Received October 31, 2017, accepted December 19, 2017, date of publication December 25, 2017,

date of current version February 14, 2018.

Digital Object Identifier 10.1109/ACCESS.2017.2787052

Decentralized PID Control Design for Magnetic

Levitation Systems Using Extremum Seeking
QIANG CHEN 1 , YING TAN2 , (Senior Member, IEEE), JIE LI1 ,
AND IVEN MAREELS2 , (Fellow, IEEE)
1 College of Mechatronics Engineering and Automation, National University of Defense Technology, Changsha 410073, China
2 Melbourne School of Engineering, The University of Melbourne, Melbourne, VIC 3010, Australia
Corresponding author: Qiang Chen ([email protected])
This work was supported by the National Key Research and Development Program of China under Grant 2016YFB1200600.
The work of Q. Chen was supported by the China Scholarship Council.

ABSTRACT A decentralized proportional–integral–differential (PID) control structure has been always the
first option for magnetic levitation systems due to its model-free nature as well as the simplicity. It is very
challenging to tune PID parameters for each suspension point to ensure that the closed-loop system is less
sensitive to uncertainties/disturbances. Even though, by a large number of experiments, a reasonably good
PID control parameter set can be obtained to levitate a single suspension point for a given set point, these
parameters are very sensitive when the set point changes or the decentralized PID controller is applied to
the half bogie where coupling exists, strong oscillations will be observed and retuning of parameters is
needed. This paper utilizes the extremum seeking (ES) method to tune PID parameters on-line to improve
the steady-state performance of the system with few oscillations. A ‘‘sampled-data’’ structure is used to
obtain a moving window approximation of the steady-state behavior, leading to a discrete-time ES tuning
for the decentralized PID parameters. By using projection method, the tuning parameters are constrained
in a compact set, in which the stability can be guaranteed and the optimal performance can be achieved.
Experiments have shown the effectiveness of the proposed method.

INDEX TERMS Magnetic levitation systems, decentralized PID control, extremum seeking for PID tuning.

I. INTRODUCTION used [4] due to its simplicity in design, parameter tuning, and
A Magnetic Levitation (Maglev) train uses electromagnetic implementation. And in practice, model-free decentralized
forces to stably levitate the train above the track at the given PID controllers are widely used in the EMS systems [5]–[7].
set-point. Compared with traditional wheel rail trains, Maglev This work tries to achieve a good steady-state perfor-
trains offer many advantages such as safety, comfort, low mance to levitate a Small-Scale Magnetic Levitation Plat-
noise, small turning radius and strong climbing ability [1]. form (SSMLP), as shown in Fig. 1.
Among different design methods, ElectroMagnetic Suspen-
sion (EMS) structure has been a popular way. It has been used
in Transrapid, HSST, CMS and so on [2], [3].
Any EMS system is open-loop unstable. Thus the control
design plays an important role in order to regulate the gap
at the desired set-point. It is worthwhile to highlight that the
model for the EMS system is highly nonlinear and uncer-
tain even for a single suspension point. As the number of
suspension points increases, the complexity of model will
increase dramatically due to mechanical coupling among
suspension points, making it infeasible to design a model-
based controller for Maglev systems. A decentralized control
structure, which implies that each suspension point has its FIGURE 1. Small scale maglev experiment platform.
own control loop and typically all suspension points share This SSMLP is a full bogie structure with EMS levi-
the identical control structure and parameters, is thus widely tation structure. A bogie consists of two half bogies with
2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 3059
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

