V3 ≤ V3 (0) − e−γ 1 t + ≤ V3 (0) + . (41)
2 2 2
i=1
γ1 γ1 γ1
Then, we have
Differentiating (30), and applying (21), we obtain
1 β1
V˙3 ≤ z1T z2 − z1T K1 z1 + z2T [τ − H(q, q̇)α1 (t)  z1 2 ≤ V3 (0) + . (42)
2 γ1

φ+2
T ˙
Hence, we can get the conclusion that zi converges to the
− K(q) − B(q)α̇1 (t)] + W̃i Γ−1
i Ŵi . (31) compact set Ωz i and W̃i converges to ΩW . The system is stabi-
i=1 lized via Lyapunov function (30) that we designed.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 GAO et al.: NEURAL NETWORK CONTROL OF A TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR USING ASSUMED MODE METHOD 759

Proof: Construct a Lyapunov function as

1  T −1
φ+2
1 1
V4 = z1T z1 + z2T Bz2 + W̃i Γi W̃i . (49)
2 2 2
i=1

The time derivation of (49) with (44) and (45)

1
V˙4 ≤ −z1T K1 z1 − z2T K2 − I z2 − z2T K2 z˜2
2
Fig. 3. Diagram of output feedback

φ+2
T 1
+ z2,i [Ŵi Ri (Ẑ) − Wi∗T Ri (Z)] +  ¯ 2
2
i=1
Remark 2: It is apparent that the control performance will
be better if the control gains K1 and K2 and the adaptive gains 
φ+2
T T
Γi are increased. Hence, we can obtain the conclusion that the − [W̃i Ri (Ẑ)zˆ2 ,i + σi W̃i Ŵi ]. (50)
tracking errors z1 and z2 will finally converge to a small neigh- i=1

borhood of zero by appropriately design parameters.

Applying the equation z̃2 = ξ2 = −εϑ(2) , where ε is an any
small positive constant, ϑ = π2 + λ̄1 π1 , and there exist positive
IV. CONTROL DESIGN FOR ADAPTIVE NN CONTROL WITH constants t∗ and h2 such that ∀t > t∗ , we have  ξ2 ≤ εh2 .
OUTPUT FEEDBACK Hence, we can use πε2 to estimate q̇, then q̇ and z2 can be
The proposed full-state feedback control (24) requires q and estimated
q̇ both to be available. On the premise of the speed sensors, a π2 π2
high-gain observer is employed to estimate q̇. The diagram of x̂2 = , ẑ2 = − α1 . (51)
ε ε
output feedback is shown in Fig. 3.
According to the Lemma from [37], πεk k+ 1 converges asymp- Using the properties from [38], we can use the following
totically to the kth order derivative of q. ξk is closed to zero in inequalities and equations:
a short time owing to the high-gain ε1 and the boundedness of q T σi
and its k derivatives. Thus, πεk k+ 1 is proper as an observer. The −σi W̃i Ŵi ≤ − ( W̃i 2 −  Wi∗ 2 ), (52)
2
estimate of the unmeasurable state vector z2 is defined as
 Ri (Ẑ) 2 ≤ li , (53)
π2 π2
q = π1 , q̇ = , ẑ2 = − α1 (43) T T
ε ε Ŵi Ri (Ẑ) = Wi∗T Rti + Wi∗T Ri (Z) + W̃i Ri (Z) (54)
where επ˙1 = π2 , επ˙2 = −λ¯1 π2 − π1 + x1 . where li > 0, > 0, and Rt is a bounded vector function.
Update the control torque (24) and adaptive law (26) Substituting the following inequalities:
τ = −z1 − K2 ẑ2 + Ŵ T R(Ẑ), (44) 
n
1 T   W ∗ 2 2  Rti 2
n
z2,i Wi∗T Rti ≤ z2 z2 + i
(55)
˙ 2 2
Ŵi = −Γi (Ri (Ẑ)ẑ 2 ,i + σi Ŵi ). (45) i=1 i=1

