See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328449173

On disturbance rejection in magnetic levitation

Article  in  Control Engineering Practice · January 2019

DOI: 10.1016/j.conengprac.2018.09.018

CITATIONS READS

0 49

3 authors, including:

Wenchao Xue
Chinese Academy of Sciences
47 PUBLICATIONS   665 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Gasoline Engine Combustion Control View project

ADRC: a new paradigm of the science of automatic control View project

All content following this page was uploaded by Wenchao Xue on 23 October 2018.

The user has requested enhancement of the downloaded file.

 Control Engineering Practice 82 (2019) 24–35

Contents lists available at ScienceDirect

Control Engineering Practice

journal homepage: www.elsevier.com/locate/conengprac

On disturbance rejection in magnetic levitation

Wei Wei a , Wenchao Xue b,c ,∗, Donghai Li d
a
School of Computer and Information Engineering, Beijing Key Laboratory of Big Data Technology for Food Safety, Beijing Technology and Business
University, 100048, Beijing, China
b
LSC, NCMIS, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China
c
School of Mathematical Sciences, University of Chinese Academy of Sciences, Beijing 100049, China
d Department of Thermal Engineering, Tsinghua University, 10084, Beijing, China

ARTICLE INFO ABSTRACT

Keywords: Magnetic levitation systems belong to an important and challenging class of control engineering problems with
Magnetic levitation system nonlinear uncertain dynamics, multiple disturbances and large sensor noise. To obtain a simple and practical
Open-loop instability solution that does not depends on the exact model information, a time-varying active disturbance rejection
Time-varying extended state observer
solution is proposed and validated in both numerical and experimental results. The proposed method is confirmed
Sensor noise attenuation
with rigorous analysis of transient performance and noise attenuation. Moreover, the proposed solution is tested
in a magnetic levitation ball system with disturbances and measurement noise. The results show that the proposed
solution is effective and practical.

1. Introduction & Graichen, 2013) have also been designed to stabilize the magnetic
levitation systems. A nonlinear magnetic levitation system model is
Magnetic levitation is a technique to suspend an object via a non- transformed into a set of piecewise linear model, and then an explicit
contact pattern. Such a method enables frictionless levitation of objects, nonlinear predictive control has been established (Ulbig, Olaru, Dumur,
and offers numerous advantages, such as elimination of lubrication & Boucher, 2010). Additionally, based on nonlinear models, different
system, lowering rotating losses, providing higher speed and long- approaches have also been proposed (Bonivento, Gentili, & Marconi,
life service. Due to its great potential in engineering applications, 2005; Glueck, Kemmetmueller, Tump, & Kugi, 2011; Xu, Hwa Chen,
including magnetic bearings, high-speed magnetically levitating trains, & Guo, 2015; Yang, Fukushima, Kanae, & Wada, 2009). Most reported
micro-manipulation of levitating objects, nanoscale positioning systems, approaches emphasize that, if better performance is expected, an accu-
and vibration isolation systems (see Ref. Golob & Tovornik, 2003; rate mathematical model is of necessity. However, it is hard to obtain
Khamesee & Shameli, 2005; Kim, Verma, & Shakir, 2007; Kummer et al., an accurate model for a complex magnetic levitation system. PID, a
2010; Lanzara, D’Ovidio, & Crisi, 2014; Phuah, Lu, & Yahagi, 2005; model free control approach, commonly used in control engineering
Schuhmann, Hofmann, & Werner, 2012 and the references therein for practice, is also employed (Abdel-Hady & Abuelenin, 2008; Berkelman
details), magnetic levitation has attracted much attention and becomes & Dzadovsky, 2013; Golob & Tovornik, 2003). Especially, for improving
increasingly popular in industries. However, inherently nonlinear, open- tracking performance, fuzzy logic based PID control (Golob & Tovornik,
loop unstable, time-varying self-inductance and mutual coefficients,
2003) and fuzzy-supervised PID control (Abdel-Hady & Abuelenin,
nonlinear and time-varying electromagnetic force, coupled with the
2008) have been proposed. However, fuzzy rules largely depend on
nonlinear actuators make the control of magnetic levitation system be
experience, which limits applications of such approaches more or less.
a challenge.
Besides the approaches mentioned above, advanced control techniques,
With an attempt to achieve desired performance, numerous magnetic
such as sliding mode control (Elahi & Nekoubin, 2011), neuronal control
levitation control techniques have been proposed. By combining linear
(Chen, Lin, & Shyu, 2009; Pan and Liu et al., 2016; Rubio et al., 2017),
quadratic Gaussian control, fault tolerant control and multi-objective
and data-driven approach (Qin, Peng, Ruan, Wu, & Gao, 2014) have also
optimization, Michail et al. proposed a linear controller to get optimum
been proposed and implemented successfully.
performance (Michail, Zolotas, Goodall, & Whidborne, 2012). Nonlin-
In this paper, we still focus on the control of magnetic levitation
ear feedback linearization (Torres, Schnitman, Junior, & Felippe de
Souza, 2012; Trumper, Olson, & Subrahmanyan, 1997), feed-forward system. By estimating and canceling the generalized disturbance, which
linearization (Morales & Sira-Ramirez, 2010), lead compensator (Weng, includes both internal uncertainties and external disturbances, in real-
Lu, & Trumper, 2002), and model predictive control (Bächle, Hentzelt, time, active disturbance rejection control (ADRC) is capable of achieving

∗ Corresponding author at: LSC, NCMIS, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China.
E-mail address: [email protected] (W. Xue).

https://doi.org/10.1016/j.conengprac.2018.09.018
Received 22 January 2018; Received in revised form 21 July 2018; Accepted 16 September 2018
Available online xxxx
0967-0661/© 2018 Elsevier Ltd. All rights reserved.
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 1. A magnetic levitation ball system.

satisfactory performance (Han, 1995, 1998, 2009). However, relatively

complex structure and more parameters limit the applications of ADRC.
For simplifying structure and reducing the number of tunable param-
eters, linear active disturbance rejection control (LADRC) has been
proposed (Gao, 2003). It is proved that LADRC is able to obtain desired
closed-loop system performance with only two tunable parameters.
Besides, those tunable parameters can be determined by an easily
acceptable bandwidth-parameterization approach (Gao, 2003). So far,
LADRC has been successfully used in several industrial sectors and does
help the extensive applications of ADRC. Recently, LADRC has also been
utilized in the control of magnetic levitation system (Zhang & Zhang,
2017).
Actually, as we known, extended state observer (ESO) is the key
part of ADRC to promote control performance. In order to improve the
system performance further, numerous modified ESOs (Pu, Yuan, Yi, &
Tan, 2015; Xue et al., 2015; Zhao & Guo, 2015) have been proposed. Fig. 2. Structure of magnetic levitation ball control system.