to improve transient responses as in [11] and [12], the ES

loops start to work when the closed loop system already
reaches the steady-state. Hence parameters of PID are tuned
on-line to minimize the tracking error as well as to eliminate
the vibrations in the steady-state. Several techniques have
been used to ensure the stability of the closed-loop as well
as achieving an optimal steady-state behaviours.
• In order to ensure the stability of the closed-loop, the sta-
bility region obtained from a large amount experiments
has been used in the tuning of PID parameters. A well-
known projection method is used to ensure that PID
FIGURE 2. Structure schematic diagram of the bogie.
parameters are still within the stability region.
• A moving window is used to approximate the steady-
two suspension points located at each corner of the half bogie.
state behaviour and generating a sampled cost function
Each suspension point has its own control loop including
representing the averaged behaviour of a given time
sensors to measure the gap and the current, signal processing
window. This will lead to a discrete-time ES tuning for
blocks and power converters. A schematic diagram of the
the decentralized PID parameters. With an hold used,
bogie is shown in Fig. 2. This SSMLP has some unique
it is shown that the steady-state behavior can reach to a
features as well.
small neighbourhood of the optimal one by tuning PID
1) The track is very thin and is fixed on the top of a support parameters carefully.
structure, which is placed on the ground. This mechanic • In order to achieve a good convergence, the initial choice
setting is prone to vibrations due to the fact that the of PID parameters is obtained from the trial-and-error
ground is not absolutely flat. method.
2) Two half bogies are connected by rigid beams. Thus
four suspension points are mechanically coupled. Experimental results for half bogie show that the pro-
3) Bogie and cabin is connected by springs, there is no posed on-line tuning for decentralized PID algorithm indeed
well-design secondary system as in many other EMS achieves better steady-state responses, which are less sensi-
systems. tive to disturbances and set-point changes.
The rest of paper is organized as follows. Section 2 pro-
In a word, compared with commercial products in mar-
vides background for decentralized PID control design
ket, the mechanical structure and electrical systems of this
including the model for single suspension point. The auto-
SSMLP are on purposely ‘‘badly’’ designed in order to serve
tuning PID based on ES is introduced in Section 3 followed
as a good testbed for testing various control algorithms.
by experimental results in Section 4. Section 5 summarizes
In this decentralized controller setting, a PID controller
the paper.
has been used to achieve the desired set-point, though the
parameters of PID controllers are tuned by trial-and-error
II. DECENTRALIZED CONTROL DESIGN
methods. However, during debugging, it is found that these
A. MODELS FOR SINGLE SUSPENSION POINT
tuned parameters are quite sensitive to disturbances, set-point
In order to understand the characteristics of the Maglev sys-
changes, and environment changes. Moreover, the stability
tem, a first-principle based model is needed. As the decen-
region for PID parameters is relatively small. It is often
tralized controller is used, only the single suspension point
observed in the experiments that vibrations would appear.
model is considered. The following assumptions are needed
It is noted that though the levitation is most challenging
to simplify the modelling procedure.
part of designing a control system for the Maglev system,
keeping a good steady-state performance is always required 1) Leakage flux can be ignored.
in Maglve transportation systems as it is closely related to the 2) The magnetic saturation can be ignored.
comfortableness of the passengers. 3) Magneto-resistance in core and track can be ignored.
Extremum Seeking (ES) is a model-free on-line opti- 4) Without considering the motions of rolling, yawing,
mization technique [8]. It can seek an optimal input pitching between electromagnets and the track.
for an input-to-steady-output mapping using input output 5) Without considering the elastic vibration or dynamic
measurements [9], [10]. It has been applied to tune param- deformation of the track.
eters of PID iteratively by using the performance mea- The schematic diagram of a single suspension point is
sured from the last trial [11], [12]. Various performance represented in Fig. 3.
requirements over a fixed time period were also discussed. The parameters of the SSMLP is shown in Table 1.
Different from [11] and [12], this paper focuses on improving Denoting δ(t) as the gap between electromagnet and track,
the steady-state performances such as reducing the tracking u(t) as the voltage applied to the system. The notion i(t) is the
error and the vibrations, which are all related to ensure the current of electromagnet coil and F(t) is the corresponding
comfort ride for passengers. Instead of repeating experiments electromagnetic force. The mass of a single suspension point

3060 VOLUME 6, 2018

Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

the bogie as 4 independent masses and many assumptions are

made.
Remark 1: Although better models including the half
bogie model and the full bogie model [13] are available,
these models are still, might be better, approximations of the
Maglev system. The overall system consists of the dynamics
of the track, the secondary suspension system, the chopper
and the power electronics. Each component is highly nonlin-
ear. Thus it is very challenging to build a relatively accurate
model so that the well-developed model-based control algo-
rithms are applicable. ◦
FIGURE 3. The Schematic diagram of suspension point system. In practical commercialized Maglev systems, simple
TABLE 1. Structural parameters of the SSMLP. model-free decentralized PID control algorithms are widely
used. The selection of PID parameters is usually by trial-and-
error methods. Large amount experiments are needed to find
a reasonably good parameters for a given set-point. In our
experiments, it is observed that the tuned PID parameters are
not robust with respect to changes of set-point or other uncer-
tainties. Vibrations always happen under such situations. This
paper tries to find an optimal set of PID parameters that can
achieve the best steady-state performance.