Theorem 2: For the dynamic model (11), the controller and

(44) with the adaptive law (45) , because the output in- n √ √

n  σi W̃i
T
2Ri (Ẑz˜2 ,i )
formation is available, for each compact set Ω0 where −
T
W̃i Ri (Ẑ)z˜2 ,i = − √ √
(q(0), q̇(0), Ŵ1 (0), Ŵ2 (0), . . . , Ŵφ+2 (0)) ∈ Ω0 , the initial i=1 i=1
2 σi
conditions are bounded, UUB of the closed-loop system trajec-

n
σi  W̃i 2  2  Ri (Ẑ) 2 1 T
n
tories are achieved. z1 , z2 , and W̃ will remain within Ωz 1 , Ωz 2 , ≤ + z˜2 z˜2 . (56)
and ΩW i=1
4 σi 2
i=1
 √ 
Ωz 1 : = z1 ∈ Rφ+2 | z1 ≤ M , (46) We can obtain
   1
M V˙4 ≤ −z1T K1 z1 − z2T (K2 − I)z2 − z2T K2 z˜2 +  ¯ 2
Ωz 2 : = z2 ∈ R φ+2
| z2 ≤ , (47) 2
λm in (B)

φ+2
σi  2  Ri (Ẑ) 2 1
φ+2
   −  W̃i 2 + z˜2 T z˜2
M 4 σi 2
ΩW : = W̃ ∈ R l×(φ+2)
| W̃ ≤ (48) i=1 i=1
λm in (Γ−1 )

φ+2
σi 2
 Rti 2
+ +  Wi∗ 2 . (57)
where M = 2(V4 (0) + γ 2 ),
β2
β2 > 0, and γ2 > 0. i=1
2 2

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 760 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 2, FEBRUARY 2019

Using (52) and z2T K2 z˜2 ≤ 12 z2T z2 + 12 (K2 z˜2 )T (K2 z˜2 ), we fur- TABLE I
PROPERTIES OF THE TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR
ther have
3 1
V˙4 ≤ −z1T K1 z1 − z2T K2 − I z2 +  ¯ 2
2 2

φ+2
σi 2li 1 T
−  W̃i 2 +λm ax K2T K2 + diag ξ ξ2
4 σi 2 2
i=1

1 
φ+2

+ σi + 2
 Rti 2  Wi∗ 2 . (58)
2
i=1

According to the Lemma from [37], applying 12 ξ2T ξ2 ≤ 12 ε2 h22 ,

we have:
3 1
V˙4 ≤ −z1T K1 z1 − z2T K2 − I z2 +  ¯ 2
2 2

φ+2
σi 2li 1 2 2
−  W̃i 2 +λm ax K2T K2 + diag ε h2
4 σi 2
i=1

1 
φ+2

+ σi + 2
 Rti 2  Wi∗ 2
2
i=1

≤ −γ3 V4 + β3 (59)
where γ3 and β3 are two constants defined as
2λm in (K2 − 32 I)
γ3 = min 2λm in (K1 ), ,
λm ax (B)

σi
min , (60)
i=1,2...φ+2 2λm ax (Γ−1
i )

1 
φ+2
 1
β3 = 2
 Rti 2 +σi  Wi∗ 2 +  ¯ 2
2 2
i=1

2li 1 2 2
+ λm ax K2T K2 + diag ε h2 . (61) Fig. 4. Tracking trajectory and control torque for open-loop system.
σi 2
To guarantee γ3 > 0, K1 and K2 are required to satisfy
A. Simulation Results for Open-Loop System
3
λm in (K1 ) > 0, λm in K2 − I > 0. (62) In this section, time domain responses of the discrete model
2
are observed through MATLAB simulations. Give the system a
Remark 3: Under the hypothesis that the position measure-
small interference
ments are perfect, by employing a high-gain observer, we de- ⎧
signed a rigorous theoretical treatment of the output feedback ⎪
⎨ 0.5 N · m, 0 s ≤ t ≤ 0.5 s
problem. The tracking error converges and remains within a τ1 (t) = τ2 (t) = −0.5 N · m, 0.5 s ≤ t ≤ 1 s . (63)
small neighborhood of zero. If the residual error is lower than ⎪
⎩
0 N · m, otherwise
expected, we can reduce the residual error by decreasing βγ or
increasing control gains K1 , K2 , the high-gain ε1 and the ap- The reference inputs of the first and the second link were
proximation accuracy of the neural networks. set to θ1d = 0.5 rad, θ2d = 0.25 rad, respectively. The tracking
trajectory θi of the first and second link for open-loop system
V. SIMULATIONS are shown as Fig. 4. We can see that the outputs of the system
are divergent.
Simulations of open-loop system will be shown in
Section V-A; a PD controller will be introduced in Section V-B;
B. Simulation Results for PD Control
neural network control will be revealed in Section V-C, and the
comparison will be given in Section V-D. The system parame- We designed two vibration control structures to reduce the
ters are specified in Table I. vibration of the two-link flexible manipulator model. At first, we

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 GAO et al.: NEURAL NETWORK CONTROL OF A TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR USING ASSUMED MODE METHOD 761

Fig. 5. Tracking trajectory and control torque for PD control. Fig. 7. Tracking trajectory and control torque for NN control.

