Professional Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
Abstract—This paper presents a new adaptive parameter esti- there are usually unknown dynamics such as friction, modeling
mation method and the corresponding control design for nonlin- uncertainties and external disturbances in such systems, which
ear servo mechanisms with friction compensation. A continuous can reduce the control precision. Among these dynamics,
friction model is employed to capture the friction dynamics
of servo mechanisms. Then, the unknown system parameters frictions are highly nonlinear, nonsmooth and difficult to be
including friction model parameters are online estimated via the modelled, while leading to significant tracking performance
derived adaptive law. Hence, a new adaptive law is proposed degradation and potential instability of servo mechanisms.
to achieve faster and more accurate parameter estimation over To improve the control performance of such systems, many
classical adaptive laws so as to suppress the undesired transient control algorithms such as PID [4], adaptive control [5], and
dynamics. For this purpose, an auxiliary filter is introduced to
extract the estimation error for driving the parameter updating sliding mode control (SMC) [6]–[8] have been proposed.
law. Moreover, an adaptive control is designed in conjunction with In addition, neural networks (NN) [9]–[11] and fuzzy logic
a robust integral of the sign of the error (RISE) feedback term systems (FLS) [12] have also been utilized to compensate the
to address the bounded disturbances and enhance the tracking unknown frictions. In these methods, the friction dynamics
precision. Simulations and experiments are given to validate the are considered as a part of unknown system nonlinearities,
efficiency of the developed control scheme.
and then approximated via NNs or FLSs. Hence, no offline
Index Terms—Adaptive control, servo mechanisms, parameter friction modeling effort is required. However, the inherent
estimation, friction dynamics, robust integral of the sign of error. nonsmooth feature of friction dynamics may make the function
approximation invalid.
Over the past decades, some control methods based on static
N OMENCLATURE friction models have been proposed (e.g., [13] and references
J Motor inertia. therein), and various friction models have also been developed
id d− axis stator current. to capture diverse friction dynamics [14]–[16]. Among these
iq q− axis stator current. friction models, LuGre model can represent stribeck effect,
ud d− axis stator voltage. hysteresis, stick-lop limit cycling, and rising static friction,
uq q−axis stator voltage. and thus it has been used in the control designs for many
np Number of pole pairs. servo systems. For example, Xie [17] proposed a sliding
ψf Rotor flux linkage. mode-observer which was incorporated into adaptive control
q Angular position. to compensate for the friction of servo actuator. In [18], an
q̇ Angular speed. adaptive feedforward NN control scheme was developed to
TL Load torque. compensate the friction effect of the pendubot. However, the
KT Torque time constant. above adaptive controllers impose a stringent assumption that
Td Load disturbance. the friction model parameters should be known, which in turn
Tf Friction torque. requires an offline modeling work to obtain these parameters.
Moreover, in the adaptive control designs, the traditional
adaptive learning approaches (e.g., gradient [19] and least-
I. I NTRODUCTION squares (LS) [20] algorithms) have been used to derive the
High performance servo mechanisms have been widely used adaptive laws to online update the unknown model/control
in practical engineering applications, and thus have drawn parameters. These methods may trigger bursting phenomena
significant attentions in the academic fields [1]–[3]. However, when the system is subject to disturbances, i.e., the estimated
parameters may go to infinity, leading to the instability of
Manuscript received Mar 9, 2019; revised Jul 23, 2019; accepted Jan control system. To address this issue, some modified parameter
4, 2020. Date of publication XXX, 2020; date of current version XXX,
2020.This work is support by the National Natural Science Foundation of estimation methods have been developed in [21], where the
China under Grant No. 61803216, Grant No. 61873115, Grant No. 61922037, boundedness of the estimated parameters is retained. However,
the China Post-Doctoral Science Foundation under Grant No. 2019M650159, the estimation error can not converge to zero due to the
and the Natural Science Foundation of Shandong Province under Grant No.
ZR2018BF022. (Corresponding author: Jing Na) included damper terms in these robust adaptive laws [22].
Shubo Wang is with the School of Automation, Qingdao University, Following this observation, a composite estimation method
Qingdao, 266071, P. R. China. (e-mail:wangshubo1130@126.com) was incorporated into adaptive control to identify the unknown
Jing Na is with the Faculty of Mechanical and Electrical Engineering,
Kunming University of Science and Technology, Kunming Yunnan 650500, parameters to achieve guaranteed tracking error convergence
China.(e-mail: najing25@163.com) [23]. Nevertheless, the parameter estimation stay around the
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
pre-set values only. To overcome this problem, a new pa- PC for Display and Data Motor Drive
Power
Supply
rameter estimation method [24] was developed for nonlinear Acquisition
qd iq* uq uD
systems, where filter operation was proposed to extract the u Position
u PI
Inverse
0
Control Park PW
estimation error to design new adaptive laws to obtain expo-
*
i ud uE IPM
u Transfor M
d
q iq
PI
mation
nential or finite-time error convergence.