Guo and Zhao propose a special kind of nonlinear ESO to ensure desired
estimation error (Zhao & Guo, 2015). Pu et al. also suggest a particular
time varying gain of nonlinear ESO to avoid the output of ESO being and TADRC are both designed. It is shown that TESO can guarantee both
large (Pu et al., 2015). But those results pay litter attention to the perfor- fast estimation of generalized disturbance and desired noise filtering.
mance analysis of ESO when system output is corrupted by sensor noise, The rest of this paper is organized as follows. In Section 2, a magnetic
which is inevitable in practice. Xue et al. (2015) start the point of using levitation ball system, including its physical model and mathematical
dynamic gain of ESO such that certain index of estimation error can be model, is described. TADRC is designed and analyzed in Section 3.
minimized in sense of stochastic. Nevertheless, the solution of optimized Numerical results have been provided in Section 4. Experimental results
problem usually needs some statistic information of measurement noise. and discussions are presented in Section 5. In the end, Section 6
Also, there are other approaches to deal with the sensor noise, such as concludes the paper.
adding a dimension of ESO (Martinez-Vazquez, Rodriguez-Angeles, &
Sira-Ramirez, 2009), providing a low-pass filter (Wei, Liang, Li, & Su, 2. Magnetic levitation ball system
2016), but the cost is making a higher order ESO or generating phase-
delays. 2.1. System description
In this paper, in order to improve the system performance one step
further, noise attenuation, a commonly discussed problem in practice, Magnetic levitation ball control system is a platform to investigate
is also taken into consideration in the design of ESO. The main objective magnetic levitation technique in laboratory. Physical system of a mag-
of this work is to propose a new ESO for ADRC in order to optimize the netic levitation ball system is shown in Fig. 1.
closed-loop system performance. Main contributions are The magnetic levitation device, shown in Fig. 1, is composed of an
electromagnet, a power amplifier, a laser sensor, and a hollow steel
• The time varying ESO (TESO) is proposed so as to deal with both ball. The experimental device is a single-axis magnetic levitation system,
generalized disturbance and measurement noise. i.e. the steel ball can just move up and down along the vertical direction.
• The convergence of TESO is confirmed theoretically, and the upper Current of the electromagnet can be adjusted so as to make the steel
bound of estimation error for TESO is depicted by the upper bound ball be stable in a given position or to drive the ball to track a desired
of generalized disturbance and measurement noise. trajectory. It is worth pointing out that, limited by the device itself, it is
• The transient performance of TESO is verified, and it can be only able to control the steel ball to move up and down. There will be
arbitrarily close to LESO. Moreover, better filtering performance no movements in other directions, since, in the laboratory, there is no
of TESO than that of LESO is proved. velocity and no force from other directions.
• Performance comparison and evaluation under different scenarios The magnetic levitation ball control system consists of two parts:
between time-varying ADRC (TADRC) and LADRC are performed the magnetic levitation device and a computer. Structure of magnetic
in both numerical and experimental cases. levitation ball control system is shown in Fig. 2.
In Fig. 2, variable x is the distance between the steel ball and the
The magnetic levitation ball control system with a contactless laser electromagnet surface, i.e. the position of the steel ball. We pick the
position measure system is taken as the experimental platform. LADRC origin to be the electromagnet surface, and the whole motion range is

25
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Table 1 Table 2
Variables and their descriptions. Nominal parameters of the magnetic levitation system.
Variable Descriptions Variables Ball mass Ball radius Resistance Coil turn K (N m2 /A2 ) Range of
(m) (R) (N) the gap
x Instantaneous gap between electromagnet surface and the steel ball
i Instantaneous current through the electromagnetic coil Value 0.094 kg 0.0125 m 13.8 Ω 2450 2.3142e−4 0–0.015 m
m The mass of ball
F The electromagnetic force
g The acceleration due to gravity
u The controlled voltage applied to electromagnetic coil
R The equivalent resistance of electromagnetic coil
L The self-inductance of electromagnetic coil
K The mutual inductance of electromagnetic coil

within the negative axis. An electromagnetic force F will be generated,

when current flow through the electromagnetic coil. By controlling Fig. 3. Structure of 3rd order LADRC.
voltage, we are able to adjust the electromagnetic force F to keep the
steel ball be in balance at different positions.
other words, self-inductance, mutual coefficients, electromagnetic force,
2.2. Model of magnetic levitation ball system
and resistance of electromagnetic coil are nonlinear and time-varying.
In addition, external time-varying disturbance signal 𝑑(𝑡) always exists.
A nonlinear model of such system derived from physical law is
Therefore, magnetic levitation system, a typical inherently nonlinear,
presented as follows
open-loop unstable and non-minimum phase system, is always operated
⎧ 𝑑2𝑥 under harsh conditions. Performance of magnetic levitation is suscep-
⎪𝑚 𝑑𝑡2 = 𝐹 (𝑖, 𝑥) − 𝑚𝑔
⎪ 𝑑𝑖 tible to different kinds of disturbance. Thus, control of the levitation
⎨𝑢(𝑡) = 𝑅𝑖 + 𝐿 (1) ball is a challenging and interesting problem. In following sections,
⎪ 𝑑𝑡
𝑖2 time-varying active disturbance rejection control has been proposed and
⎪𝐹 (𝑖, 𝑥) = 𝐾
⎩ 𝑥2 confirmed.
Variables in Eq. (1) and their descriptions are given in Table 1.
The first equation in Eq. (1) describes the dynamics of magnetic lev- 3. Time-varying active disturbance rejection control for a mag-
itation ball. The third equation characterizes the nonlinear relationship netic levitation ball system
among F, i, and x. If a ball is in stable levitation state, i.e. F (𝑖𝑒 , 𝑥𝑒 )-
mg = 0, the point (𝑖𝑒 , 𝑥𝑒 ) is an equilibrium point or working point. 3.1. Brief introduce to linear active disturbance rejection control
If disturbance exists, the balance will be broken, and the closed-loop
control system will adjust electromagnetic force to make the ball be in
For system (3), control law of LADRC can be designed as
a new balance.
For 𝐹 (𝑖, 𝑥), we have Taylor series expansion, i.e. 𝑘𝑝 (𝑦𝑟 − 𝑧𝑙1 ) − 𝑘𝑑 𝑧𝑙2 − 𝑧𝑙3
𝑢= (4)
𝑏0
𝐹 (𝑖, 𝑥) = 𝐹 (𝑖𝑒 , 𝑥𝑒 ) + 𝑘𝑖 (𝑖 − 𝑖𝑒 ) + 𝑘𝑥 (𝑥 − 𝑥𝑒 ) + 𝐹ℎ.𝑜.𝑡 (𝑖, 𝑥)
where 𝑘𝑝 , 𝑘𝑑 and 𝑏0 are tunable parameters, 𝑦𝑟 is the desired output,
𝜕𝐹 | 2𝐾𝑖𝑒 𝜕𝐹 | 2𝐾𝑖2𝑒
where 𝐹 (𝑖𝑒 , 𝑥𝑒 ) = 𝑚𝑔, 𝑘𝑖 = |
𝜕𝑖 |(𝑖𝑒 ,𝑥𝑒 )
= , 𝑘𝑥 = |
𝜕𝑥 |(𝑖𝑒 ,𝑥𝑒 )
=− and and 𝑧𝑙1 , 𝑧𝑙2 , 𝑧𝑙3 are states of linear ESO (LESO), which can be designed
𝑥2𝑒 𝑥3𝑒
𝐹ℎ.𝑜.𝑡 (𝑖, 𝑥) is the high order term of 𝐹 (𝑖, 𝑥). as (Gao, 2003)
Then, system (1) can be rewritten as ⎧𝑧̇ = 𝑧 + 𝛽 (𝑦 − 𝑧 )
⎪ 𝑙1 𝑙2 1 𝑙1
𝑚𝑥̈ = 𝑘𝑖 (𝑖 − 𝑖𝑒 ) + 𝑘𝑥 (𝑥 − 𝑥𝑒 ) + 𝐹ℎ.𝑜.𝑡 (𝑖, 𝑥) ⎨𝑧̇ 𝑙2 = 𝑧𝑙3 + 𝛽2 (𝑦 − 𝑧𝑙1 ) + 𝑏0 𝑢 (5)
⎪𝑧̇ 𝑙3 = 𝛽3 (𝑦 − 𝑧𝑙1 )
Considering that control signal is voltage signal 𝑢 and ubiquitous ex- ⎩
ternal disturbance signal 𝑑(𝑡) does disturb the desired system dynamics, where 𝛽1 , 𝛽2 and 𝛽3 are tunable gains of LESO, 𝑦 is the system output, and
we have 𝑢 is the control input. 𝑧𝑙1 , 𝑧𝑙2 and 𝑧𝑙3 estimate 𝑥1 , 𝑥2 and the generalized
𝑥̈ = 𝑓 (𝑖, 𝑥, 𝑑, 𝑡) + 𝑏𝑢 (2) disturbance 𝑓 (𝑖, 𝑥1 , 𝑢, 𝑑, 𝑡), respectively. Control block diagram of 3rd
order LADRC is given in Fig. 3.
2𝐾𝐿𝑖 𝑑𝑖 2𝐾𝑖2𝑒 2𝐾𝑖2𝑒 1
where 𝑓 (𝑖, 𝑥, 𝑑, 𝑡) = − 2 𝑒 𝑑𝑡 − − (𝑥 − 𝑥𝑒 ) + 𝐹 (𝑖, 𝑥) + 𝑑(𝑡), 𝑏
𝑚 ℎ.𝑜.𝑡
= Substituting control law (4) into system (3), we have the closed-loop
𝑥𝑒 𝑚𝑅 𝑚𝑥2𝑒 𝑚𝑥3𝑒
2𝐾𝑖𝑒
. system
𝑥2𝑒 𝑚𝑅 {
Let 𝑥 = 𝑥1 , 𝑥̇ 1 = 𝑥2 , 𝑦 = 𝑥1 +𝑛, and the magnetic levitation ball system 𝑥̇ 1 = 𝑥2
can be written as 𝑥̇ 2 = −𝑘𝑝 (𝑥1 − 𝑦𝑟 ) − 𝑘𝑑 𝑥2 + 𝑘𝑝 𝑒1 + 𝑘𝑑 𝑒2 + 𝑒3
⎧𝑥̇ = 𝑥 where estimation errors 𝑒1 , 𝑒2 , 𝑒3 are defined as
⎪ 1 2
⎨𝑥̇ 2 = 𝑓 (𝑖, 𝑥1 , 𝑢, 𝑑, 𝑡) + 𝑏0 𝑢 (3)
⎪𝑦 = 𝑥1 + 𝑛 ⎧𝑒 = 𝑥 − 𝑧
⎩ ⎪ 1 1 𝑙1
⎨𝑒2 = 𝑥2 − 𝑧𝑙2
where n is the measurement noise, y is the measurement output, ⎪𝑒3 = 𝑓 − 𝑧𝑙3
⎩
𝑓 (𝑖, 𝑥1 , 𝑢, 𝑑, 𝑡) = 𝑓 (𝑖, 𝑥1 , 𝑑, 𝑡) + ( 𝑏 − 𝑏0 )𝑢, and it can be viewed as gen-
eralized disturbance including unknown dynamics, omitted dynamics, Obviously, if an ESO is convergent, the closed-loop system is stable.
linearization error and the external disturbance, 𝑏0 is the estimation of Additionally, better linearization can be obtained, if smaller estimation
𝑏 and it is a tunable parameter. errors 𝑒1 , 𝑒2 and 𝑒3 can be guaranteed. Desired performance of an ESO
Nominal parameters of the system are given in Table 2. guarantees the closed-loop system performance. Therefore, an advanced
For system (1), we have made an assumption that all conditions, such design of an ESO will help optimize the system performance. Next, a
as the magnetic field, resistance of electromagnetic coil, are ideal. In modified ESO is discussed.