B. DECENTRALIZED PID CONTROLLER

m1 needs to balance a quarter of the overall weight of the The control objective for the decentralized PID controller is
bogie. to track the desired set-point r with no steady-state error.
1 Though the control input is voltage, the current through the
(m + M ).
m1 = (1) electromagnet coil generates the magnetic force. Due to the
4
influence of the coil inductance, the change of the current
According to Newton’s second Law, it has
always have a lag behind that of the voltage change. More-
m1 δ̈(t) = m1 g − F(t), (2) over, it is desired to have a quick current so that there will be
enough magnetic force to balance the gravity as the current
where the electromagnetic force F(t) can be expressed starts from zero. Therefore, a cascaded control [14], [15] is
as [13]: designed, as shown in Fig. 4.
µ0 AN 2 i(t) 2
 
F(t) = . (3)
2 δ(t)
On the other hand, the control input u satisfies the following
condition [13]:
u(t) µ0 AN 2 µ0 AN 2 i(t)
= Ri(t) + i̇(t) − δ̇(t), (4) FIGURE 4. Block diagram of cascade PID control.
2 2δ(t) 2δ 2 (t)
With the consideration of the cascaded control struc-
The state-space representation of the single-point model is ture, the single-point model (5) can be re-written as inter-
thus obtained as connected two subsystems 61 and 62 .
 
x2
µ0 AN 2 i(t) 2
   
0
 g− 0
 µ AN 2 x
3
 61 : δ̈(t) = g − , (7)
ẋ =  [ ]2  +
 0 
 u (5) 2m1 δ(t)
2m1 x1 1
The input is current i(t) and the output is gap δ(t). The PID
   
 x2 x3 2R  x 1
− x1 x3 µ0 AN 2
x1 µ0 AN 2 controller is designed for (61 ).
 
y= 1 0 0 x (6) e(t) = r − δ(t) (8)
Z t
where Iexp (t) = kp · e(t) + kd · ė(t) + ki · e(τ ) dτ, (9)
T T 0
= δ(t) δ̇(t) .
 
x = x1 x2 x3 i(t)
where Iexp (t) is the expected current for (61 ), and the param-
A single-point model is thus obtained. It is a third-order eters kp , ki , kd are proportional, integral and derivative gains
nonlinear system with the output y as the gap and input u is respectively. The outer-loop controller will drive the gap to
the voltage. obviously, this model is too simplistic, as it views the desired set-point.

VOLUME 6, 2018 3061

 Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

On the other hand, 62 :

δ̇(t)i(t) 2R 1
62 : i̇(t) = − δ(t)i(t) + δ(t)u(t).
δ(t) µ0 AN 2 µ0 AN 2
(10)
The input is voltage u(t) and the output is current i(t), which
is the input of 61 . A P controller is designed for (62 ). It takes
the following form:
eI (t) = Iexp (t) − i(t) (11)
u(t) = kc · ei (t), (12)
where the parameter kc is the gain of the inner-loop. The faster
inner-loop will drive the current quickly to desired reference
current to generate enough magnetic force.
Usually, the inner loop gain kc is firstly selected to ensure
fast enough current response speed. For example, in the
experiments of the SSMLP, kc = 50, the response time of
current loop is about 10ms. Then a trial-and-error method has
been used to tune PID parameters for single suspension point
using a large number of experiments for levitating a single
point.
After experiments, a good parameter set is thus obtained:
kp = 2000, ki = 0.1, and kd = 20. As can be seen in Fig. 5,
a reasonable tracking performance is obtained when the set-
point r = 6mm and r = 4mm for suspension point 1. From
a large amount of experiments, we have obtained a set of
reasonable PID parameters that can ensure the boundedness
of tracking errors as well as the boundedness of control
input. This set will be used later to ensure the boundedness FIGURE 5. Experiment results of suspension point 1 using cascaded PID
of trajectories of the SSMLP when searching an optimal parameters obtained by trial-and-error (single-point). (a) Setting
PID parameter set. It is noted that as the SSMLP is open- expectation gap to 6mm. (b) Setting expectation gap to 4mm.