It is noted that in the above schemes, the unavoidable iD iaU V W
id Park Clark
ib
measurement noise or modeling uncertainties will make the Transf iE Transf
ormati ormati ic
on on
control error converge to a compact set only. In this respect, a
new continuously control scheme with a robust integral of the
Position Calculation of
sign of the error (RISE) term was proposed to accommodate angular and velocity
Encoder PMSM
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
Finite-Time Parameter Estimation To facilitate the parameter estimation, the second equation
in (6) can be written as
Adaptive law Filter operation
H
°kx2 f x2 f x2 P lP M f M Tf ẋ2 = ΘΨ(x2 , u) + ∆ (7)
°
4ˆ * <d r M ®
°̄kM f M f M ®
4̂ Q lQ M f ª¬ x2 x2 f / k º¼
¯°
where Θ = [θ1 , θ2 , θ3 , θ4 , θ5 , θ6 ] is the unknown param-
eter vector, and Ψ(x2 , u) = [u, −x2 , −(tanh(β1 x2 ) −
Reference
Control x1, x2 tanh(β2 x2 )), − tanh(β3 x2 ), −x2 , −1]T is the known regres-
Servo Command Servo
Controller Mechanisms Output sor vector.
xd The filtered variables x2f and Ψf of the dynamics x2 and
Closed-loop Servo System
Ψ are defined as [24]
{
k ẋ2f + x2f = x2 , x2f (0) = 0
Fig. 2. Structure of the proposed adaptive controller. (8)
k Ψ̇f + Ψf = Ψ, Ψf (0) = 0
where k > 0 is a filter constant. The following auxiliary
where γ1 , γ2 , γ3 , β1 , β2 and β3 are friction model parameters. variable ∆f (only used to analysis) is also defined:
The above friction model (5) has piecewise continuous feature,
˙ f + ∆f = ∆, ∆f (0) = 0
k∆ (9)
which can contribute to the smoothness of the derived control
signals when it is involved in the control designs. Moreover, Since ∆ is bounded, we know ∆f is also bounded.
the friction model coefficients γ1 , γ2 , γ3 can be online updated By applying low-pass filter operation given in (8) on both
via the proposed adaptive law to cover more realistic friction sides of (7) and considering the first equation of (8), we have
behaviors. The modeling errors of the friction model (5) will
x2 − x2f
be further compensated in the following control via a robust ẋ2f = = ΘΨf + ∆f (10)
term. k
Then, model (4) can be expressed as Based on (10), we define matrix P1 and vector Q1 as
{
Ṗ1 = −lP1 + Ψf ΨTf , P1 (0) = 0
ẋ1 =x2 [ ] (11)
ẋ2 =θ1 u − θ2 x2 − θ3 (tanh(β1 x2 ) − tanh(β2 x2 )) (6) Q̇1 = −lQ1 + Ψf (x2 − x2f )/k , Q1 (0) = 0
− θ4 tanh(β3 x2 ) − θ5 x2 − θ6 + ∆ where l > 0 is a constant.
The solution of (11) is derived as
where θ1 = K1 /J, θ2 = K2 /J, θ3 = γ1 /J, θ4 = γ2 /J, ∫ t
θ5 = γ3 /J, and θ6 = Tl are the unknown parameters to be
e−l(t−r) ΨTf (r)Ψf (r)dr
P1 =
estimated in this paper, and ∆ = −Td denotes the lumped 0
∫ t (12)
disturbances and modeling uncertainties.
[ ]
Q1 = e−l(t−r) ΨTf (r) (x2 − x2f )/k dr
Assumption 1: The disturbance ∆ and its first two-order
0
derivatives ∆ ˙ and ∆ ¨ are bounded.
Substituting (10) into (12), one has
The aim of this paper is to estimate the unknown parameters
θi , i = 1, ..., 6 and design an adaptive controller to achieve Q 1 = P 1 ΘT − ∆
¯ (13)
position tracking for servo system (6), i.e., the output x1 tracks ∫t
where ∆¯ = − e−ℓ(t−r) ΨT (r)∆f (r)dr is bounded since the
a desired trajectory xd . 0 f
lumped disturbance ∆ is bounded and the regressors Ψ and Ψf
are locally bounded. Thus, we have ∥∆∥¯ ≤ ε∆ for a constant
III. A DAPTIVE C ONTROL D ESIGN W ITH G UARANTEED ε∆ > 0.
PARAMETER E STIMATION Another auxiliary vector H1 is further defined as
In this part, a novel parameter estimation method is pro- H1 = P1 Θ̂T − Q1 (14)
posed for system (6). Fig.2 shows the structure of the devel-
oped control system, which consists of a feedback control with where Θ̂ is the estimate of the unknown parameter Θ.
a robust term to obtain tracking response, and a new adaptive Substituting (13) into (14), it is verified that H1 equals to
law used for parameter estimation.
H1 = P1 Θ̂T − P1 ΘT = −P1 Θ̃T + ∆
¯ (15)
where Θ̃ = Θ − Θ̂ denotes the estimation error.