26
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

3.2.2. Convergence of TESO

Define the estimation errors of TESO as
⎧𝑒 = 𝑥 − 𝑧
⎪ 1 1 1
⎨𝑒2 = 𝑥2 − 𝑧2 (10)
⎪𝑒3 = 𝑓 − 𝑧3
⎩

According to system (3) and TESO (6), dynamics of estimation errors

Fig. 4. Structure of 3rd order TADRC.
can be written as

̇ = 𝑨𝑒 (𝑡)𝒆(𝑡) + 𝑩 2 𝑓̇ (∙) − 𝑩 𝑛 (𝑡)𝑛(𝑡)

𝒆(𝑡) (11)
3.2. TADRC design and analysis
where
3.2.1. TADRC design approach
As discussed in Section 3.1, the closed-loop system performance ⎡ −𝛽1 1 0⎤
can be greatly improved, if the states and generalized disturbance can 𝒆(𝑡) = [𝑒1 , 𝑒2 , 𝑒3 ]𝑇 , 𝑨𝑒 (𝑡) = ⎢−𝛽2 𝜌(𝑡) 0 1⎥ = 𝑨𝑒 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 1),
⎢ ⎥
be estimated and canceled timely and effectively. Generally, enlarging ⎣−𝛽3 𝜌(𝑡) 0 0⎦
the bandwidth of a LESO generates faster convergence of estimation ⎡−𝛽1 1 0⎤
errors, but it also results in much pollution by sensor noise. Therefore, 𝑨𝑒 = ⎢−𝛽2 0 1⎥ ,
⎢ ⎥
compromise should be made in the design of an ESO. ⎣−𝛽3 0 0⎦
Actually, better estimation of generalized disturbance is the domi-
nated work in initial state, and better noise filtering is the main task ⎡0⎤ ⎡1⎤ ⎡0⎤ ⎡ 𝛽1 ⎤
in steady state. In order to keep the estimation performance of ESO 𝑩 𝑒 = ⎢𝛽2 ⎥ , 𝑪 𝑒 = ⎢0⎥ , 𝑩 2 = ⎢0⎥ , 𝑩 𝑛 (𝑡) = ⎢𝛽2 𝜌(𝑡)⎥ .
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
in transient process and achieve better noise filtering ability in steady ⎣𝛽3 ⎦ ⎣0⎦ ⎣1⎦ ⎣𝛽3 𝜌(𝑡)⎦
state, an ESO with time-varying gain, i.e. time-varying ESO (TESO) is
proposed, Since all variables of the magnetic levitation ball system are limited
by physical conditions, we have following reasonable assumption.
⎧𝑧̇ = 𝑧 + 𝛽 (𝑦 − 𝑧 )
⎪ 1 2 1 1
⎨𝑧̇ 2 = 𝑧3 + 𝛽2 𝜌(𝑡)(𝑦 − 𝑧1 ) + 𝑏0 𝑢 (6) Assumption 1. ||𝑓̇ (∙)|| ≤ 𝐷𝑓 , |𝑛(𝑡)| ≤ 𝐷𝑛 , where 𝐷𝑓 and 𝐷𝑛 are positive
⎪𝑧̇ 3 = 𝛽3 𝜌(𝑡)(𝑦 − 𝑧1 )
⎩ constants.
where 𝑧1 , 𝑧2 and 𝑧3 are states of TESO, which estimate 𝑥1 , 𝑥2 and the
1
generalized disturbance of system (3), 𝑦 is the output of system (3), Theorem 1. If Assumption 1 holds and let < 𝜌∞ ≤ 𝜌0 < 3, then estimation
3
𝛽1 , 𝛽2 , 𝛽3 are tunable gains of TESO, 𝑏0 is another tunable parameter, 𝑢 error 𝒆(𝑡) satisfies
is the control input, and 𝜌(𝑡) is a smoothly decreasing function.
In this paper, we choose 𝜌(𝑡) as ‖𝒆(𝑡)‖ ≤ 𝛾1 exp(−𝛾2 𝑡) ‖ ‖
‖𝒆(𝑡0 )‖ + 𝛾3 (𝐷𝑓 + 𝐷𝑛 )
𝜌(𝑡) = (𝜌0 − 𝜌∞ )𝑒−𝑙𝑡 + 𝜌∞ (7)
where 𝛾1 , 𝛾2 , 𝛾3 are positive constants.
where 𝑙 is the decreasing rate and
Proof. Firstly, we will show that there exists a positive constant matrix
⎧𝜌0 ≜ 𝜌(0) > 0
⎪ 𝑷 𝑒 and a positive constant 𝑐0 such that
⎨𝜌∞ ≜ 𝑡→∞
lim 𝜌(𝑡) > 0, ∀𝑡 ≥ 0 (8)
⎪𝜌(𝑡)
⎩̇ ≤0 𝑨𝑒 𝑇 (𝑡)𝑷 𝑒 + 𝑷 𝑒 𝑨𝑒 (𝑡) ≤ −𝑐0 𝑷 𝑒
Obviously, gains of 𝑧2 and 𝑧3 in TESO are varying with respect to
time. 𝜌∞ is the parameter determining the final gains of 𝑧2 and 𝑧3 . It Define 𝑄1 (𝑠) = 𝑪 𝑇𝑒 (𝑠𝑰 − 𝑨𝑒 )−1 𝑩 𝑒 , it can be verified that
can be seen that 𝜌∞ determines the noise attenuation level. 1 − 𝑗3
𝜔𝑜
(𝑗𝜔 + 𝜔𝑜 )3 − (𝑗𝜔)3 − 3𝜔𝑜 (𝑗𝜔)2 𝜔
TADRC control law is designed as 𝑄1 (𝑗𝜔) = =1− 𝜔
(𝑗𝜔 + 𝜔𝑜 )3 (1 − 𝑗 𝜔𝑜 )3
𝑘𝑝 (𝑦𝑟 − 𝑧1 ) − 𝑘𝑑 𝑧2 − 𝑧3
𝑢= (9) = 1 − cos 3𝜃(cos(3𝜃) − 𝑗 sin(3𝜃))(1 + 𝑗3 tan 𝜃)
𝑏0 = 1 − cos 2𝜃(cos(3𝜃) − 𝑗 sin(3𝜃))(cos 𝜃 + 𝑗3 sin 𝜃)
Remark 1. LESO (5) is the particular case of TESO (6) when 𝜌0 = 𝜌∞ = and 𝜃 = arctan(−
𝜔𝑜
), 𝜃 ∈ (− 𝜋2 , 0).
̇ ≡ 0.
1, 𝜌(𝑡) 𝜔
Note that the Nyquist plot of {1 − cos2 𝜃(cos(3𝜃) − 𝑗 sin(3𝜃))(cos 𝜃
Remark 2. Let 𝛽1 = 3𝜔𝑜 , 𝛽2 = 3𝜔2𝑜 , 𝛽3 = 𝜔3𝑜 . Here, 𝜔𝑜 is the observer + 𝑗3 sin 𝜃)} lies in the closed disk 𝐷𝑛 (−0.5, 1.5), whose diameter is the
bandwidth. line segment connecting the points −0.5 + 𝑗0 and 1.5 + 𝑗0. Thus the
1+𝛼 𝑄 (𝑠)
circle criterion means that 𝑄2 (𝑠) ≜ 1+𝛼4 𝑄1 (𝑠) is strictly positive real when
3 1
Remark 3. Control law (4) of LADRC and control law (9) of TADRC are [𝛼3 , 𝛼4 ] ⊂ (− 32 , 2).
similar, and the key difference between TADRC and LADRC is the gain Since 𝑄2 (𝑠) has the following state–space realization
design of extended state observer.
{
𝜻̇ = (𝑨𝑒 − 𝛼4 𝑩 𝑒 𝑪 𝑇𝑒 )𝜻 + 𝑩 𝑒 𝑢∗
Remark 4. The denominators (i.e., 𝑏0 ) of control law (4) and (9) are the 𝑦∗ = (𝛼3 − 𝛼4 )𝑪 𝑇𝑒 𝜻 + 𝑢∗
estimation of the control input gain of the system. Therefore, 𝑏0 should
be chosen to be nonzero constant or time-varying variables to ensure There exists a positive matrix 𝑷 𝑒 and a vector 𝑳 such that
that control input always works. {
(𝑨𝑒 − 𝛼4 𝑩 𝑒 𝑪 𝑇𝑒 )𝑇 𝑷 𝑒 + 𝑷 𝑒 (𝑨𝑒 − 𝛼4 𝑩 𝑒 𝑪 𝑇𝑒 ) = −𝑳𝑳𝑇 − 𝑐0 𝑷 𝑒
√
Control block diagram of TADRC can be shown in Fig. 4. 𝑷 𝑒 𝑩 𝑒 = (𝛼3 − 𝛼4 )𝑪 𝑒 − 2𝑳