Fig. 6. Tip position for PD control. Fig. 8. Tip position for NN control.

C. Simulation Results for Neural Network Control

applied two separate PD controllers to achieve the tip tracking
control performance of the two flexible links. Select the initial We choose the RBFNN system parameters as: the center pa-
states as zero. Design the PD controller rameters is selected as 1 or −1, 28 nodes are employed for each
Si (Zi ). Variance of centers η = 2 as well as the initial weights
τi = −Kp (θi − θid ) − Kd θ˙i , i = 1, 2 (64) Ŵi = 0 (i = 1, 2, . . . , 256), σ = 0.0002, Γ = 100I256×256 . We
introduce a neural network controller with K1 = 2I6×6 , K2 =
where Kp and Kd are proportional gain and differential gain.
8I6×6 . The tracking trajectory θi and the control torque τi of the
In the simulation model, the reference inputs of the first and
first and second link are shown as Fig. 7. Tip position in 3-D
the second link were set to θ1d = 0.5 rad, and θ2d = 0.25 rad,
of the two links is shown as Fig. 8. We can conclude that the
respectively. We introduce a PD controller with Kp = 5, Kd =
angular displacement signal θi is stable at 1 s, and the overshoot
1.5. The tracking trajectory θi and the control torque τi of the
and vibration are removed in the tracking process.
first and second link are shown as Fig. 5. We can conclude that
the angular displacement signal θi is stable at 3 s, but there is
a large overshoot within the range of − 0.2 to 0.2 rad in the D. Comparison Between PD and NN Simulation Results
tracking process. Tip position in 3-D of the two links is shown Notice that between the start of the motion of the second link
in Fig. 6. There is a large vibration of the first link within the and the stability of the second link in Figs. 7 and 5, there is a
range of − 0.05 to 0.05 m in the tracking process. visible vibration in the first link. The vibration is caused by the

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 762 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 2, FEBRUARY 2019

TABLE II
PROPERTIES OF THE QUANSER EXPERIMENTAL PLATFORM

Fig. 9. Control framework of the experimental flexible link.

Fig. 10 shows working principle diagram of flexible experi-

ment platform. The 2 DOF serial flexible ruler rotates are driven
by a DC motor (servo device). In the process of rotation, the
end deflection and the angular displacement signal are mea-
sured by strain sensor and optical encoder (4096 times per turn
in orthogonal mode), respectively. The derivatives of the an-
gular displacement signal and the end deflection are obtained
by a derivative filter. The control algorithm is designed by the
Fig. 10. Diagram of the experimental flexible link. Simulink module in PC. Then, the control signal is send to the
SRV02 servo device through Q8-usb data acquisition device and
voltPAQ-X1 amplifier. The relevant parameters of this platform
inertia of the first link when the second link stops its motion. are tabulated in Table II.
Using the PD controller, this vibration ranges from − 0.1 to The purpose of the experiment is to design the controller to
0.1 rad, which cannot be neglected in practical engineering. In track the robot position and minimize the end vibration in the
addition, there exists an overshoot within the range of − 0.2 horizontal plane. The planar tracking or position adjusting of the
to 0.2 rad for the first link in tracking the desired trajectory end member is obtained by the motion of a series mechanism
process. These vibrations will bring irreparable consequences with two flexible joints. For the MIMO system, the decoupling
to practical projects. Therefore, the NN controller is employed controller is used to suppress the vibration due to the flexibility
in the two-link flexible manipulator. Seen from the simulation and the coupling effect between the two links.
results, we can conclude that each of the two flexible links is able
to track the desired angular position trajectories in the form of B. Experimental Results
square waves. However, compared with the PD controller, under In this section, the PD controller and NN controller designed
the NN controller, the flexible manipulator reaches the desired by MATLAB Simulink models and implemented on the Quanser
position in a shorter time. Furthermore, the NN controller works two-link flexible plant are presented. In order to compare the
better in vibration suppression and the overshoot is removed in control performance adequately and intuitively, the superim-
the tracking process. posed curves of the two controllers are given.
A decoupled model is employed in the control system, which
VI. EXPERIMENT is to say that link coupling of the serial mechanism is neglected.
In the experiment, we chose square waves as input signals.
A. Introduction of the Experimental Testbed Through proper amplitude, degrees to radians and time-delay,
In order to verify the feasibility of the control strategy in en- the square waves are converted to analog outputs. After power
gineering practice, the simulation experiment is implemented amplification, the outputs are applied to SRV02 to drive the
in the Quanser Flexible Link System. From Fig. 9, the flexible flexible manipulator. Both drives are commanded independently
rotating manipulator consists of double motors driven gapless of each other, each using a separate control loop. The rotation
harmonic gear reducer and two series connected flexible links. angle θ and the deflection α measured by the encoder and strain
The first stage flexible link is connected to the first motor and gauge sensor can be observed in real time plot. The derivatives
equipped with a strain sensor to measure the end displacement. of the angular displacement signal and the end deflection θ̇, α̇
In addition, the end is connected to the second motor. The sec- are obtained by a derivative filter.
ond flexible link with strain sensor is connected to the second We designed two vibration control structures to reduce the vi-
motor. The displacement of the two motors is measured by an bration of the two-link flexible manipulator. At first, we applied
incremental photoelectric encoder. two separate PD controllers to achieve the control performance.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 GAO et al.: NEURAL NETWORK CONTROL OF A TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR USING ASSUMED MODE METHOD 763