A. Parameters Estimation Error Extraction It is shown in (15) that the variable H1 calculated based on
In the traditional adaptive control, the adaptive laws are P1 , Q1 given in (12) contains the estimation error Θ̃. Thus, as
derived by the gradient algorithm to minimize the tracking shown in [22], this variable can be used in the updating law to
error. It is clear that if the estimated parameters converge to retain the estimation error convergence and also enhance the
the true values fast (e.g., finite time (FT)), the convergence control performance in the adaptive control system.
of tracking error can be enhanced. Differing to these existing Moreover, as widely recognized in adaptive control, the
results, we will introduce a novel parameter updating law with proper excitation condition should be imposed to guarantee
the estimation error to obtain better estimation response. the convergence of the parameter estimation. In this line, we
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
assume that the derived matrix P1 is positive definite, such where Ψd = [u, −xd , −(tanh(β1 ẋd ) −
that its minimum eigenvalue λmin (P1 ) > ϱ > 0. As proved in tanh(β2 ẋd )), − tanh(β3 ẋd ), −ẋd , −1]T is the regressor
[22], this condition is related to the standard persistent exaction with the desired trajectory xd , and Γ > 0, σ > 0 are learning
(PE) condition widely used in adaptive control, which can be gains.
fulfilled in our case studies. It is noted that the variable r is not available for imple-
menting the online learning (22) directly. Hence, inspired by
B. Adaptive Controller Design with RISE the Newton-Leibniz formula and the idea given in [31], we
write equation (22) as the following form, which can be online
To address the position tracking control by introducing a calculated:
smooth RISE term to address the effect of modeling uncer- ∫ t( )
tainties and disturbances, we let z1 = x1 − xd be the position P T (τ )H1 (τ )
Θ̂(t) =Θ̂(0) + σ H1 (τ ) + 1 dτ
tracking error, and define a first-order sliding surface z2 and 0 ∥H1 (τ )∥
∫ t
the associated error r as [ (24)
+ ΓΨ̇d (τ )z2 (τ )|t0 − Γ Ψ̈d (τ )z2 (τ )
z2 = ż1 + k1 z1 , r = ż2 + k2 z2 (16) ] 0
− k2 Ψ̇d (τ )z2 (τ ) dτ
where k1 and k2 are positive constants. Note that the filter
It is found in (24) that the immeasurable variable r can be
error r is used to facilitate the stability analysis only and it is
replaced by the term ż2 + k2 z2 , where ż2 can be online
not used in the control design since equation (16) depends on
calculated by using the Newton-Leibniz formulation. Hence,
the immeasurable acceleration variable ẍ.
the estimated parameter Θ̂ can be obtained based on (24).
From (6) and (16), one has
Specifically, since the desired trajectory xd is used in the
r =θ1 u − θ2 x2 − θ3 (tanh(β1 x2 ) − tanh(β2 x2 )) adaptive law (24), the derivatives of the regressor Ψd can be
− θ4 tanh(β3 x2 ) − θ5 x2 − θ6 + ∆ − ẍd + k1 ż1 + k2 z2 calculated online, which is free of measurement noise and thus
(17) smooth, helping to obtain smoother parameter estimation.
To reduce the effect of noise in the measurements x, the Remark 1: To avoid potential singularity in the proposed
desired trajectory xd is adopted to replace the system state x control (20) (for the case with θ̂1 = 0), the initial condition
in the control design. Thus, the error variable r given in (17) θ̂1 (0) should be properly selected and/or the project operation
can be rewritten by using the desired trajectories xd as: [21] can be applied, while the subsequent stability analysis
does not change. Specifically, since the unknown parameter
r = θ1 u − θ2 ẋd − θ3 (tanh(β1 ẋd ) − tanh(β2 ẋd ))
(18) θ1 in the studied servo system is a positive constant, we can
− θ4 tanh(β3 ẋd ) − θ5 ẋd − θ6 − ẍd + S + ∆ choose the initial condition as θ̂1 (0) > 0.
where the auxiliary function S(x1 , x2 , xd , ẋd ) is defined as Remark 2: Differing from the existing adaptive control
schemes, a novel leakage term M , containing the parameter
S = − θ2 (x2 − ẋd ) − θ3 [(tanh(β1 x2 ) − tanh(β2 x2 )) estimation error information H1 , is imposed on the traditional
− (tanh(β1 ẋd ) − tanh(β2 ẋd ))] − θ4 (tanh(β3 x2 ) (19) ˙
adaptive law Θ̂ = ΓΨ̇d r in (22), which can guarantee the
− tanh(β3 ẋd )) − θ5 (x2 − ẋd ) + k1 ż1 + k2 z2 convergence of parameter estimation and the control error. In
particular, a sliding mode term given in (23) is used, such
Then, the controller can be designed as that the finite-time convergence of H1 can be guaranteed
(
1 as shown in the next subsection. In this respect, this new
u= (θ̂2 + θ̂5 )ẋd + θ̂3 (tanh(β1 ẋd ) − tanh(β2 ẋd )) leakage term improves the results over other adaptive methods
θ̂1 ) (20) (e.g., e-modification and σ-modification [32]), which cannot
+ θ̂4 tanh(β3 ẋd ) + θ̂5 ẋd + ẍd + θ̂6 − µs retain the convergence of Θ̂ to Θ. This convergence property
allows online modeling of uncertain systems (i.e., accurate
where θ̂i is the estimate of unknown parameters θi , which will estimation of unknown parameters including the friction model
be online updated with the lumped vector Θ̂, and µs denotes parameters), apart from achieving motion tracking.
the RISE feedback term, which is defined as
µs = (ks + 1)z2 − (ks + 1)z2 (0) C. Practical Implementation
∫ t
[ ] The implementation of the proposed control algorithm is
+ (ks + 1)k2 z2 (σ) + λsgn(z2 (σ)) dσ given as follows:
0
(21) 1) Select the adaptive law parameters κ, Γ, l, σ and initial
where ks and λ are positive constants, and sgn(·) is the signum condition Θ̂(0).
function. 2) Calculate the tracking error z1 = x1 −xd , and the filtered
The parameter updating law for Θ̂ can be designed as errors z1 .z2 based on (16).