27
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Note that − 23 < lim𝑡→∞ 𝜌(𝑡) − 1 ≤ 𝜌(𝑡) − 1 ≤ 𝜌0 − 1 < 2, then Proof. Define 𝜼 = 𝒆 − 𝒆, we have change rate of 𝜼
𝜼̇ = 𝑨𝑒 𝜼 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 1)𝒆 − 𝑩 𝑒 (𝜌(𝑡) − 1)𝑛(𝑡)
𝑨𝑒 𝑇 𝑷 𝑒 + 𝑷 𝑒 𝑨𝑒 = (𝑨𝑒 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 1))𝑇 𝑷 𝑒 + 𝑷 𝑒 (𝑨𝑒 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 1))
≤ −𝑳𝑳𝑇 − 𝑐0 𝑷 𝑒 − 𝑪 𝑒 𝑩 𝑇𝑒 𝑷 𝑒 (𝜌(𝑡) − 1 − 𝛼4 ) By the proof of Theorem 1, we confirm that 𝑨𝑒 is Hurwitz. Addition-
ally, 𝜌(𝑡) satisfies
− 𝑷 𝑒 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 1 − 𝛼4 )
√
≤ −𝑐0 𝑷 𝑒 − 𝑳𝑳𝑇 − 𝑪 𝑒 [(𝛼3 − 𝛼4 )𝑪 𝑒 − 2𝑳]𝑇 sup ||𝜌(𝑡) − 𝜌(𝑡0 )|| ≤ (𝑡∗ − 𝑡0 ) sup |𝜌(𝑡)|
̇ .
𝑡∈[𝑡0 ,𝑡∗ ] ∗
𝑡∈[𝑡0 ,𝑡 ]
× (𝜌(𝑡) − 1 − 𝛼4 ) − [(𝛼3 − 𝛼4 )𝑪 𝑒
√ If 𝜌(𝑡0 ) = 1, it follows that sup𝑡∈[𝑡0 ,𝑡∗ ] |𝜌(𝑡) − 1| ≤ (𝑡∗ − 𝑡0 ) sup𝑡∈[𝑡0 ,𝑡∗ ] |𝜌(𝑡)|.
̇
− 2𝑳]𝑪 𝑇𝑒 (𝜌(𝑡) − 1 − 𝛼4 ) According to the knowledge of linear differential equations, we have
√
= −𝑐0 𝑷 𝑒 − ( 2(𝛼4 − (𝜌(𝑡) − 1))𝑪 𝑒 − 𝑳)𝑇 sup ‖ ‖ ∗
√ ‖𝒆(𝑡) − 𝒆(𝑡)‖ ≤ 𝛾4 (𝑡 − 𝑡0 ) sup∗ |𝜌(𝑡)|
̇ □
𝑡∈[𝑡0 ,𝑡∗ ] 𝑡∈[𝑡0 ,𝑡 ]
× ( 2(𝛼4 − (𝜌(𝑡) − 1))𝑪 𝑒 − 𝑳)
≤ −𝑐0 𝑷 𝑒 Obviously, Theorem 2 confirms that by tuning |𝜌(𝑡)|
̇ , 𝑡 ∈ [𝑡0 , 𝑡∗ ], the
estimation error of TESO can be arbitrarily close to that of LESO when
Let 𝑉 (𝒆) = 𝒆𝑇 𝑷 𝑒 𝒆, then 𝑡 ∈ [𝑡0 , 𝑡∗ ].
𝑑𝑉 (𝒆)
= 𝒆𝑇 (𝑷 𝑒 𝑨𝑒 + 𝑨𝑒 𝑇 𝑷 𝑒 )𝒆 − 2𝒆𝑇 𝑷 𝑒 𝑩 𝑒 𝜌(𝑡)𝑛 + 2𝒆𝑇 𝑷 𝑒 𝑩 2 𝑓̇ (∙) 3.2.4. Steady performance of TESO
𝑑𝑡
≤ 𝑐0 𝑉 (𝒆) + 2 ‖𝒆‖ ‖ ‖ ‖ ‖ ‖ ‖
‖𝑷 𝑒 ‖ (‖𝑩 𝑒 ‖ + ‖𝑩 2 ‖)(𝐷𝑓 + 𝐷𝑛 ) In this section, we proceed to verify the filtering performance of
TESO in steady state. Firstly, define 𝒆(𝑡) = [𝑒1 (𝑡), 𝑒2 (𝑡), 𝑒3 (𝑡)]𝑇 , we have,
which leads to
√ {
̇ = 𝑨𝑒,∞ 𝒆(𝑡) + 𝑩 2 𝑓̇ (∙) − 𝑩 𝑛,∞ 𝑛(𝑡)
𝒆(𝑡)
𝑑 𝑉 (𝒆) 𝑐 √ ‖𝒆‖ ‖ ‖ ‖ ‖ ‖ ‖
≤ − 0 𝑉 (𝒆) + √ ‖𝑷 𝑒 ‖ (‖𝑩 𝑒 ‖ + ‖𝑩 2 ‖)(𝐷𝑓 + 𝐷𝑛 ) 𝒆(0) = 𝒆(0)
𝑑𝑡 2 𝑉 (𝒆)
where 𝑨𝑒,∞ = 𝑨𝑒 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌∞ − 1), 𝑩 𝑛,∞ = lim𝑡→∞ 𝑩 𝑛 (𝑡). Note that
Let the maximal eigenvalue and minimal eigenvalue of 𝑷 𝑒 be 𝑐1 and
the ESO’s parameters in 𝑨𝑒,∞ is smaller than those in 𝑨𝑒 (𝑡), then
𝑐2 , respectively. Then, we have
√ 𝒆(𝑡) can be viewed as the desired steady performance of ESO with
√
𝑑 𝑉 (𝒆) 𝑐 √ 𝑐2 ‖ ‖ ‖ ‖ smaller bandwidth. Next, we will show the estimation error of TESO
≤ − 0 𝑉 (𝒆) + ( 𝑩 + 𝑩 )(𝐷 + 𝐷𝑛 )
𝑑𝑡 2 𝑐1 ‖ 𝑒 ‖ ‖ 2 ‖ 𝑓 𝒆(𝑡) converges to 𝒆(𝑡) as 𝑡 → ∞.