loop unstable, limiting the search space of the PID param-

eters is very important to prevent unstable behaviour from
happening. levitated in the presence of set-point changes or disturbance
Due to mechanic coupling between suspension point 1 and happens.
suspension point 2, even there is no current applied to sus- These experiments clearly suggest that as the overall
pension point 2, it is also levitated, though it could not reach system is too complex, a simple cascaded PID controller
the desired set-point. Moreover, it is notices that the initial might not track well in a large region of interests even
position of suspension point 1 and suspension point 2 are though the uniform boundedness of the gap can be ensured.
δ1,0 = 10.2mm and δ2,0 = 9.8mm respectively when both of As vibrations are undesired steady-state behaviours, it is
them are at rest on the same track. Obviously, the mechanic important to find a better set of cascaded PID parame-
coupling is strong. Besides, due to the uneven distribution ters to remove the vibrations and achieve a better tracking.
of gravity and other mechanical and electrical differences As the model cannot be used, a data-driven method is thus
between these suspension points. Thus it is not surprising needed to find an optimal set of parameters. It is noted
that the tuned PID parameters are not robust when more dis- that to simplify our analysis, only PID parameters are tuned
turbances coming in such as the set-point changes as shown on-line for the given performance indexes. The kc in the
in experiments. More precisely, as the parameters are well- inner loop is kept as a constant as experiments have shown
tuned for suspension point 1, the performance of suspension that the performance of the SSMLP is not sensitive to this
point 1 can be ensured when set-point changes. However, parameter.
when these parameters applied to suspension point 2, the set-
point changes from r = 6mm to r = 4mm, vibrations happen III. DECENTRALIZED PID CONTROL DESIGN BASED ON
at suspension point 2, leading to oscillations at suspension EXTREMUM SEEKING
point 1 as well, as shown in Fig. 6. A. AUTO-TUNING PID USING EXTREMUM SEEKING
Similar vibrations can be observed from experiments when Killingsworth and Krstic [11], [12] proposed an auto-tuning
half bogie (two suspension points with two cascaded PID con- PID method using ES algorithm in order to achieve some
trollers for both suspension point 1 and suspension point 2) is optimal performances that are closely related to overshoot

3062 VOLUME 6, 2018

Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

B. DECENTRALIZED OPTIMAL PID BASED ON EXTREMUM

SEEKING
Let θ = [θ1 , θ3 , θ3 ]T = [kp , ki , kd ]T , the following cost
function for each suspension point is proposed:
Z t+T
1
J (θ ) = lim sup (α1 e2 (τ ) + α2 ė2 (τ ))dτ (13)
T t→∞ t
where T is called a sampling period. The positive parameters
(α1 , α2 ) are weighting coefficients between the tracking error
and vibrations.
Remark 2: As the measurement e(t) is very small while
ė(t) can be quite large due to vibrations, the weight α1 and α2
can be used to normalize the tracking error and its derivative.
Different choices of α1 and α2 will lead to slightly different
performance. Due to space limitation, this paper will not
discuss the effect of performance when these two parameters
change. How to choose these two parameters in a systematic
way will be our future work. ◦
The cost function (13) is a static mapping of tun-
ing parameter
n θ. On the other hand, denoting o the set
S := (kp , ki , kd ) [k p , k̄p ] × [k i , k̄i ] × [k d , k̄d ] .

As clearly indicated in Section II, when PID tuning param-
eters are in this set, the cost function J (θ) is always bounded,
i.e., unstable performance cannot happen.
Although there exist various modelling uncertainties and
measurement noises, the closed-loop system can be re-written
as

FIGURE 6. Experiment results of suspension point 2 using cascaded PID

ẋ = f (x, θ)
parameters for suspension point 1. (a) Setting expectation gap to 6mm. y = J (θ). (14)
(b) Setting expectation gap to 4mm.

Next it is assumed that the optimal set of parameters exists as

shown in Assumption 1
and rise time. By repeating the experiments of step response, Assumption 1: There exists a unique θ ∗ ∈ S such that the
the performance of a fixed finite interval is recorded and will cost function (13) coming from the system (14) is minimized,
be used to tune PID parameter for the next trial. A discrete- i.e.,
time ES algorithm was implemented until the optimal PID
parameters for the given cost are obtained. It is noted that ∇(θ ∗ ) = 0, ∀θ ∈ S (15)
the role of ES in [11] and [12] is to improve the transient ψ ∇ J (θ )ψ < 0, ∀ψ ∈ S.
T 2 ∗
(16)
response. It is highlighted that the cost J (θ ) cannot be measured on-
It is noted that the proposed auto-tuning PID cannot line. Hence, the following measurable cost is used:
directly applied to our problem. In the levitation of SSMLP, 1 (k+1)T
Z
the transient response is not the major consideration in the Jk (θ) = (α1 e2 (τ ) + α2 ė2 (τ ))dτ (17)
T kT
design. In the major commercial Maglev transportation sys-
tems, the transient response time is always much shorter than As lim sup Jk (θ) = J (θ), thus Jk (θ) can approximate the cost
the steady-state response. A good steady-state performance is k→∞
J (θ ) when k is sufficiently large.
always needed as it is related to the comfort of the ride, which The following projection operator [18], [19] is used for any
is one of the key design factors in Maglev transportation θ ∈ R.
systems. Moreover, vibrations always happen during steady- 
state due to unexpected disturbances coming from set-point θ̄ if θ > θ̄

changes, environment changes, self-excited vibration [16] Proj(θ, θ, θ̄) = θ if θ ≤ θ ≤ θ̄ (18)
and other factors [17]. 