˙ 3) Choose the controller parameters k1 , k2 , ks , and calcu-
Θ̂ = Γ(Ψ̇d r − M ) (22) late the control action u and parameter updating law Θ̂
based on (20) and (24).
with
P1T H1 4) Apply the control action on the servo mechanisms and
M = σH1 + σ (23) record the input/output measurements.
∥H1 ∥
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
0.2
Reference Adaptive Gradient Reference Adaptive Gradient
0.1
Position tracking
Position tracking
0.1
0.05
0 0
-0.05
−0.1
-0.1
−0.2 -0.15
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
1.5 1.5
Reference Adaptive Gradient Reference Adaptive Gradient
1 1
Speed tracking
Speed tracking
0.5 0.5
0 0
−0.5 -0.5
−1 -1
−1.5 -1.5
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time(s) Time (s)
Fig. 3. Simulation output tracking with constant parameters. Fig. 5. Simulation output tracking with varying parameters
10
True Value Adaptive Gradient
θ̂1
10 5
θ̂1
5 0
True Value Adaptive Gradient
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0 5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
θ̂2
5 0
-5
θ̂2
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
1
-5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
θ̂3
0
1
-1
θ̂3
θ̂4
0
20 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-1
θ̂4
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
!
"
be achieved with the proposed control (20). In particular, the
, tracking errors of adaptive controllers decrease gradually as
#
long as the parameter estimates converge to constants, while
$%& PID control has a larger tracking error than others. We can
,
also see that the control error of the ANDSC is smaller than
'
()*
PID control and the gradient based adaptive control. This is
+
-
reasonable since NNs are adopted to handle the nonlinearities
in the ANDSC scheme.
Fig. 7. Structure of the two-axis servo system. Case 2: To further justify the control performance, a fast
varying sinusoid signal xd = 0.8 sin(2πt/5.5) is adopted as
the reference signal. We can find from Fig.10 that satisfactory
1) The proposed adaptive control: It is noted that there control response can also be achieved in this case. It is noted
is two terms of x2 associated with the parameters θ1 that the tracking error of the proposed adaptive control is
and θ5 in the regressor Ψ. Then, we can define the smaller than the other control methods (gradient, ANDSC, and
regressor vector Ψ as Ψ = [u, −x2 , −(tanh(β1 x2 ) − PID), since fast convergence of the estimated parameters can
tanh(β2 x2 )), − tanh(β3 x2 ), −1]T to reduce the computation- be retained as shown in Fig. 8. Nevertheless, we can also find
al costs. The parameters used in the proposed control are set from Fig.10 (a) that the control error of the ANDSC is smaller
as: k1 = 10, k2 = 1, ks = 12, k = 0.1, l = 10, σ = 2, than PID control and gradient based adaptive control. More-
β = 0.01 and Γ = diag([2 1 1 1 2]), The initial condition is over, Fig.10(b) depicts the control signals of three controllers.
Θ̂(0) = [6, 0, 0, 0, 0]. One can find that the control signal of the proposed adaptive
2) Gradient based adaptive control: This control is the same is smooth, while the control signal of the ANDSC method has
as the developed control, while the updating law (22) is set obvious oscillations due to the use of high-order derivatives of
with σ = 0, such that only the first term derived with the the virtual controls. From these comparative results, one can
gradient algorithm is used. All the control parameters and claim that the developed control scheme (20) with finite-time
conditions are the same as 1). Hence, the essential difference learning algorithm (22) can effectively handle the unknown
to the proposed control in 1) is that the leakage term of H1 disturbances and frictions.
used in the adaptive law (22) is switched off. To further quantitatively compare the performance of d-
3) Adaptive neural dynamic surface control (ANDSC) [33]: ifferent controllers, the following performance indices are
We define z1 = x1 − xd and z2 = x2 − s1 with µ1 ṡ1 + s1 = calculated: 1) Maximum absolute value of the tracking error
α1 and α1 = −k1 z1 − θ̂z1 ΦT1 Φ1 /2 − ε̂1 tanh(z1 /ω1 ). Then Me = max {|z1 (i)|}; 2) Average tracking error Ae =
the controller is u = −k2 z2 − θ̂z2 ΦT2 Φ2 /2 − ε̂2 tanh(z2 /ω2 ) 1
∑N i=1,...,N
and the adaptive laws are θ̂i = Γi [zi2 ΦTi Φi − σi θ̂i ]/2. The N √ z1∑
i=1 (i); 3) Standard deviation of the tracking error
N
parameters are set as k1 = 9, k2 = 4, Γ1 = Γ2 = 100, Se = N i=1 [z1 (i) − Ae ]2 . These performance indices of
1
µ1 = 0.01, Γa1 = Γa2 = 10, σ1 = σa1 = σa2 = 0.01, and two different controllers for the above two cases are shown in
ω1 = ω2 = 1. Table I.