Thus the comparison principle shows

Theorem 3. Under Assumption 1, for estimation errors of TESO, we have
√
√ 𝑐 √ 2 𝑐2 lim𝑡→∞ ‖ ‖
‖𝒆(𝑡) − 𝒆(𝑡)‖ = 0.
𝑉 (𝒆) ≤ exp(− 0 𝑡) 𝑉 (𝒆(𝑡0 )) + √ (‖ ‖ ‖ ‖
‖𝑩 𝑒 ‖ + ‖𝑩 2 ‖)(𝐷𝑓 + 𝐷𝑛 )
2 𝑐0 𝑐1
Proof. Define 𝝃 = 𝒆 − 𝒆, then
Consequently,
√ √ 𝝃̇ = 𝑨𝑒,∞ 𝝃 − 𝑩 𝑒 𝑪 𝑇𝑒 (𝜌(𝑡) − 𝜌∞ )𝒆 − 𝑩 𝑒 (𝜌(𝑡) − 𝜌∞ )𝑛(𝑡)
𝑐2 𝑐 2 𝑐2
‖𝒆(𝑡)‖ ≤ √ exp(− 0 𝑡) ‖𝒆(𝑡 )‖ + √ (‖𝑩 ‖ + ‖𝑩 ‖)(𝐷 + 𝐷𝑛 ) By the proof of Theorem 1, we know that 𝑨𝑒,∞ is Hurwitz. Note that
𝑐1 2 ‖ 0 ‖ 𝑐0 𝑐1 ‖ 𝑒 ‖ ‖ 2 ‖ 𝑓
lim𝑡→∞ (𝜌(𝑡) − 𝜌∞ ) = 0, according to the property of linear differential
√
𝑐2
√
2 𝑐2 equation, we have lim𝑡→∞ ‖ ‖
‖𝒆(𝑡) − 𝒆(𝑡)‖ = 0. □
√ (‖𝑩 ‖ + ‖ ‖
𝑐0
If we let 𝛾1 = √ , 𝛾2 = , 𝛾3 = ‖𝑩 2 ‖), we have
𝑐1 2 𝑐0 𝑐1 ‖ 𝑒 ‖
Furthermore, noise filtering ability of TESO will be checked. Consid-
ering that 𝒆(𝑡) → 𝒆(𝑡), as 𝑡 → ∞, i.e. in steady state, for the comparison
‖𝒆(𝑡)‖ ≤ 𝛾1 exp(−𝛾2 𝑡) ‖ ‖
‖𝒆(𝑡0 )‖ + 𝛾3 (𝐷𝑓 + 𝐷𝑛 ) □ of 𝒆(𝑡) and 𝒆(𝑡), we can just compare the performance between 𝒆(𝑡) and
𝒆(𝑡).
Theorem 1 clearly shows that the estimation error of TESO is
Assuming that Laplace transformation of 𝒆(𝑡) = [𝑒1 (𝑡), 𝑒2 (𝑡), 𝑒3 (𝑡)]𝑇 ,
bounded, and the upper bound is related to the bound of generalized
𝒆(𝑡) = [𝑒1 (𝑡), 𝑒2 (𝑡), 𝑒3 (𝑡)]𝑇 and 𝑛(𝑡) be 𝑬(𝑠) = [𝐸 1 (𝑠), 𝐸 2 (𝑠), 𝐸 3 (𝑠)]𝑇 , 𝑬(𝑠) =
disturbance and measurement noise, respectively. Moreover, since the
[𝐸 1 (𝑠), 𝐸 2 (𝑠), 𝐸 3 (𝑠)]𝑇 and 𝑁(𝑠), respectively, i.e.
limitation on the time varying parameter 𝜌(𝑡) is so mild that the
practitioners only need to keep 𝜌(𝑡) ∈ (1∕3, 3). ℒ [𝑒1 (𝑡)] = 𝐸 1 (𝑠), ℒ [𝑒2 (𝑡)] = 𝐸 2 (𝑠), ℒ [𝑒3 (𝑡)] = 𝐸 3 (𝑠)
ℒ [𝑒1 (𝑡)] = 𝐸 1 (𝑠), ℒ [𝑒2 (𝑡)] = 𝐸 2 (𝑠), ℒ [𝑒3 (𝑡)] = 𝐸 3 (𝑠)
3.2.3. Transient performance of TESO and ℒ [𝑛(𝑡)] = 𝑁(𝑠).
Generally, estimation and cancellation of the generalized distur- Then, the transfer functions between 𝑬(𝑡), 𝑬(𝑡) and 𝑛(𝑡) can be
bance in initial state is crucial for the transient performance of ADRC. defined as
Reported literatures have shown that LESO based LADRC can ensure ⎡ 𝐸 1 (𝑠) ⎤
satisfactory transient performance. Thus, it is natural to check whether ⎢ ⎥
TESO can also keep desired transient performance. ⎢ 𝑁(𝑠) ⎥ ⎡𝛽1 𝑠2 + 𝛽2 𝑠 + 𝛽3 ⎤
⎢ 𝐸 2 (𝑠) ⎥ −1 ⎢ 𝛽 𝑠2 + 𝛽 𝑠 ⎥ 𝑁(𝑠) and
According to estimation errors of LESO, i.e. 𝑒1 , 𝑒2 , 𝑒3 , we have ⎢ ⎥= 3 2 3
⎢ 𝑁(𝑠) ⎥ 𝑠 + 𝛽1 𝑠 + 𝛽2 𝑠 + 𝛽3 ⎢⎣ ⎥
2
𝛽3 𝑠2 ⎦
̇ = 𝑨 𝒆 + 𝑩 𝑓̇ (∙) − 𝑩 𝑛(𝑡)
𝒆(𝑡) (12) ⎢ 𝐸 (𝑠) ⎥
𝑒 2 𝑙𝑒 ⎢ 3 ⎥
⎣ 𝑁(𝑠) ⎦
where 𝒆 = [𝑒1 , 𝑒2 , 𝑒3 ]𝑇 , 𝑩
𝑙𝑒 = [𝛽1 , 𝛽2 , 𝛽3 ]𝑇 .
Actually, 𝒆(𝑡) stands for the de-
⎡ 𝐸 1 (𝑠) ⎤
sired transient performance of ESO with ideal bandwidth parameterized. ⎢ 𝑁(𝑠) ⎥ ⎡ 1 𝛽 𝑠2 + 𝛽 𝑠 + 𝛽 ⎤
For the transient performance of TESO, we have following theorem. ⎢ 𝐸 (𝑠) ⎥ −𝜌 ⎢ 𝜌∞ 1 2 3
⎥
⎢ 2 ⎥= ∞
⎢ 𝛽 𝑠 2+𝛽 𝑠 ⎥ 𝑁(𝑠)
⎢ 𝑁(𝑠) ⎥ 𝑠 + 𝛽1 𝑠 + 𝜌∞ 𝛽2 𝑠 + 𝜌∞ 𝛽3 ⎢
3 2 2
2
3
⎥
Theorem 2. Let 𝜌(𝑡0 ) = 1 and 𝒆(0) = 𝒆(0), then estimation error of TESO ⎢ 𝐸 (𝑠) ⎥ ⎣ 𝛽 3 𝑠 ⎦
⎢ 3 ⎥
defined in (11) and estimation error of LESO defined in (12) satisfies ⎣ 𝑁(𝑠) ⎦
sup ‖ ‖ ∗
‖𝒆(𝑡) − 𝒆(𝑡)‖ ≤ 𝛾4 (𝑡 − 𝑡0 ) sup∗ |𝜌(𝑡)|
̇ Thus, we have
𝑡∈[𝑡0 ,𝑡∗ ] 𝑡∈[𝑡0 ,𝑡 ] / /
𝐸 (𝑠) 𝐸 2 (𝑠) 𝐸 (𝑠) 𝐸 3 (𝑠)
where 𝑡∗ is any given positive constant. lim 2 = 𝜌∞ < 1 and lim 3 = 𝜌∞ < 1
𝑠→∞ 𝑁(𝑠) 𝑁(𝑠) 𝑠→∞ 𝑁(𝑠) 𝑁(𝑠)