Fig. 11. Tracking trajectory with PD and NN control. Fig. 13. Control torques with PD and NN control.

ever, the end-effect (α1 and α2 ) of the flexible manipulator using

PD controller remains 0.6◦ . And the links reach the desired po-
sition in 2 s, with a large overshoot whose max amplitude is 4◦ .
In order to eliminate the flexible vibration, we designed the
NN controller in the control system. The inputs of the radial
basis function (RBF) network contain the tracking errors of two
links angle position, the vibration of the tip payload, and their
derivatives. Given the same inputs as the PD models, we choose
the RBF neural network parameters as: the center parameters
is selected as 1 or −1, 28 nodes are employed for each Si (Zi ).
Variance of centers η = 2 as well as the initial weights Ŵi = 0
(i = 1, 2, . . . , 256). σ = 0.0002, Γ = 100I256×256 . We selected
the control parameters as K1 = 10, K2 = 0.3 in the first link,
and K1 = 12, K2 = 0.7 in the second link. The servo angle θi ,
the end deflection αi , and the control inputs τi are revealed in
Figs. 11, 12, and 13, respectively. Seen from the real time plot,
the links reach the desired position in 1 s and the large overshoot
is reduced to 1◦ . More importantly, the elastic vibration (α1 and
α2 ) are decreased from 0.6◦ to 0.05◦ .

C. Comparison Between PD and NN Experiment Results

Fig. 12. Elastic vibration with PD and NN control. Notice that between the start of the motion of the first link and
the stability of the first link in Fig. 11, there is a vibration range
from −4 to 4◦ in the second link and a vibration range from −2
In the Simulink model, the reference inputs of the first and the to 2◦ in the first link under PD controller. The vibration is caused
second link were set to 15π 10π
180 and 180 , respectively. The control by the inertia of the second link when the first link stops its mo-
parameters are selected as Kp = 5, Kd = 0.01 in the first link, tion, and the magnitude of the vibration is much bigger than that
and Kp = 8, Kd = 0.01 in the second link. The servo angle of other vibrations in tracking process. In practice, it will result
θi , the end deflection αi , and the control inputs τi are revealed in serious consequences. Therefore, we would design a more
in Figs. 11, 12, and 13, respectively. Seen from the real time effective control strategy like neural network control to solve
plot, each of the two flexible links is able to track the desired the inertial issue in this work. The tracking performance of NN
angular position trajectories in the form of square waves. How- controller is shown in Fig. 11. Compared with the PD controller,