4) PID : Finally, a well tuned PID control is also tested.
TABLE I
The PID control with gains kp = 30, ki = 0.05, kd = 10 C OMPARISON RESULTS OF PERFORMANCE INDICES
are tuned based on a reference signal xd = 0.4 sin(2πt/4) by
using a trial-and-error method to make a tradeoff between the .
steady-state performance and transient response. In general, xd = 0.4 sin(2πt/4) xd = 0.8 sin(2πt/5.5)
a large proportional gain kp may increase convergent speed PID ANDSC Gradient Adaptive PID ANDSC Gradient Adaptive
while trigging oscillations in the control system. Too large
Me 0.1156 0.0764 0.1595 0.0601 0.1312 0.1021 0.2016 0.0990
integral gain ki may lead to large overshoots in the transient
stage. The derivative gain kd can enhance convergence rate but Ae 0.0970 0.0192 0.0313 0.0148 0.0980 0.0203 0.0337 0.0136
it is sensitive to the measurement noise. The derived control Se 6.9× 1.4 × 3.01× 1.28× 8.3× 1.48× 1.9 × 1.25×
gains are then used for all experiments to compare their ability 10−4 10−5 10−4 10−5 10−4 10−4 10−4 10−4
and generality under different references.
Case 1: It is known that the effects of friction, ripple From Table I, one can see that the proposed adaptive control
disturbance are more notable at the low speed regime. Hence, obtains smaller Me , Ae and Se in all cases, which denote
a sinusoidal wave xd = 0.4 sin(2πt/4) is firstly used as the better control performance. In particular, all the performance
desired trajectory to test the control performance. Figs.8-9 indices of the proposed adaptive control are better than the
depict the experiment results. As it can be seen, the estimated gradient based adaptive control. This is due to the fact that the
parameters with adaptive law (22) converge to constants near introduced leakage term M in the adaptive law (22) contains
the nominal values, while the gradient based adaptive law can the estimation error information H1 , and thus it can drive
not obtain convergent estimates, i.e., the estimated parameters the estimated parameters converging to their actual values
have significant oscillations. Moreover, Fig.9 provides the fast. Moreover, it is also clear that the control performance
control errors and control actions u of these four controllers. of the ANDSC is close to the proposed adaptive control,
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
6 0.1
Adaptive
5 Adaptive
θ̂1
Gradient
4 0
0 2 4 6 8 10 12 14 16 18 20
-0.1
0 Gradient
0 2 4 6 8 10 12 14 16 18 20 0
1.5
-0.1
1
0 2 4 6 8 10 12 14 16 18 20
θ̂3
0.5
0 0.1
0 2 4 6 8 10 12 14 16 18 20
ANDSC
0
0.5
θ̂4
-0.1
0 0 2 4 6 8 10 12 14 16 18 20
0 2 4 6 8 10 12 14 16 18 20 0.1
0.05 PID
0
θ̂5
0
-0.05 -0.1
0 2 4 6 8 10 12 14 16 18 20
0 2 4 6 8 10 12 14 16 18 20
Time(s)
Time(s)
2
Adaptive
while its computational costs are more demanding due to 0
new adaptive law (22) over the generic adaptive control with -2
0 2 4 6 8 10 12 14 16 18 20
the gradient algorithm. Specifically, it cannot only retain the 2
(b)
VII. C ONCLUSION
This paper proposed and validated a novel adaptive control Fig. 9. Experimental results for reference signal xd = 0.4 sin(2πt/4): (a)
Tracking error, and (b) Control signal u
scheme for servo mechanisms with friction dynamics. An
auxiliary filter is introduced to extract the estimation error for
driving the parameter updating law, where a sliding mode term Select a Lyapunov function candidate as
was incorporated into the adaptive law to achieve FT parameter 1 1 1 1
estimation convergence. Moreover, the desired trajectory is VL (y) = z12 + z22 + r2 + P2 + Θ̃Γ−1 Θ̃T (38)
2 2 2 2
used to replace the actual system state to reduce the effect which satisfies the inequalities
of the noise injected into the online learning. The external
disturbance and modeling uncertainties can be compensated U1 (y) ≤ VL (y) ≤ U2 (y) (39)
by using a RISE term in the control design. The effectiveness The continuous, positive definite functions U1 (y) and U2 (y)
of the proposed method is validated by simulations and exper- are defined as
iments. The salient feature of the proposed adaptive control
over existing adaptive control methods lies in that it cannot U1 (y) = λ1 ∥y∥, U2 (y) = λ2 ∥y∥ (40)
only achieve satisfactory tracking performance, but also allow where λ1 and λ2 are positive constants.
the online modeling of uncertain systems since it can estimate The time derivative of VL can be written as
the unknown system parameters. It is also noted that the ˙T
rotation velocity of the servo mechanism is required in the V̇ =z ż + z ż + rṙ + Ṗ + Θ̃Γ−1 Θ̃
L 1 1 2 2 2
proposed control implementation. Thus, future work will focus = z1 (z2 − k1 z1 ) + z2 (r − k2 z2 ) + r(Θ̃Ψ̇d + Ñ + Nd
on the output feedback control design for servo mechanisms − z2 − (ks + 1)r − λsgn(z2 )) − rNd + rλsgn(z2 ) − C1
with other nonsmooth dynamics.