28
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Table 3
Signals and their descriptions.
Signals Descriptions (Unit: mm)
⎧
⎪𝑦𝑟1 = {
−10.02
⎪ −9.06 10𝑘 ≤ 𝑡 < 10𝑘 + 10(𝑘 = 0, 2, 4, 6 …)
𝑦𝑟 Reference input ⎨ 𝑟2
𝑦 =
⎪ −10.98 10𝑘 ≤ 𝑡 < 10𝑘 + 10(𝑘 = 1, 3, 5, 7 …)
⎪𝑦𝑟3 = 0.96 sin(3.14𝑡) − 10.02
⎩
𝑑 Disturbance signal 𝑑=3

Table 4
Parameters of LADRC and TADRC in simulations.
Controllers Parameters
𝑏0 𝜔𝑜 𝜔𝑐 𝜌0 𝜌∞ 𝑙
LADRC 6 460 46 – – –
TADRC 6 460 46 1 0.35 1

Table 5
Fig. 5. Control block diagram of LADRC/TADRC for a magnetic levitation ball system. Comparisons of performance indexes in simulations.
Reference signals Controller ITAE IAE RMSE
Step LADRC 1.350 × 10−3 6.499 × 10−4 3.802 × 10−6
which quantitatively reveals that TESO has better filtering performance TADRC 1.257 × 10−3 6.329 × 10−4 2.225 × 10−6
than LESO at high frequency range. Therefore, TESO is able to obtain Improvement 6.689% 2.616% 41.478%
Square LADRC 7.475 × 10−3 1.092 × 10−3 1.999 × 10−5
better filtering performance in steady state.
TADRC 7.031 × 10−3 1.049 × 10−3 2.056 × 10−6
So far, we have confirmed that TESO is convergent and capable of Improvement 5.940% 3.938% 89.715%
achieving desired estimation performance in initial state and getting Sinusoidal LADRC 1.364 × 10−3 8.234 × 10−4 1.289 × 10−4
better noise filtering effect in steady state. TADRC 1.305 × 10−3 8.075 × 10−4 9.774 × 10−5
Improvement 4.326% 1.931% 24.174%
Next, the tuning procedures of TADRC is summarized.

3.2.5. Tuning procedures of TADRC Table 6

Parameters of LADRC and TADRC in experiments.
From the analysis above, we can see that 𝑏0 , 𝑘𝑝 , 𝑘𝑑 , 𝛽1 , 𝛽2 , 𝛽3 , 𝜌0 , 𝜌∞ , 𝑙
Controllers Parameters
are tunable parameters of TADRC. Tuning procedures of TADRC can be
summarized as follows Groups 𝑏0 𝜔𝑜 𝜔𝑐 𝜌0 𝜌∞ 𝑙

Step 1. Design controller as shown in Eq. (9), and TESO according LADRC – 1700 600 10 – – –
to Eq. (6); A 1700 600 10 1 0.5 1
TADRC
Step 2. Let 𝑘𝑝 = 𝜔2𝑐 , 𝑘𝑑 = 2𝜔𝑐 , 𝛽1 = 3𝜔𝑜 , 𝛽2 = 3𝜔2𝑜 , 𝛽3 = 𝜔3𝑜 . Here, 𝜔𝑐 B 1700 600 10 1 0.35 1
is the control bandwidth, and 𝜔𝑜 is the observer bandwidth;
Step 3. Choose parameters of 𝜌(𝑡), i.e. 𝜌0 , 𝜌∞ , 𝑙, and 𝑏0 ;
Step 4. Adjust 𝑏0 , 𝜔𝑐 , 𝜔𝑜 , 𝜌0 , 𝜌∞ , 𝑙 to achieve desired performance.
where 𝑒𝑦 = 𝑦𝑟 − 𝑦 is tracking error of the closed-loop system.
According to the analysis and design procedures of TESO, numerical Parameters of TADRC and LADRC utilized in simulations are given
simulations and experiments have been performed in Sections 4 and 5, in Table 4.
respectively. In simulations, disturbance signal given in Table 3 has been intro-
duced from the 2nd seconds to the end. Responses and control signals
4. Numerical simulation for step, square, and sinusoidal signals are shown in Fig. 6.
Tracking performance are shown in Fig. 6(a), (c), and (e), the corre-
In this section, numerical simulations have been performed. For the sponding control signals are given in Fig. 6(b), (d), and (f), respectively.
purpose of confirming TADRC and comparing TADRC and LADRC, noise From the system response, we can see that both LADRC and TADRC are
has been introduced, and the polluted system output can be simulated able to track the desired outputs, even if disturbance exists. However,
as shown in Fig. 5. Here, the controlled plant is magnetic levitation from all control signals, we can see that TADRC is superior to LADRC on
ball system. 𝑦𝑟 is reference input, y is system measurement output noise filtering. It confirms that TADRC is able to achieve both satisfied
including measurement noise n, d is the external disturbance, and the tracking and noise filtering performance. Comparisons of performance
instantaneous position of ball is 𝑥1 . Reference inputs, disturbance signals index values are listed in Table 5.
and their descriptions are shown in Table 3. From system tracking responses, control signals, and the compar-
Gaussian noise, whose variance is 1, is taken to simulate the mea- isons of performance index values, we can see clearly that TADRC is
surement noise. Euler method is utilized in time steps of 0.001, and the able to get more desired performance. Next, this will be confirmed by
integral of time-multiplied absolute-value of error (ITAE) (Pan et al., experimental results.
2016), integral of absolute-value of error (IAE) (Pan et al., 2016), and
root mean square error (RMSE) (Rubio, 2018) are utilized to depict the 5. Experimental study
performance quantitatively.
𝑡 𝑡
5.1. Set up of experiments
| |
ITAE = 𝑡 |𝑒𝑦 | 𝑑𝜏 = 𝑡 |𝑦 − 𝑦|| 𝑑𝜏
∫0 | | ∫0 | 𝑟 Experiments are implemented on a laboratory test bench, i.e. a
𝑡 𝑡
| | |𝑦𝑟 − 𝑦| 𝑑𝜏 magnetic levitation ball control system with contactless laser position
IAE = |𝑒 | 𝑑𝜏 =
∫0√| 𝑦 | ∫0 | √ |
measure system (see Figs. 1 and 2). LESO based LADRC and TESO based
𝑡 𝑡 TADRC are realized in MATLAB/SIMULINK 2012 working environment.
1 1
RMSE = 𝑒2𝑦 𝑑𝜏 = (𝑦 − 𝑦)2 𝑑𝜏
𝑡 0∫ 𝑡 0 𝑟
∫ Sample frequency is 1000 Hz.

29
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 6. Responses and control signals in simulations.

Experiments can be divided into three steps. Firstly, step signal is Remark 6. Considering that when 𝜌0 = 𝜌∞ = 1, TESO is equivalent to
taken as the reference signal. Secondly, step disturbance is introduced LESO. For convenience, we just take 𝜌0 = 1 in all experiments. This is
to verify the robustness of LADRC and TADRC. From the first two also in accordance with Theorem 1.
steps, tracking and filtering performance of LESO and TESO have
been analyzed and a better group of parameters for TESO has been Remark 7. In order to make comparisons under the same condition,
𝑏0 , 𝜔𝑜 and 𝜔𝑐 of LADRC and TADRC are chosen to be same values in all
determined. Furthermore, with this group of parameters, square and
cases.
sinusoidal signals are taken as reference signals, experimental results
have been given to confirm TESO. According to Ref. Gao (2003) and Six cases (listed in Table 7) have been considered.
Section 3.2.5, we can choose the parameters of LADRC and TADRC.
5.2. Experimental results for different cases
Those values are shown in Table 6.
5.2.1. Experimental results for Case I
Remark 5. According to Theorem 2, small change rate of 𝜌(𝑡) is able In this case, reference input 𝑦𝑟 = −10.02 mm, no disturbance is
to minimize the difference between 𝒆(𝑡) and 𝒆(𝑡) in transient process. introduced. Parameters chosen in this case are shown in Table 6.
Therefore, in the experiments, l is fixed to be 1. Experimental results are given in Fig. 7.