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 764 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 15, NO. 2, FEBRUARY 2019

the vibration in the second link is reduced within the range of [11] D. Wang, H. He, and D. Liu, “Adaptive critic nonlinear robust control: A
−1◦ to 1◦ and the vibration in the first link is reduced within the survey,” IEEE Trans. Cybern., vol. 47, no. 10, pp. 3429–3451, Oct. 2017.
[12] W. He, X. He, and C. Sun, “Vibration control of an industrial moving strip
range of −0.5◦ to 0.5◦ in the tracking process. Furthermore, un- in the presence of input deadzone,” IEEE Trans. Ind. Electron., vol. 64,
der the NN controller, the flexible manipulator reaches the stable no. 6, pp. 4680–4689, Jun. 2017.
state in a shorter time. More importantly, seen form Fig. 12, the [13] W. He and S. S. Ge, “Robust adaptive boundary control of a vibrating
string under unknown time-varying disturbance,” IEEE Trans. Control
end-effect (α1 and α2 ) of the flexible are decreased from 0.6◦ to Syst. Technol., vol. 20, no. 1, pp. 48–58, Jan. 2012.
0.05◦ which is to say that the neural network controller in the [14] B. Luo, T. Huang, H. N. Wu, and X. Yang, “Data-driven h ∞ control
model works better in vibration suppression. for nonlinear distributed parameter systems,” IEEE Trans. Neural Netw.
Learn. Syst., vol. 26, no. 11, pp. 2949–2961, Nov. 2015.
[15] H. Wu and J. Wang, “Static output feedback control via PDE boundary and
VII. CONCLUSION ODE measurements in linear cascaded ODE–beam systems,” Automatica,
vol. 50, no. 11, pp. 2787–2798, 2014.
In this paper, the n-dimensional discretized model of the [16] B. Xu, D. Wang, Y. Zhang, and Z. Shi, “Dob based neural control of
flexible hypersonic flight vehicle considering wind effects,” IEEE Trans.
two-link flexible manipulator is achieved via the AMM which Ind. Electron., vol. 64, no. 11, pp. 8676–8685, Nov. 2017.
realizes the conversion from PDE to ODE and facilitates the [17] S. Dai, S. He, H. Lin, and C. Wang, “Platoon formation control with
analysis of dynamic uncertainty and the design of controller. prescribed performance guarantees for USVs,” IEEE Trans. Ind. Electron.,
vol. 65, no. 5, pp. 4237–4246, May 2018.
Subsequently, both full-state and output feedback NN controller [18] T. Jiang, J. Liu, and W. He, “Boundary control for a flexible manipulator
are investigated to track the desired trajectory and suppress the based on infinite dimensional disturbance observer,” J. Sound Vibration,
flexible vibration. UUB of the closed-loop system is realized vol. 348, pp. 1–14, 2015.
[19] C. Y. Chang, “Adaptive fuzzy controller of the overhead cranes with
by the Lyapunov’s stability. Furthermore, extensive simulations nonlinear disturbance,” IEEE Trans. Ind. Informat., vol. 3, no. 2, pp. 164–
and experiments on Quanser platform are implemented to 172, May 2007.
demonstrate the feasibility of the proposed NN controller, [20] M. M. Fateh, “Robust control of flexible-joint robots using voltage control
strategy,” Nonlinear Dyn., vol. 67, no. 2, pp. 1525–1537, 2012.
and the final performance has been compared with that of [21] S. Zhang, Y. Dong, Y. Ouyang, Z. Yin, and K. Peng, “Adaptive neural
PD control strategy. In the future, we will further study control for robotic manipulators with output constraints and uncertainties,”
more efficient modeling and control techniques to expand the IEEE Trans. Neural Netw. Learn. Syst., vol. pp, no. 99, pp. 1–11, 2018,
doi: 10.1109/TNNLS.2018.2803827.
applicability of the proposed method in practice and solve the [22] S. Tong, T. Wang, and Y. Li, “Fuzzy adaptive actuator failure compen-
vibration problems of various flexible structures. sation control of uncertain stochastic nonlinear systems with unmodeled
dynamics,” IEEE Trans. Fuzzy Syst., vol. 22, no. 3, pp. 563–574, Jun. 2014.
[23] S. C. Tong, Y. M. Li, G. Feng, and T. S. Li, “Observer-based adaptive fuzzy
ACKNOWLEDGMENT backstepping dynamic surface control for a class of MIMO nonlinear
systems,” IEEE Trans. Syst., Man, Cybern. B, vol. 41, no. 4, pp. 1124–
The authors thank the AE and the anonymous reviewers for 1135, Aug. 2011.
their constructive comments which helped improve the quality [24] C. Yang, Y. Jiang, Z. Li, W. He, and C.-Y. Su, “Neural control of bimanual