θ if θ < θ
Hence, the goal of ES algorithm is to find an PID parameter
set to achieve an optimal steady-state performance, which is a for any given θ < θ̄.
balance between minimizing the tracking error and minimiz- For each suspension point, on-line PID tuning using ES is
ing the vibrations. shown in Fig. 7.

VOLUME 6, 2018 3063

 Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

• Assuming that for a given θ, J (θ) (instead of Jk (θ)) is

measurable. By using the proof technique from discrete-
time ES [20] and using non-local stability analysis
method [10], it is not difficult to show that semi-global
practical convergence to the optimal parameter θ ∗ . The
role of projection is to show the boundedness of the cost
function.
• By using the closeness between J (θ ) and Jk (θ) when
k → ∞, we can conclude the inequality (20).
This completes the proof. ◦
Remark 3: As discussed in [10], the ES algorithm pro-
posed is a ‘‘gradient-decent’’ like algorithm. The step size
of gradient is related to the amplitude of an while δn plays
an important role in averaging analysis. It represents how
close between the averaged trajectories and actual trajectories
would be.
Remark 4: Different from [21], only a static mapping is
FIGURE 7. An on-line PID tuning for SSMLP with ES.
considered. Thus the singular perturbation technique [22] is
no longer needed. Hence the convergence speed is relatively
The updating law in closed-loop can be described as fast. As discussed in [23], the higher dither frequency will
lead to a faster convergence speed. However, when a large
ςn (k) = Hn · ςn (k − 1) + Jk (θ(k)) − Jk−1 (θ(k − 1)) frequency is selected, it will affect the computation of ėi as
ηn (k) = ςn (k) · sin(ωn k) it will have very large cost function. Although by tuning a
σn (k) = Ln · σjn (k − 1) + (1 − Ln ) · ηjn (k − 1) sufficiently small α2 relatively to α1 , it is possible to obtain
the fast convergence, the small vibration might still exist as
θ̂n (k) = θ̂n (k − 1) − γn · σn (k − 1)
  the cost function (17) weights less in terms of vibrations.
θn (k) = Proj θ̂n (k) + an · sin(ωn k), θ n , θ̄n , (19) As in the setting of the SSMLP, it is desired to have less
vibrations. In this work, the frequency of dither signal is
where the cost Jk (·) is defined in (17) and n = 1, 2, 3.
selected to be lower, leading to a slower convergence. Obvi-
It is noted that by using ‘‘sampled-data’’-like structure,
ously there is a design trade-off between the convergence
a discrete-time ES is used. The notion ‘‘k’’ is the number
speed and the performance. Different applications might have
of iterations. As three parameters are tuned on-line, three ES
different balances. ◦
loops are used. The dither signal an sin(ωn k) has different fre-
Remark 5: Instead of using the exact same PID parame-
quency ωn and amplitude an to ensure the enough persistent
ter in the classic decentralized control design, the proposed
excitation condition. Three high-pass filters and three low-
method use the exact same tuning parameters of ES for
pass filters are used with their corresponding tuning parame-
each suspension point with the possibility of having differ-
ters. The tuning parameters γn are positive small parameters
ent PID parameters. As long as a local optimal parameter
as well.
set exists for each suspension point, the ES algorithm can
Denoting the parameters θ 0 = [kp,0 , ki,0 , kd,0 ]T ∈ S,
find it. ◦
which is obtained from the trial-and-error tuning such that
Remark 6: Although Theorem 1 shows the convergence
one suspension point can levitate to a neighbourhood of the
when the iteration number goes to infinity, usually, the updat-
desired set point, though oscillations might happen.
ing law will stop if |Jk (θ )−Jk−1 (θ)| ≤  for some pre-defined
The following theorem show that the proposed on-line PID
threshold  > 0. ◦
tuning can reach to a neighbourhood of the optimal solution.
Remark 7: It is worthwhile to highlight that though this
Theorem 1: Assume that Assumption 1 holds the cost
decentralized auto-tuning or on-line PID control algorithm is
function (13) coming from the system (14). Let ν, α1 , and α2
designed for the SSMLP, it can be applied to the commercial
be positive constants. Let ωn is selected such that ωn 6=
Maglev transport systems when the different conditions of
ωj , ∀n, j = 1, 2, 3, n 6 = j. Then here exists a∗ > 0 such that
tracks, change of loads and self-excited vibration happen.
for every an ∈ (0, a∗ ), n = 1, 2, 3, there exists γn∗ = γn∗ (an )
Usually, under these changes, PID parameters might have
such that for any γn ∈ (0, γn∗ ), such that the trajectories of the
to be re-tuned or an extra block (model-based or model-
closed-loop system (19) satisfy:
free) is needed to improve the performance. With the pro-
lim sup θ(k) − θ ∗  ≤ ν.
 