30
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 7. Step response of LADRC and TADRC for Case I.

Table 7 fixed, i.e. parameters of Group A will be taken in the following two
Description for the cases considered in experiments.
cases. Tracking and filtering performance of the closed-loop system will
Cases Description
be checked.
I Step response i with control parameters of Group A
II Step response ii with control parameters of Group B
III Step response iii with disturbance given in Table 3 5.2.5. Experimental results for Case V
IV Step response iv with disturbance given in Table 3 ⎧
V Square response with control parameters of Group A
⎪−9.06 mm 10𝑘 ≤ 𝑡 < 10𝑘 + 10
⎪ (𝑘 = 0, 2, 4, 6...)
VI Sinusoidal response with control parameters of Group A In this section, square signal 𝑦𝑟 = ⎨
⎪−10.98 mm 10𝑘 ≤ 𝑡 < 10𝑘
⎪ +10(𝑘 = 1, 3, 5, 7...)
⎩
is taken as the reference signal. System responses and control inputs are
From the experimental results shown in Fig. 7, we can see that,
shown in Fig. 11.
with the help of TESO, TADRC is able to suppress sensor noise more
Fig. 11 confirms that both reference tracking and noise filtering
effectively than LADRC. In Case II, 𝜌∞ is decreased for checking the
performance are guaranteed by TADRC.
filtering performance.

5.2.6. Experimental results for Case VI

5.2.2. Experimental results for Case II
In this section, 𝑦𝑟3 = 0.96 sin(3.14𝑡) − 10.02 is taken as the reference
In this section, we keep the parameters be the same as those taken
input.
in Case I, but decrease 𝜌∞ to be 0.35 to check the filtering performance.
Experimental results are shown in Fig. 8. In other words, the reference signal will be changed from −10.98 mm
From Fig. 8, we can see that TADRC is also effective in noise filtering. to −9.06 mm in sinusoidal way. Experimental results have been given
Additionally, by comparing Figs. 8 and 7, we can see clearly that better in Fig. 12.
filtering performance can be achieved by choosing smaller 𝜌∞ . Fig. 12 shows that TADRC is also able to get better tracking and
In order to check LADRC and TADRC’s ability in disturbance rejec- filtering performance than LADRC when reference signal is sinusoidal
tion, disturbance signal is introduced in following two experiments. signal.
For the purpose of making a comparison, besides figures given above,
5.2.3. Experimental results for Case III ITAE, IAE and RMSE values for different cases are also calculated and
In this section, disturbance signal d given in Table 3 has been shown in Table 8. Obviously, by TESO, performance of the closed-loop
introduced and choose the same parameters as those taken in Case system is improved. ITAE, IAE and RMSE values presented in Table 8
I. Disturbance signal is added at 10 s, and experimental results are are also able to confirm the advantages of TESO over LESO.
presented in Fig. 9.
From Fig. 9(a) we can see that system response is robust to distur- 5.3. Discussion
bance. At the same time, noise filtering performance of TESO is still an
advantage over LESO. In this paper, parameters chosen in experiments and are different
from those adopted in the numerical simulations. 𝜔𝑜 taken in the simu-
5.2.4. Experimental results for Case IV lations are smaller than the value utilized in the experiments. It confirms
Here, the same disturbance as chosen in Case III has also been that simulation is relatively ideal, and larger observer bandwidth should
introduced at 10 s, and control parameters are taken the same as those be assigned in order to estimate uncertainties faster and more accurate.
chosen in Case II. Fig. 10 presents the experimental results. In addition, even if controller bandwidth 𝜔𝑐 is relatively smaller in
By comparing Figs. 9 and 10, we can confirm that smaller 𝜌∞ is able experiments, system performance has also been guaranteed, since the
to depress much measurement noise. However, smaller 𝜌∞ brings more observer bandwidth is enough.
oscillations and larger oscillation amplitudes. For attenuating measurement noise and improving the tracking per-
Thus, in order to achieve desired tracking and measurement noise formance, TESO has been proposed. Experiments have been performed
filtering performance, a proper 𝜌∞ should be chosen. on a laboratory test bench. Six cases have been considered. Figs. 7 and
From the comparisons among experimental results shown in Figs. 7– 8 confirm that taking smaller 𝜌∞ is able to get much better performance
10, we can see clearly that fixing 𝜌∞ to be 0.5 is a better choice. on sensor noise suppression. However, form Figs. 9 and 10, we can
Therefore, in following experiments, controller parameters have been see clearly that such disposal results in relatively worse disturbance

31
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 8. Step response of LADRC and TADRC for Case II.

Fig. 9. Step response of LADRC and TADRC for Case III.

Fig. 10. Step response of LADRC and TADRC for Case IV.

rejection performance, i.e. more oscillations and larger oscillation am- Furthermore, a technique, which is able to enlarge observer band-
plitudes, since observer bandwidth has been narrowed down. Therefore, width automatically when disturbance signal comes in again, is helpful
smaller observer bandwidth in steady state, i.e. by taking smaller 𝜌∞ , to improve the performance of active disturbance rejection based control
can guarantee better noise filtering performance, but it is achieved at the design approach. This is an interesting problem in forthcoming research.
cost of sacrificing the ability of disturbance estimation. Thus, for better Secondly, parameter l, the change rate of 𝜌(𝑡), is also a key param-
reference signal tracking, disturbance rejection and noise attenuation eter which determines the disturbance rejection ability. According to
ability, proper 𝜌∞ should be chosen, i.e. there is a compromise. Theorem 2, small change rate will guarantee the transient performance.

32
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 11. Step response of LADRC and TADRC for Case V.

Fig. 12. Step response of LADRC and TADRC for Case VI.

Therefore, in all cases, l is chosen the same value. However, what if l Table 8
Comparisons of performance indexes for different cases.
larger?
Cases Controller ITAE IAE RMSE
Here, for Case III, l is changed to be 2, and other parameters do not
change. Comparisons are shown in Fig. 13. Cases I LADRC 30.082 2.296 7.068 × 10−4
TADRC 24.991 1.625 5.697 × 10−4
Fig. 13(a) is the partial response when 𝑙 = 1, and Fig. 13(b) presents
Improvement 16.924% 29.225% 19.397%
the partial response when 𝑙 = 2. Obviously, smaller value of 𝑙 does Cases II LADRC 31.673 2.214 7.430 × 10−4
promote disturbance rejection ability. TADRC 19.914 1.854 4.555 × 10−4
Thirdly, from the analysis made above, a conclusion can be made, Improvement 37.126% 16.260% 38.695%
Cases III LADRC 30.988 2.804 8.391 × 10−4
i.e. for TESO, 𝜌∞ is the parameter which determines the noise at-
TADRC 24.690 2.392 7.122 × 10−4
tenuation level, and 𝑙 is the parameter which is able to promote the Improvement 20.324% 14.693% 15.123%
ability of disturbance rejection. Therefore, all parameters in TESO have Cases IV LADRC 34.603 2.732 3.900 × 10−3
clear physical explanation, which makes TESO easier to be accepted by TADRC 26.832 2.250 2.800 × 10−3
engineers. Improvement 22.458% 17.643% 28.205%
Cases V LADRC 47.970 2.712 1.363 × 10−3
Finally, in all cases, except 𝜌∞ , parameters chosen in experiments TADRC 44.093 2.490 1.358 × 10−3
(presented in Table 6) are almost the same, which shows that both Improvement 8.082% 8.186% 0.367%
LADRC and TADRC are of strong robust to disturbance, noise and the CASE VI LADRC 54.636 3.654 1.200 × 10−3
changed reference inputs. TADRC 52.408 3.021 1.100 × 10−3
Improvement 4.078% 17.323% 8.333%

6. Conclusion

In this paper, a time-varying active disturbance rejection control

(TADRC) solution is proposed to stabilize an open-loop unstable mag- better steady state performance. Overall, the proposed solution leads
netic levitation system. Both the existing LADRC and the new TADRC to better tracking and disturbance rejection performance, clear insight
are designed, implemented and tested. It is shown that both techniques on the design principle and the ease of tuning. Such characteristics,
are able to provide consistent performance over the entire operating not obtained from a detailed and accurate mathematical model, but
range. In particular, TESO is able to optimize the existing LESO with from the idea of disturbance active estimation and rejection, make the