and presentation of this paper. robots with guaranteed global stability and motion precision,” IEEE Trans.
Ind. Informat., vol. 13, no. 3, pp. 1162–1171, Jun. 2017.
[25] S. Dai, M. Wang, and C. Wang, “Neural learning control of marine surface
REFERENCES vessels with guaranteed transient tracking performance,” IEEE Trans. Ind.
Electron., vol. 63, no. 3, pp. 1717–1727, Mar. 2016.
[1] W. Chen, “Dynamic modeling of multi-link flexible robotic manipulators,” [26] H. Dinh, R. Kamalapurkar, S. Bhasin, and W. Dixon, “Dynamic neural
Comput. Struct., vol. 79, no. 2, pp. 183–195, 2001. network-based robust observers for uncertain nonlinear systems,” Neural
[2] W. He, Z. Li, and C. L. Philip Chen, “A survey of human-centered in- Netw., vol. 60, pp. 44–52, 2014.
telligent robots: Issues and challenges,” IEEE/CAA J. Automatica Sinica, [27] S. Dai, C. Wang, and M. Wang, “Dynamic learning from adaptive neural
vol. 4, no. 4, pp. 602–609, Oct. 2017. network control of a class of nonaffine nonlinear systems,” IEEE Trans.
[3] T. Zebin and M. Alam, “Modeling and control of a two-link flexible Neural Netw. Learn. Syst., vol. 25, no. 1, pp. 111–123, Jan. 2014.
manipulator using fuzzy logic and genetic optimization techniques,” J. [28] Z. Li, Q. Ge, W. Ye, and P. Yuan, “Constrained multilegged robot sys-
Comput., vol. 7, no. 3, pp. 578–585, 2012. tem modeling and fuzzy control with uncertain kinematics and dynamics
[4] Y. Li, B. Tang, Z. Shi, and Y. Lu, “Experimental study for trajectory incorporating foot force optimization,” IEEE Trans. Syst., Man, Cybern.
tracking of a two-link flexible manipulator,” Int. J. Syst. Sci., vol. 31, Syst., vol. 46, no. 10, pp. 1338–1351, Jan. 2016.
no. 1, pp. 3–9, 2000. [29] K. S. Narendra and K. Parthasarathy, “Identification and control of dy-
[5] W. He, Y. Ouyang, and J. Hong, “Vibration control of a flexible robotic ma- namical systems using neural networks,” IEEE Trans. Neural Netw., vol. 1,
nipulator in the presence of input deadzone,” IEEE Trans. Ind. Informat., no. 1, pp. 4–27, Mar. 1990.
vol. 13, no. 1, pp. 48–59, Feb. 2017. [30] H. Gomi and M. Kawato, “Neural network control for a closed-loop system
[6] L. Celentano and A. Coppola, “A computationally efficient method for using feedback-error-learning,” Neural Netw., vol. 6, no. 7, pp. 933–946,
modeling flexible robots based on the assumed modes method,” Appl. 1993.
Math. Comput., vol. 218, no. 8, pp. 4483–4493, 2011. [31] T. W. Chow and Y. Fang, “A recurrent neural-network-based real-time
[7] L. Celentano, “Modeling of flexible robots with varying cross section and learning control strategy applying to nonlinear systems with unknown
large link deformations,” J. Dyn. Syst. Meas. Control, vol. 138, no. 2, pp. dynamics,” IEEE Trans. Ind. Electron., vol. 45, no. 1, pp. 151–161, Feb.
021010, 2015. 1998.
[8] R. K. Moolam, “Dynamic modeling and control of flexible manipulators,” [32] W. T. Miller III, R. P. Hewes, F. H. Glanz, and L. G. Kraft III, “Real-time
Ph.D. dissertation, Dept. Mech. Eng., Politecnico Di Milano, Milan, Italy, dynamic control of an industrial manipulator using a neural network-based
2014. learning controller,” IEEE Trans. Robot. Autom., vol. 6, no. 1, pp. 1–9,
[9] C. Sun and Y. Xia, “An analysis of a neural dynamical approach to solv- Feb. 1990.
ing optimization problems,” IEEE Trans. Autom. Control, vol. 54, no. 8, [33] F. Jin and B. Guo, “Lyapunov approach to output feedback stabilization
pp. 1972–1977, Aug. 2009. for the euler–bernoulli beam equation with boundary input disturbance,”
[10] X. He, W. He, J. Shi, and C. Sun, “Boundary vibration control of vari- Automatica, vol. 52, pp. 95–102, 2015.
able length crane systems in two dimensional space with output con- [34] X. Zhang, W. Xu, S. S. Nair, and V. Chellaboina, “PDE modeling and
straints,” IEEE/ASME Trans. Mechatronics, vol. 22, no. 5, pp. 1952–1962, control of a flexible two-link manipulator,” IEEE Trans. Control Syst.
Oct. 2017. Technol., vol. 13, no. 2, pp. 301–312, Mar. 2005.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.
 GAO et al.: NEURAL NETWORK CONTROL OF A TWO-LINK FLEXIBLE ROBOTIC MANIPULATOR USING ASSUMED MODE METHOD 765