Θ̃P1T H1
− Θ̃Ψ̇d r + σ Θ̃H1 + σ
∥H1 ∥
A PPENDIX A
P ROOF OF T HEOREM 1 = − k1 z1 − k2 z2 − (ks + 1)r2 + z1 z2 + rÑ − C1
2 2
¯ − σ (H1 − ∆) H1
¯ T
Proof: We define a vector y as − σ Θ̃(P1 Θ̃ − ∆)
√ ∥H1 ∥
y = [z, P2 ] (37) (41)
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
0.1
Gradient
. This together with the fact that xd , ẋd and ẍd are bounded
0
implies that x1 , ẋ1 ∈ L∞ in Ξ.
-0.1
0 2 4 6 8 10 12 14 16 18 20 Given that r ∈ L∞ in Ξ, similar analysis as given in [30]
0.1
ANDSC
can be used to prove that µ̇s ∈ L∞ in Ξ. Then, we can also
0
verify that ṙ ∈ L∞ . Since ż1 , ż2 , and ṙ ∈ L∞ in Ξ, the
-0.1
0 2 4 6 8 10 12 14 16 18 20
definitions for U (y) and z(t) can be used to prove that U (y) is
0.1 uniformly continuous in Ξ. Hence, following similar analysis
as shown in [25], [30], we can verify that U (y) → 0 and thus
PID
0
-0.1
0 2 4 6 8 10 12 14 16 18 20
z → 0 for t → 0 and y(0) ∈ Ξ. Then, from the definition of
Time(s) r, z1 , z2 , we can obtain that z1 → 0 for t → ∞.
In the final step, we will prove the finite-time convergence
(a)
of the parameter updating law Θ̂ in (22) and (23).
2 Select the Lyapunov function VΘ̃ as
Adaptive
0
1 T −1 −1
VΘ̃ = H P P H1 (46)
-2
0 2 4 6 8 10 12 14 16 18 20 2 1 1 1
2
The time derivative of VΘ̃ is
Control Signal u (vol)
Gradient
0
∂H1 P1 ( ′)
V̇Θ̃ =H1T P1−1 = H1T P1−1 Γ (Ψ̇d r − M ) + ∆
-2
0 2 4 6 8 10 12 14 16 18 20
2 ∂t
′
=H1T P1−1 ΓΨ̇d r − H1T P1−1 ΓσH1 − Γσ∥H1 ∥ + H1T P1−1 Γ∆
ANDSC
0
[ ′ ]
-2
0 2 4 6 8 10 12 14 16 18 20 ≤ − ζ(Γ)ζ(σ) − ζ̄(Γ)∥P1−1 ∥∥Ψ̇d ∥∥r∥ − ζ̄(Γ)∥P1−1 ∆ ∥ ∥H1 ∥
√
2 [
2
0
PID
≤− −1
ζ(Γ)ζ(σ) − ζ̄(Γ)ζ̄(P1−1 )∥Ψ̇d ∥∥r∥
ζ̄(P1 )
-2 ′ ]√
0 2 4 6 8 10
Time(s)
12 14 16 18 20
− ζ(Γ)ζ(P1−1 )∥∆ ∥ VΘ̃
(47)
′
(b) where ∆ = −P1−1 Ṗ1 P1−1 ∆ ¯ + P −1 ∆
1
¯˙ is a bounded variable,
which can be verified by considering the fact that the regressor,
Fig. 10. Experimental results for reference signal xd = 0.8 sin(2πt/5.5): ¯ are all bounded.
(a) Tracking error, and (b) Control signal u the matrix P1 , and the residual error ∆
Since r will converge to a small compact set, and the
regressor Ψ̇d is bounded, then there exists a time T and large
Using the Young’s inequality, one has learning gain Γ, such that for all t > T it can be claimed that
1 2 1 2 ′
z1 z2 ≤ z + z , ζ(Γ)ζ(σ) > ζ̄(Γ)ζ̄(P1−1 )(∥Ψ̇d ∥∥r∥ + ∥∆ ∥) (48)
2 1 2 2
2 2 (42)
ρ (∥z∥)∥z∥
rÑ ≤ ρ(∥z∥)∥z∥|r| ≤ ks |r|2 + . √ T1 and a constant ϵ > 0, such
Thus, there is a time instant
4ks that the condition V̇Θ̃ ≤ −ϵ VΘ̃ is true for all t > T1 . There-
Substituting (42) into (41), yields fore, according to the convergence property of sliding mode
( ) ( ) control in [34], the finite-time convergence of the estimation
1 2 1 2
V̇L ≤ − k1 − z 1 − k2 − z − r2 − C1 error P1−1 H1 to zero can be claimed. Then, by recalling the
2 2 2 fact given in (15), one can verify that Θ̃ = P1−1 H1 = P1−1 ∆ ¯
( ) (43)
ρ2 (∥z∥)∥z∥2 σ in finite-time. This completes the proof.
+ − σϱ − ∥Θ̃∥2 + C1
4ks 2ζ
σζε2 R EFERENCES
where C1 = 2 ∆ + σε∆ is a positive constant for design
parameter ζ > 2ϱ. [1] S. Li and Z. Liu, “Adaptive speed control for permanent-magnet syn-
Then, the derivative of VL can be obtained as chronous motor system with variations of load inertia,” IEEE Transac-
tions on Industrial Electronics, vol. 56, no. 8, pp. 3050–3059, 2009.