(20) posed algorithm, as the PID parameter are tuned on-line,
k→∞ as long as the changes are slowly time-varying, it is always
Sketch of Proof: The convergence analysis of the proposed possible to find an optimal parameter set under the current
ES consists of the following steps. situation. ◦

3064 VOLUME 6, 2018

Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

IV. EXPERIMENTS RESULTS

A. SINGLE POINT EXPERIMENTS
Due to space limitation, this paper will not present dif-
ferent experimental results for different sets of parameters.
It is noted that the different combinations of tuning param-
eters have been tested. What presented here is the best
observed performance. However, we will share our tuning
experiences.
 From a large amount of experiments, we have S :=
(kp , ki , kd ) |[500, 2500] × [0.0001, 0.5] × [10, 25] .
Noting that e(t) is relatively small as the working range for
the gap is within [0, 11mm] and ė(t) can be very large if vibra-
tions happen, α1 is set to be relatively larger than α2 . A simple
normalization technique is used in the experiments to make
e(t) and ė(t) are compatible by selecting α1 and α2 . That is,
α1 e(t) and α2 ė(t) are in the similar ranges. In the experiments,
α1 = 10000 and α2 = 100.
In the experiments, the sampling period T is selected to be
1 second. Obviously, when a larger sampling period is used,
the cost function Jk (·) will be in a larger range, the step size
of the ES algorithms has to tuned relatively large, leading to
a relatively fast convergence speed and less computational
cost, but with a larger ultimate bound. On the other hand, if a
smaller T is used, it requires more computations. It also needs
smaller tuning parameters with a relatively slow convergence
speed and a smaller ultimate bound. In this experiments,
the sampling period T is selected to be 1 second to achieve
the desired performance. FIGURE 8. Experiment results of suspension point 2 with r = 4mm.
With the consideration of eliminating the vibrations,
the dither frequency is selected to be small. In the exper-
iments, ω1 = π8 , ω2 = 12 π
, and ω3 = 16 π
. The dither
amplitudes are set to be relatively small. As kp ∈ [500, 2500], B. HALF BOGIE EXPERIMENT
which is much larger than the ranges of other two parameters, Instead of testing only single suspension point, half bogie is
the step size of for tuning kp is the largest among three tested (suspension point 1 and suspension point 2). With a
parameters. As ki ∈ [0.0001, 0.5], the step size of the gradient decentralized PID controller is implemented with θ 0 obtained
should be the smallest. from the experiments for single point, it is observed from
Thus in our experiments γ1 = 2, a1 = 5 (kp ) is selected. experiments that severe vibrations appear as shown in Fig. 9.
We also choose γ2 = 0.0001, a2 = 0.0001 (ki ) and In this experiments, a special tuning procedure is intro-
γ3 = 0.01, a3 = 0.02 (kd ). The parameters for low pass filters duced to reduce the possibility that the vibration could hap-
and high pass filters are set as H1 = 0.9245, L1 = 0.1404, pen. It is noted that the focus of this work is to optimize the
H2 = 0.9615, L2 = 0.3747, H3 = 0.949 and L3 = 0.2701. steady-state behavior, thus the transient response is ignored
Other parameters have been tested in the experiments with in this procedure. Firstly levitate suspension point 1 (I ).
slightly different performances. After suspension point 1 is stably levitated, suspension point
In order to compare the performance of decentralized PID 2 begins to levitate (II ). When suspension point 2 reaches its
controller, the first experiment focused on levitating suspen- steady-state, vibration happens and causes suspension point 1
sion point 2 only. The decentralized PID controller with the vibration. Then ES is introduced to tune PID parameters on-
ES is used to levitate suspension point 2 to r = 4mm, line to ensure the stability of the overall system (III ) with an
which had a lot of vibrations when used the PID param- optimal steady-state behaviour. The tuning of ES parameters
eters from the tuning of suspension point 1, as shown is similar to what discussed above. It is noticed that the PID
in Fig. 6b. Experiment result for suspension point 2 is shown parameters of suspension point 1 and suspension point 2 are
in Fig. 8. different after ES tuning. When the set-point are r = 4mm
It can be seen that when ES is used to auto-tune PID param- for suspension point 1 and suspension point 2, as shown
eters, the cost function decreases along iterations. Compared in Fig. 9, the PID parameters are kp = 1317, ki = 0.0964,
with Fig. 6b, a good steady-state performance is achieved kd = 12.44 for suspension point 1 and kp = 1370,
with vibrations greatly reduced. ki = 0.0958, kd = 12.73 for suspension point 2.