[35] L. Zhang and J. Liu, “Observer-based partial differential equation bound- Chen Zhou received the B.Eng. degree in con-
ary control for a flexible two-link manipulator in task space,” IET Control trol theory and engineering from the School of
Theory Appl., vol. 6, no. 13, pp. 2120–2133, 2012. Automation and Electrical Engineering, Univer-
[36] A. D. Luca and B. Siciliano, “Closed-form dynamic model of planar sity of Science and Technology Beijing, Beijing,
multilink lightweight robots,” IEEE Trans. Syst., Man, Cybern., vol. 21, China, in 2017. He is currently working toward
no. 4, pp. 826–839, Jul./Aug. 1991. the M.S. degree with the School of Electrical
[37] S. Behtash, “Robust output tracking for non-linear systems,” Int. J. Con- and Computer Engineering, Georgia Institute of
trol, vol. 51, no. 6, pp. 1381–1407, 1990. Technology, Atlanta, GA, USA.
[38] R. M. Sanner and J.-J. E. Slotine, “Gaussian networks for direct adaptive His research interests include neural net-
control,” IEEE Trans. Neural Netw., vol. 3, no. 6, pp. 837–863, Nov. 1992. work control, vibration control, machine learn-
ing, computer systems, and software.

Hejia Gao (S’16) received the B.Eng. degree

in intelligence science and technology from the
School of Automation and Electrical Engineer-
ing, University of Science and Technology Bei-
jing, Beijing, China, in 2016. She is currently
working toward the Ph.D. degree with the School
of Automation and Electrical Engineering, Uni-
versity of Science and Technology Beijing.
Her research interests include neural network
control, vibration control, flexible robots, rein-
forcement learning, etc.

Wei He (S’09–M’12–SM’16) received the B.Eng.

and M.Eng. degrees in control theory and ap-
plication from the College of Automation Sci-
ence and Engineering, South China University
of Technology, China, in 2006 and 2008, respec-
tively, and the Ph.D. degree from the Depart-
ment of Electrical and Computer Engineering,
National University of Singapore, Singapore, in
2011.
He is currently a full Professor with the School
of Automation and Electrical Engineering, Uni-
versity of Science and Technology Beijing, Beijing, China. He has co-
authored 2 books published in Springer and published more than 100
international journal and conference papers. His current research in-
terests include robotics, distributed parameter systems, and intelligent Changyin Sun received the Bachelor’s de-
control systems. gree in applied mathematics from the College
Dr. He has been awarded a Newton Advanced Fellowship from the of Mathematics, Sichuan University, Chengdu,
Royal Society, U.K. He is a recipient of the IEEE SMC Society Andrew China, in 1996, and the M.S. and Ph.D. de-
P. Sage Best Transactions Paper Award in 2017. He serves as an Asso- grees in electrical engineering from Southeast
ciate Editor of IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARN- University, Nanjing, China, in 2001 and 2004,
ING SYSTEMS, IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, respectively.
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, He is currently a Professor with the School of
and IEEE ACCESS, and an Editor of IEEE/CAA JOURNAL OF AUTOMATICA Automation, Southeast University. His research
SINICA, Neurocomputing, and Journal of Intelligent and Robotic Systems. interests include intelligent control, flight control,
He is the member of the IFAC TC on Distributed Parameter Systems, pattern recognition, optimal theory, etc.
IFAC TC on Computational Intelligence in Control, and IEEE CSS TC on Prof. Sun is the Associate Editor of IEEE TRANSACTIONS ON NEURAL
Distributed Parameter Systems. NETWORKS AND LEARNING SYSTEMS.

Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on January 31,2024 at 03:38:58 UTC from IEEE Xplore. Restrictions apply.