You are on page 1of 11

1818 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO.

4, AUGUST 2019

Dynamics Modeling and Inversion-Based


Synchronized Model Predictive Control for a
Fabry–Perot Spectrometer
Zhi Li , Jinjun Shan , Senior Member, IEEE, and Ulrich Gabbert

Abstract—Three piezoelectric actuators (PEAs) are as- A critical requirement for FPS is to adjust the two optical
sembled in a Fabry–Perot spectrometer (FPS) to adjust the plates to move in parallel in order to obtain the best measurement
spacing between two optical plates. The measurement of condition, namely, the motion of the actuating PEAs should be
the FPS requires that the actuated optical plates move in
parallel with nanometer resolution. To meet this require- synchronized. However, the inherent hysteresis effect of PEAs
ment, the motion of the three PEAs should be controlled and the system dynamics makes it a very challenging task. To
synchronously. Thus, the challenging tasks in control of overcome these problems, the modeling and control approach
the Fabry–Perot (F–P) system lie in two aspects: control of the dynamic system driven by multiple PEAs should be
of each PEA precisely at the nanometer level and control
investigated.
of the three PEAs to achieve a synchronized motion. To this
end, a novel dynamic model is first proposed to describe In the literature, for modeling and control of the dynamic
the inherent hysteresis effect and the system dynamical system actuated by a single PEA, many research studies have
behaviors in the F–P system. Based on the proposed model, been reported, which mainly focus on the modeling of the
an inversion-based synchronized model predictive control hysteresis and the inverse compensation, the modeling of the
(SMPC) design is developed. The experimental results vali-
system dynamics, and the controller design. In terms of hystere-
date the proposed control scheme.
sis modeling and the inverse compensation, a Prandtl–Ishlinskii
Index Terms—Fabry–Perot spectrometer (FPS), hystere- (PI) model was utilized to describe the hysteresis effect of a PEA
sis, model predictive control (MPC), piezoelectric actuators in [7], and an analytical inverse compensator was constructed
(PEAs), system dynamics.
to mitigate the hysteresis effect. In [8], the hysteresis effect was
I. INTRODUCTION represented by a Bouc–Wen model and an inverse multiplicative
structure was developed to compensate for the hysteresis effect.
FABRY–PEROT spectrometer (FPS), which is an optical
A instrument used for measuring the wavelengths of light,
can provide multispectral mappings for atmospheric studies.
Similar research results for different hysteresis models and
inverse compensation methods, such as the Preisach model, least
squares support vector machine-based hysteresis model, can be
The key technique for FPS is adjusting the spacing between found in [9]–[13]. System dynamics modeling is also a very
two parallel highly reflecting optical plates inside FPS at the important factor for the applications of PEA. A lumped system
nanometer scale. To satisfy the requirement, multiple piezoelec- model [14]–[16] is commonly used to describe the systems
tric actuators (PEAs), which can convert electrical signals into dynamics in the piezo-actuated systems. One of the advantages
mechanical motions are employed. In contrast to the traditional of the lumped system model is that it can be easily integrated
electric motors, the PEAs have advantages, such as high reso- with the hysteresis model to describe both the hysteresis effects
lution, large bandwidth, and good temperature stability [1]–[6]. and the system dynamics of a piezo-actuated system.
Manuscript received July 30, 2018; revised January 11, 2019; ac- For control of the system considering both the hysteresis
cepted July 7, 2019. Date of publication July 10, 2019; date of current effect and the system dynamics, the available control strategies
version August 14, 2019. This work was supported in part by the National can be classified into two categories: without and with an
Natural Science Foundation of China under Grant 11472090, and in part
by the Alexander von Humboldt Foundation. Recommended by Technical inverse compensator. In the first category, the hysteresis effect
Editor C. Manzie. (Corresponding author: Jinjun Shan.) is considered as a system disturbance or uncertainty, and then
Z. Li is with the State Key Laboratory of Synthetical Automation for the control strategies, such as the sliding mode control [16], the
Process Industries, Northeastern University, Shenyang 110819, China
(e-mail:, gavinlizhi@gmail.com). disturbance observer-based control design [17], and the adaptive
J. Shan is with the Department of Earth and Space Science and neural network dynamic surface control [18], are applied to
Engineering, York University, Toronto, ON M3J 1P3, Canada (e-mail:, deal with both the hysteresis effects and the system dynamics
jjshan@yorku.ca).
U. Gabbert is with the Institute of Mechanics, Otto-von-Guericke to guarantee a precise positioning performance. In the other
Universität Magdeburg, 39106 Magdeburg, Germany (e-mail:, ulrich. category, the hysteresis effect is properly described first and its
gabbert@ovgu.de). inverse compensator is constructed to mitigate the hysteresis
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. effect, and then the available linear control methods can be
Digital Object Identifier 10.1109/TMECH.2019.2927752 directly applied to deal with the system dynamics. For example,