VOLUME 6, 2018 3065

 Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

[6] R.-J. Wai, J.-D. Lee, and K.-L. Chuang, ‘‘Real-time PID control strategy
for Maglev transportation system via particle swarm optimization,’’ IEEE
Trans. Ind. Electron., vol. 58, no. 2, pp. 629–646, Feb. 2011.
[7] R. Song and Z. Chen, ‘‘Optimization and design of Maglev system PID
controller based on particle swarm optimization algorithm,’’ J. Southwest
Jiaotong Univ., vol. 50, no. 1, pp. 36–43, 2015.
[8] Y. Tan, W. H. Moase, C. Manzie, D. Nešić, and I. M. Y. Mareels,
‘‘Extremum seeking from 1922 to 2010,’’ in Proc. 29th Chin. Control Conf.
(CCC), 2010, pp. 14–26.
[9] M. Krstić and H.-H. Wang, ‘‘Stability of extremum seeking feedback
for general nonlinear dynamic systems,’’ Automatica, vol. 36, no. 4,
pp. 595–601, 2000.
[10] Y. Tan, D. Nešić, and I. Mareels, ‘‘On non-local stability properties of
extremum seeking control,’’ Automatica, vol. 42, no. 6, pp. 889–903, 2006.
[11] N. Killingsworth and M. Krstić, ‘‘Auto-tuning of PID controllers via
extremum seeking,’’ in Proc. Amer. Control Conf., 2005, pp. 2251–2256.
[12] N. J. Killingsworth and M. Krstić, ‘‘PID tuning using extremum seeking:
Online, model-free performance optimization,’’ IEEE Control Syst. Mag.,
vol. 26, no. 1, pp. 70–79, Feb. 2006.
[13] X. Zhang, Y. Tan, Y. G. Li, and I. M. Mareels, ‘‘On the modelling com-
plexity and utility of an EMS system,’’ in Proc. 49th IEEE Conf. Decision
Control (CDC), Dec. 2010, pp. 3523–3528.
[14] X.-H. Shi and L.-H. She, ‘‘Dynamic uncoupling capability simulation
of multi-controllers Maglev system,’’ J. Syst. Simul., vol. 18, no. 7,
pp. 1954–1670, 2006.
[15] A. Sakalli, T. Kumbasar, E. Yesil, and H. Hagras, ‘‘Analysis of the per-
formances of type-1, self-tuning type-1 and interval type-2 fuzzy PID
controllers on the magnetic levitation system,’’ in Proc. IEEE Int. Conf.
Fuzzy Syst. (FUZZ), Jul. 2014, pp. 1859–1866.
[16] J. Li, J. Li, D. Zhou, P. Cui, L. Wang, and P. Yu, ‘‘The active control of
Maglev stationary self-excited vibration with a virtual energy harvester,’’
IEEE Trans. Ind. Electron., vol. 62, no. 5, pp. 2942–2951, May 2015.
[17] D. Zhou, C. H. Hansen, J. Li, and W. Chang, ‘‘Review of coupled vibration
problems in EMS Maglev vehicles,’’ Int. J. Acoust. Vibrat., vol. 15, no. 1,
p. 10, 2010.
FIGURE 9. Experiment results of half bogie: setting the expectation gap [18] N. I. Marcos, M. Guay, and D. Dochain, ‘‘Output feedback adaptive
to 4mm and 4mm. extremum seeking control of a continuous stirred tank bioreactor with
Monod’s kinetics,’’ J. Process Control, vol. 14, no. 7, pp. 807–818, 2004.
[19] D. Popovic, M. Jankovic, S. Magner, and A. R. Teel, ‘‘Extremum seeking
It is observed that the performance of the proposed algo- methods for optimization of variable cam timing engine operation,’’ IEEE
Trans. Control Syst. Technol., vol. 14, no. 3, pp. 398–407, May 2006.
rithm is effective as it can greatly reduce the vibrations and [20] J.-Y. Choi, M. Krstic, K. B. Ariyur, and J. S. Lee, ‘‘Extremum seeking
achieve better steady-sate response. control for discrete-time systems,’’ IEEE Trans. Autom. Control, vol. 47,
no. 2, pp. 318–323, Feb. 2002.
[21] S. Z. Khong, Y. Tan, C. Manzie, and D. Nešić, ‘‘Multi-agent source seeking
V. CONCLUSION via discrete-time extremum seeking control,’’ Automatica, vol. 50, no. 9,
In the SSMLP, the decentralized PID control structure is pp. 2312–2320, 2014.
used to levitate the gap to the desired set-point. Due to the [22] H. K. Khalil, Nonlinear Systems, vol. 2. Englewood Cliffs, NJ, USA:
Prentice-Hall, 1996, pp. 1–5.
existence of mechanic couplings among suspension points, [23] Y. Tan, D. Nešić, and I. Mareels, ‘‘On the choice of dither in extremum
tuning PID parameters becomes very challenging. The trial- seeking systems: A case study,’’ Automatica, vol. 44, no. 5, pp. 1446–1450,
and-error method cannot always work. This paper uses ES 2008.