ρ2 (∥z∥)∥z∥2 [2] W. Sun, Y. Liu, and H. Gao, “Constrained sampled-data arc for a class of
V̇L ≤ −λ3 ∥z∥2 + ≤ −U (y) (44) cascaded nonlinear systems with applications to motor-servo systems,”
4ks
IEEE Transactions on Industrial Informatics, vol. 15, no. 2, pp. 766–
where λ3 = min{k2 − 21 , k3 − 12 , 1, σ(ϱ − 1/2ζ)} is a positive 776, 2019.
[3] J. Na, Q. Chen, X. Ren, and Y. Guo, “Adaptive prescribed performance
constant, and U (y) = c∥z∥2 with a positive constant c is a motion control of servo mechanisms with friction compensation,” IEEE
continuous positive function over a set Ξ defined as Transactions on Industrial Electronics, vol. 61, no. 1, pp. 486–494, 2014.
√ [4] P. Dupont, “Avoiding stick-slip through pd control,” IEEE Transactions
Ξ ≡ {y(t)|∥y∥ ≤ ρ−1 (2 λ3 ks )} (45) on Automatic Control, vol. 39, no. 5, pp. 1094–1097, 1994.
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2971056, IEEE
Transactions on Industrial Informatics
10
[5] T. Gao, Y. Liu, L. Liu, and D. Li, “Adaptive neural network-based control [29] Y. Zhang, B. Xian, and S. Ma, “Continuous robust tracking control for
for a class of nonlinear pure-feedback systems with time-varying full magnetic levitation system with unidirectional input constraint,” IEEE
state constraints,” IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 5, Transactions on Industrial Electronics, vol. 62, no. 9, pp. 5971–5980,
pp. 923–933, 2018. 2015.
[6] X. Chen and T. Hisayama, “Adaptive sliding-mode position control [30] C. Makkar, G. Hu, W. Sawyer, and W. Dixon, “Lyapunov-based tracking
for piezo-actuated stage,” IEEE Transactions on Industrial Electronics, control in the presence of uncertain nonlinear parameterizable friction,”
vol. 55, no. 11, pp. 3927–3934, 2008. IEEE Transactions on Automatic Control, vol. 52, no. 10, pp. 1988–
[7] C. Milosavljevic, B. Perunicic-Drazenovic, and B. Veselic, “Discrete- 1994, 2007.
time velocity servo system design using sliding mode control approach [31] N. Sharma, S. Bhasin, Q. Wang, and W. E. Dixon, “Rise-based adaptive
with disturbance compensation,” IEEE Transactions on Industrial Infor- control of a control affine uncertain nonlinear system with unknown
matics, vol. 9, no. 2, pp. 920–927, 2013. state delays,” IEEE Transactions on Automatic Control, vol. 57, no. 1,
[8] Q. Chen, S. Xie, M. Sun, and X. He, “Adaptive nonsingular fixed- pp. 255–259, 2012.
time attitude stabilization of uncertain spacecraft,” IEEE Transactions [32] J. J. E. Slotine and W. Li, Applied nonlinear control. NJ,USA:Prentice-
on Aerospace and Electronic Systems, vol. 54, no. 6, pp. 2937–2950, Hall, 1991.
2018. [33] J. Na, X. Ren, G. Herrmann, and Z. Qiao, “Adaptive neural dynamic
[9] B. Xu, D. Yang, Z. Shi, Y. Pan, B. Chen, and F. Sun, “Online recorded surface control for servo systems with unknown dead-zone,” Control
data-based composite neural control of strict-feedback systems with Engineering Practice, vol. 19, no. 11, pp. 1328–1343, 2011.
application to hypersonic flight dynamics,” IEEE Transactions on Neural [34] S. P. Bhat and D. S. Bernstein, “Continuous finite-time stabilization of
Networks and Learning Systems, vol. 29, no. 8, pp. 3839–3849, 2018. the translational and rotational double integrators,” IEEE Transactions
[10] H. Gao, W. He, C. Zhou, and C. Sun, “Neural network control of a two- on Automatic Control, vol. 43, no. 5, pp. 678–682, 1998.
link flexible robotic manipulator using assumed mode method,” IEEE
Transactions on Industrial Informatics, vol. 15, no. 2, pp. 755–765,
2019.
[11] S. Wang, H. Yu, J. Yu, J. Na, and X. Ren, “Neural-network-based
adaptive funnel control for servo mechanisms with unknown dead-zone,”
IEEE Transactions on Cybernetics, pp. 1–12, 2018.