1083-4435 © 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: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DYNAMICS MODELING AND INVERSION-BASED SMPC FOR A FABRY–PEROT SPECTROMETER 1819

in [3], the Preisach model is utilized to describe the hysteresis


effect of PEA, and the inverse of the Preisach model is developed
to mitigate the hysteresis effects. Then, an adaptive control
method is applied to achieve the desired tracking performance.
Similar strategies can also be found in [19] and [20].
It can be seen from the literature that the modeling and control
of the system actuated by the single PEA is well developed. For
the system with multiple PEAs, most of the current research
works focus on the multiaxis piezo-actuated positioning system,
such as a 2-DOF flexure-based mechanism [21] and a 3-DOF
piezotube scanner [22]. However, the mechanism of the F–P
system is quite different from the multiaxis piezo-actuated posi-
Fig. 1. Image of the developed FPS.
tioning system in this application. In the F–P system, three PEAs
are placed parallell along a single axis. Therefore, the available
modeling and control strategies for the multiaxis systems cannot
objective of the F–P system. Therefore, in this application, an
be applied directly, especially for the modeling of the system
inversion-based synchronized model predictive control (SMPC)
dynamics. In our previous works [5] and [6], the modeling and
approach, which is a combination of the MPC approach and the
control of coupled hysteresis effects in the F–P system have
cross-coupling error concept, will be developed to deal with the
been well addressed. However, the system dynamics of the F–P
nonlinear hysteresis effects and system dynamics in the F–P
system, which plays an important role when the F–P system is
system and to achieve a synchronized motion of the driven
operated at higher input frequencies, is ignored in above works,
PEAs.
and the experimental tests in [5] are only conducted at low input
frequencies. Thus, a novel dynamic model will be proposed to
address the system dynamics of the F–P system in this paper. The II. EXPERIMENTAL PLATFORM
proposed dynamic model is capable of describing the coupling Fig. 1 shows the developed FPS in the Spacecraft Dynamics
effects and the system dynamics with the mechanical loads Control and Navigation Laboratory at York University. The
imposed on the three PEAs. device was enclosed into a sealed tube and was launched on
As indicated before, one of the control purposes of the F–P a stratospheric balloon. So far, two successful fights: one in
system is the synchronization control of the three PEAs. The Kiruna, Sweden, and the other one in Alice Springs, Australia,
synchronization objective is to force the three PEAs to have a have been made for the research goal of measuring aerosol,
synchronized motion in order to achieve a parallel movement surface albedo, and pressure at O2 A-band.
of the driven optical plate, which is a critical requirement to The key component of the FPS is the tunable system (F–
guarantee accurate measurement of FPS. Toward this target, a P system), which is used to adjust the spacing between two
synchronization strategy should be considered in the controller parallel highly reflecting optical plates, shown in Fig. 2(a). It
design. A commonly used approach for synchronizing the should be noted that this device is only a prototype of the F–P
motion of multiple actuators is to synthesize the cross-coupling system, which is used for the purpose of validating the proposed
error [23], [24], which is a combination of a tracking error and controller. Thus, only partial components, e.g., the optical plates
a synchronization error, in the controller design. It is noted that are not installed, are shown in Fig. 2(a). Fig. 2(b) shows a side
the cost of the synchronization strategy would sacrifice the fast view of the F–P system. The F–P system is driven by three PEAs,
responses of certain controlled actuators to achieve a synchro- which attach to the front structure for holding the optical plate
nized motion with the slow ones. In terms of the controller of the F–P system, to adjust the spacing between the front plate
design, a model predictive control (MPC) approach is selected and the back plate. Three PEAs (P-887.51) with 15 μm nominal
in this application. MPC [25], [26] is an optimization-based travel range and three capacitive sensors (D-015.00) are utilized.
approach that computes the next control action by minimizing A voltage amplifier (E-503.00) and capacitive sensor monitor
the difference between the predicted output of a system and a (E-509.C3A) are employed to amplify the control signal and to
desired tracking signal. In [27], an internal model based repet- obtain the measurement from the three capacitive sensors. All
itive control law integrated with MPC is proposed to deal with the above equipments are from Physik Instrumente. An xPC
the periodic signal tracking problem on a piezoelectric platform. real-time target computer is used to control the amplifier and
In [20], an inversion-based MPC with an integral-of-error state read from the sensor monitor through a National Instruments
variable is proposed to improve the tracking performance of the PCI-6289 card. A host computer with the software MATLAB is
PEA. In [28] and [29], a MIMO MPC controller is designed set up to control the target computer.
to cope with a 2-DOF piezoelectric tube scanner. It is noted It is noted that the operating frequencies of the F–P system
that above MPC strategies can run at 5–20 kHz in real-time, depends on the design of the F–P system and its working
which make the MPC controller applicable in the tracking principle. For different types of F–P systems, such as the
control of fast piezoelectric systems. However, the available MEMS-based F–P system[30], pyroelectric detector-based F–P
MPC strategies cannot guarantee a synchronized performance system, the operating frequencies could be in the range from
of the controlled actuators, which cannot meet the control 0.01 Hz to 1 MHz. In this design, the F–P system in the two

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
1820 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 4, AUGUST 2019

positioning performance and even instability when used in


the closed-loop control system. To capture these dynamical
behaviors, a dynamic model should be proposed first.