algorithm to find optimal PID parameters on-line to achieve

good steady-state performance and avoid the vibrations. The
ES is implemented in a decentralized way and the initial
value of PID parameters come from trial-and-error methods.
Experimental results validate the effectiveness of the pro-
posed method.

REFERENCES
[1] L. Yan, ‘‘Development and application of the Maglev transportation sys- QIANG CHEN received the B.S. and M.S.
tem,’’ IEEE Trans. Appl. Supercond., vol. 18, no. 2, pp. 92–99, Jun. 2008. degrees in automation from the National Uni-
[2] H.-W. Lee, K.-C. Kim, and J. Lee, ‘‘Review of Maglev train technologies,’’ versity of Defense Technology, Changsha, China,
IEEE Trans. Magn., vol. 42, no. 7, pp. 1917–1925, Jul. 2006. in 2012 and 2014, respectively, where he is cur-
[3] Y. Luguang, ‘‘Progress of the Maglev transportation in China,’’ IEEE
rently working toward the Ph.D. degree at the
Trans. Appl. Supercond., vol. 16, no. 2, pp. 1138–1141, Jun. 2006.
Maglev Engineering Center.
[4] G. He, J. Li, and P. Cui, ‘‘Decoupling control design for the module sus-
pension control system in Maglev train,’’ Math. Problems Eng., vol. 2015, His research interests include magnetic levi-
no. 1, pp. 1–13, 2015. tation control technologies and linear propulsion
[5] H. Liu, X. Zhang, and W. Chang, ‘‘PID control to Maglev train system,’’ technologies.
in Proc. Int. Conf. Ind. Inf. Syst., 2009, pp. 341–343.

3066 VOLUME 6, 2018

Q. Chen et al.: Decentralized PID Control Design for Maglev Systems

YING TAN received the Ph.D. degree in electri- IVEN MAREELS received the M.S. degree in
cal engineering from the National University of electromechanical engineering from Ghent Uni-
Singapore in 2002. versity, Belgium, in 1982 and the Ph.D. degree in
In 2002, she joined the Department of systems engineering from The Australian National
Chemical Engineering, McMaster University, as a University, Canberra, Australia, in 1987.
Postdoctoral Fellow. Since 2004, she has been Since 1996, he has been a Professor of Electrical
with the Department of Electrical and Elec- and Electronic Engineering at the Department of
tronic Engineering, The University of Melbourne, Electrical and Electronic Engineering, The Univer-
where she is currently an Associate Professor sity of Melbourne. His current research interests
and Reader. Her current research interests are are control and instrumentation for irrigation net-
iterative learning control, extremum seeking control, nonlinear systems works, control theory and automation, mathematical modelling of engineer-
and applications in rehabilitation robotic systems, and engine operation ing systems, signal processing and data modelling, and systems engineering.
optimizations.

JIE LI received the Ph.D. degree in automation

from the National University of Defense Technol-
ogy, Changsha, China, in 1999.
He held a post-doctoral position with The
Hong Kong University of Science and Technology,
Hong Kong. He is currently a Professor with
the Maglev Engineering Center, National Univer-
sity of Defense Technology. His research interests
include the magnetic levitation technologies and
robotics.

VOLUME 6, 2018 3067