33
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Fig. 13. Comparison of TADRC for Case III in different l.

solution more effective and attractive for engineering applications than Khamesee, M., & Shameli, E. (2005). Regulation technique for a large gap magnetic field
the model-based approach. for 3D non-contact manipulation. Mechatronics, 15(9), 1073–1087.
Kim, W., Verma, S., & Shakir, H. (2007). Design and precision construction of novel
Nevertheless, much better performance can be guaranteed if the
magnetic-levitation-based multi-axis nanoscale positioning systems. Precision Engi-
parameters of TESO, especially for 𝜌∞ and l, are able to adjust adaptively neering, 31(4), 337–350.
or actively. From this point of view, self-learning ability is preferable in Kummer, M., Abbott, J., Kratochvil, B., Borer, R., Sengul, A., & Nelson, B. (2010).
determining the parameters of TESO. Therefore, a self-learning TESO is OctoMag: an electromagnetic system for 5-DOF wireless micromanipulation. IEEE
the work in the near future. Transactions on Robotics, 26(6), 1006–1017.
Lanzara, G., D’Ovidio, G., & Crisi, F. (2014). UAQ4 levitating train: Italian maglev
transportation system. IEEE Vehicular Technology Magazine, 9(4), 71–77.
Acknowledgments Martinez-Vazquez, Daniel L., Rodriguez-Angeles, A., & Sira-Ramirez, Hebertt (2009).
Robust GPI observer under noisy measurements. In Proc. International conference on
This work was supported in part by National Natural Science Founda- electrical engineering, computing science and automatic control.
Michail, K., Zolotas, A. C., Goodall, R. M., & Whidborne, J. F. (2012). Optimised
tion of China (61873005, 61403006), Key program of Beijing Municipal
configuration of sensors for fault tolerant control of an electro-magnetic suspension
Education Commission, China (KZ201810011012), and the support system. International Journal of Systems Science, 43(10), 1785–1804.
Project of High-level Teachers in Beijing Municipal Universities in the Morales, R., & Sira-Ramirez, H. (2010). Trajectory tracking for the magnetic ball levitation
Period of 13th Five-year Plan, China (CIT&TCD201704044), system via exact feedforward linearization and GPI control. International Journal of
Control, 3(6), 1155–1166.
Pan, Y., Liu, Y., Xu, B., & Yu, H. (2016). Hybrid feedback feedforward: an efficient design
Appendix A. Supplementary data of adaptive neural network control. Neural Networks, 76(4), 122–134.
Pan, Y., Sun, T., & Yu, H. (2016). Composite adaptive dynamic surface control using
Supplementary material related to this article can be found online at online recorded data. International Journal of Robust and Nonlinear Control, 26(18),
https://doi.org/10.1016/j.conengprac.2018.09.018. 3921–3936.
Phuah, J., Lu, J., & Yahagi, T. (2005). Chattering free sliding mode control in magnetic
levitation system. IEEE Transactions on Electronics, Information and Systems, 125(4),
References 600–606.
Pu, Z. Q., Yuan, R. Y., Yi, J. Q., & Tan, X. M. (2015). A class of adaptive extended state
Abdel-Hady, F., & Abuelenin, S. (2008). Design and simulation of a fuzzy-supervised PID observers for nonlinear disturbed systems. IEEE Transactions on Industrial Electronics,
controller for a magnetic levitation system. Studies in Informatics and Control, 17(3), 62(9), 5858–5869.
315–328. Qin, Y. M., Peng, H., Ruan, W. J., Wu, J., & Gao, J. C. (2014). A modeling and control
Bächle, Thomas, Hentzelt, Sebastian, & Graichen, Knut (2013). Nonlinear model predictive approach to magnetic levitation system based on state-dependent ARX model. Journal
control of a magnetic levitation system. Control Engineering Practice, 21(9), 1250– of Process Control, 24(1), 93–112.
1258. Rubio, J. J. (2018). Robust feedback linearization for nonlinear processes control. ISA
Berkelman, P., & Dzadovsky, M. (2013). Magnetic levitation over large translation and Transactions, 74(3), 155–164.
rotation ranges in all directions. IEEE/ASME Transactions on Mechatronics, 18(1), 44– Rubio, J. J., Zhang, L., Lughofer, E., Cruz, P., Alsaedi, A., & Hayat, T. (2017). Modeling
52. and control with neural networks for a magnetic levitation system. Neurocomputing,
Bonivento, C., Gentili, L., & Marconi, L. (2005). Balanced robust regulation of amagnetic 227(3), 113–121.
levitation system. IEEE Transactions on Control Systems Technology, 3(6), 1036–1044. Schuhmann, T., Hofmann, W., & Werner, R. (2012). Improving operational performance
Chen, S. -Y., Lin, F. -J., & Shyu, K. -K. (2009). Direct decentralized neural control for of active magnetic bearings using Kalman filter and state feedback control. IEEE
nonlinear MIMO magnetic levitation system. Neurocomputing, 72(13–15), 3220–3230. Transactions on Industrial Electronics, 59(2), 821–829.
Elahi, T., & Nekoubin, A. (2011). Design of magnetic levitation system controller using Torres, L. H. S., Schnitman, L., Junior, C. A. V. V., & Felippe de Souza, J. A. M. (2012).
sliding mode control. International Review on Modelling and Simulations, 4(4), 1550– Feedback linearization and model reference adaptive control of a magnetic levitation
1557. system. Studies in Informatics and Control, 21(1), 67–74.
Gao, Z. (2003). Scaling and bandwidth-parameterization based controller tuning. In Porc. Trumper, D. L., Olson, S. M., & Subrahmanyan, P. K. (1997). Linearizing control of
American control conference. magnetic suspension systems. IEEE Transactions on Control Systems Technology, 5(4),
Glueck, T., Kemmetmueller, W., Tump, C., & Kugi, A. (2011). A novel robust position 427–438.
estimator for self-sensing magnetic levitation systems based on least squares identifi- Ulbig, A., Olaru, S., Dumur, D., & Boucher, P. (2010). Explicit nonlinear predictive control
cation. Control Engineering Practice, 9(2), 146–157. for a magnetic levitation system. Asian Journal of Control, 12(3), 434–442.
Golob, M., & Tovornik, B. (2003). Modeling and control of the magnetic suspension Wei, W., Liang, B., Li, D. H., & Su, W. J. (2016). Improving the efficiency of extended
system. ISA Transactions, 42(1), 89–100. state observer under noisy measurements by low-pass filter. In Proc. Chinese control
Han, J. (1995). A class of extended state observers for uncertain systems. Control and and decision conference.
Decision, 10, 85–88, (In Chinese). Weng, Ming-Chih, Lu, Xiaodong, & Trumper, David L. (2002). Vibration control of flexible
Han, J. (1998). Auto-Disturbance rejection control and its applications. Control and beams using sensor averaging and actuator averaging methods. IEEE Transactions on
Decision, 13, 19–23, (In Chinese). Control Systems Technology, 10(4), 568–577.
Han, J. (2009). From PID to active disturbance rejection control. IEEE Transactions on Xu, Jinquan, Hwa Chen, Ye, & Guo, Hong (2015). Robust levitation control for maglev
Industrial Electronics, 56(3), 900–906. systems with guaranteed bounded airgap. ISA Transactions, 59(11), 205–214.

34
W. Wei et al. Control Engineering Practice 82 (2019) 24–35

Xue, W. C., Bai, W. Y., Yang, S., Song, K., Huang, Y., & Xie, H. (2015). ADRC with adaptive Zhang, J. H., & Zhang, W. A. (2017). Linear active disturbance rejection control to
extended state observer and its application to air–fuel ratio control in gasoline engines. magnetic levitation ball system and parameters tuning. Journal of System Science and
IEEE Transactions on Industrial Electronics, 62(9), 5847–5857. Mathematical Science, 37(8), 1741–1756.
Yang, Z. J., Fukushima, Y., Kanae, S., & Wada, K. (2009). Robust non-linear output- Zhao, Z. L., & Guo, B. Z. (2015). On active disturbance rejection control for nonlinear
feedback control of a magnetic levitation system by K-filter approach. IET Control systems using time-varying gain. European Journal of Control, 23(2), 62–70.
Theory & Applications, 3(7), 852–864.

35

View publication stats