[12] L. Cao, H. Li, N. Wang, and Q. Zhou, “Observer-based event-triggered
adaptive decentralized fuzzy control for nonlinear large-scale systems,”
IEEE Transactions on Fuzzy Systems, pp. 1–1, 2018. Shubo Wang (M’19) received the B.S. degree in
[13] J. Moreno, R. Kelly, and R. Campa, “Manipulator velocity control physics from Binzhou University, Shandong, China,
using friction compensation,” IEE Proceedings - Control Theory and in 2008; the M.S degree in control science and
Applications, vol. 150, no. 2, pp. 119–126, 2003. engineering from the School of Information Science
[14] J. Amin, B. Friedland, and A. Harnoy, “Implementation of a friction and Engineering, Central South University, Hunan,
estimation and compensation technique,” IEEE Control Systems, vol. 17, China, 2011; and the Ph.D. degree in control sci-
no. 4, pp. 71–76, 1997. ence and engineering from the Beijing Institute of
[15] P. Dahl, “A solid friction model,” DTIC Document, Tech. Rep., 1968. Technology, Beijing, China, in 2017. Since 2017, He
[16] C. De Wit, H. Olsson, K. Astrom, and P. Lischinsky, “A new model has been with the School of Automation, Qingdao
for control of systems with friction,” IEEE Transactions on Automatic University, where he became an Associate Professor
Control, vol. 40, no. 3, pp. 419–425, 1995. in 2019. He has coauthored one monograph and
[17] W.-F. Xie, “Sliding-mode-observer-based adaptive control for servo more than 30 international journal and conference papers.
actuator with friction,” IEEE Transactions on Industrial Electronics, His current research interests include adaptive control, adaptive parameter
vol. 54, no. 3, pp. 1517–1527, 2007. estimation, neural network, motor control, nonlinear control and applications.
[18] D. Xia, L. Wang, and T. Chai, “Neural-network-friction compensation-
based energy swing-up control of pendubot,” IEEE Transactions on
Industrial Electronics, vol. 61, no. 3, pp. 1411–1423, 2014.
[19] M. Bodson and S. Sastry, Adaptive control: stability, convergence and
robustness. Prencice Hall, 1989.
[20] K. S. Narendra and A. M. Annaswamy, Stable adaptive systems. Courier
Corporation, 2012.
[21] P. A. Ioannou and J. Sun, Robust adaptive control. Prentice-Hall, Inc.,
1995.
[22] J. Na, Y. Huang, X. Wu, G. Gao, G. Herrmann, and J. Z. Jiang, Jing Na (M’15) received the B.Eng. and Ph.D.
“Active adaptive estimation and control for vehicle suspensions with degrees from the School of Automation, Beijing
prescribed performance,” IEEE Transactions on Control Systems Tech- Institute of Technology, Beijing, China, in 2004
nology, vol. 26, no. 6, pp. 2063–2077, 2018. and 2010, respectively. From 2011 to 2013, he was
[23] G. Zhang, J. Chen, and Z. Li, “Identifier-based adaptive robust con- a Monaco/ITER Postdoctoral Fellow at the ITER
trol for servomechanisms with improved transient performance,” IEEE Organization, Saint-Paul-ls-Durance, France. From
Transactions on Industrial Electronics, vol. 57, no. 7, pp. 2536–2547, 2015 to 2017, he was a Marie Curie Intra-European
2010. Fellow with the Department of Mechanical Engi-
[24] J. Na, M. N. Mahyuddin, G. Herrmann, X. Ren, and P. Barber, “Robust neering, University of Bristol, U.K. Since 2010, he
adaptive finite-time parameter estimation and control for robotic sys- has been with the Faculty of Mechanical and Elec-
tems,” International Journal of Robust and Nonlinear Control, vol. 25, trical Engineering, Kunming University of Science
no. 16, pp. 3045–3071, 2015. and Technology, Kunming, China, where he became a Professor in 2013. He
[25] B. Xian, D. Dawson, M. de Queiroz, and J. Chen, “A continuous has coauthored one monograph and more than 100 international journal and
asymptotic tracking control strategy for uncertain nonlinear systems,” conference papers. His current research interests include intelligent control,
IEEE Transactions on Automatic Control, vol. 49, no. 7, pp. 1206–1211, adaptive parameter estimation, nonlinear control and applications for robotics,
2004. vehicle systems and wave energy convertor, etc.
[26] L. Lu, Z. Chen, B. Yao, and Q. Wang, “Desired compensation adaptive He is currently an Associate Editor of the IEEE Transactions on Indus-
robust control of a linear-motor-driven precision industrial gantry with trial Electronics, the Neurocomputing, and has served as the Organization
improved cogging force compensation,” IEEE/ASME Transactions on Committee Chair of DDCLS 2019, international program committee Chair of
Mechatronics, vol. 13, no. 6, pp. 617–624, 2008. ICMIC 2017. Dr Na has been awarded the Best Application Paper Award of
[27] T. Dierks and S. Jagannathan, “Neural network control of mobile robot the 3rd IFAC International Conference on Intelligent Control and Automation
formations using rise feedback,” IEEE Transactions on Systems, Man, Science (IFAC ICONS 2013), and the 2017 Hsue-shen Tsien Paper Award.
and Cybernetics, Part B: Cybernetics, vol. 39, no. 2, pp. 332–347, 2009.
[28] Y. Xu and K. Mohseni, “Bioinspired hydrodynamic force feedforward
for autonomous underwater vehicle control,” IEEE/ASME Transactions
on Mechatronics, vol. 19, no. 4, pp. 1127–1137, 2014.
1551-3203 (c) 2019 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: University of Exeter. Downloaded on June 13,2020 at 21:48:26 UTC from IEEE Xplore. Restrictions apply.