III. MODELING AND IDENTIFICATION OF THE HYSTERETIC


BEHAVIORS AND THE SYSTEM DYNAMICS FOR
THE F–P SYSTEM

The F–P system is driven by three PEAs and each PEA


shows the hysteresis effect and dynamic behaviors. Besides,
the performance of the driven PEAs depends on each other and
they show coupling dynamics as seen in Fig. 3(b) and (c). To
describe this complex behavior, a model should consist of two
components: the hysteresis submodel and the system dynamics
submodel. Fig. 4 shows the diagram of the proposed model,
in which the three hysteresis models P1 − P3 are employed to
describe the hysteretic behavior in each PEA, and nine lumped
system models G11 − G33 are utilized to describe the system
dynamics, including the coupling dynamics of the F–P system.
To depict the hysteretic behavior for a PEA, many hysteresis
models are available in the literature, such as the PI model
[7], Preisach model [31], and Bouc–Wen model [8]. In this
Fig. 2. Experimental platform. (a) Prototype of the F–P system application, considering the model accuracy and simplicity, the
(b) Schematic plot of the F–P system. PI model is selected to represent the hysteretic behavior in PEAs.
The numerical expression of the PI model [7] is written as
flights were operated in a scan mode at 0.033 Hz. However, it 
n

consumes a long time for a full-range scan and causes a low P [u](t) = p0 u(t) + pi Fr i [u](t) (1)
i=1
scan efficiency. The faster scan, e.g., with a scan mode up to
100 Hz, is desired in practical applications for the developed where u(t) denotes the input, Fr i [u](t) = ω(t) is the play
FPS. operator defined as
To investigate the behaviors of the F–P system with higher
ω(0) = Fr [u](0) = fr (u(0), ω−1 ) (2)
input frequencies, the system responses of the F–P system with
different input frequencies were collected, which is shown in ω(t) = Fr [u](t) = fr (u(t), ω(ti )) (3)
Fig. 3. Note that the input signals are u1 = 0.6 sin(2fr πt), u2 =
where ω−1 denotes the initial value and ti < t ≤ ti+1 , 0 ≤ i ≤
u3 = 0, where fr = 1, 10, 50, 100, u1 , u2 , and u3 denote the
N − 1, with
input to PEA 1, PEA 2, and PEA 3, respectively. Different  
input frequencies can excite the dynamic characteristics of the fr (u, ω) = max u − r, min(u + r, ω) (4)
system. Fig. 3(a) illustrates the hysteresis effects and dynamics
between the input u1 and output y1 , and it can be seen that with pi (i = 0, 1, . . . , n) is the weight of the play operator, n defines
different input frequencies the shape of the hysteresis loop goes the number of the play operators.
differently. To describe the dynamics of the F–P system, especially with
Note that the performance of the three PEAs is not in- coupling dynamics, a transfer function matrix based on the
dependent to each other and dynamic coupling effects exist diagram in Fig. 4 is expressed as
among the three PEAs. For example, Fig. 3(b) and (c) shows ⎡ ⎤ ⎡ ⎤⎡ ⎤
y1 G11 G12 G13 v1
the coupling dynamics behaviors under the input condition ⎢ ⎥ ⎢ ⎥⎢ ⎥
u1 = 0.6 sin(2fr πt), and u2 = u3 = 0. If the coupling effects ⎢ y2 ⎥ = ⎢ G21 G22 G23 ⎥ ⎢ v2 ⎥ (5)
⎣ ⎦ ⎣ ⎦⎣ ⎦
do not exist, when the input u2 and u3 are equal to zero, the
y3 G31 G32 G33 v3
output y2 and y3 should also be zero. However, from the input
and output (I/O) responses in Fig. 3(b) and (c), the outputs y2 and where Gj k , j = 1, 2, 3, k = 1, 2, 3 has the following expression
y3 show dynamic responses according to the input excitation.
KH
In addition, it is noted that the motion of the PEA 3 is different Gj k = (6)
from the other two PEAs. The reason is that the circular flexure s2 + aH s + bH
was much stiffer in the mounting area of PEA 3 than the other where KH , aH , and bH are the coefficients to be identified.
two, which leading to the complex dynamical coupling effects The next step is to identify the parameters in the proposed
together with the hysteresis nonlinearities. model. From the structure in Fig. 4, the proposed model shows a
The hysteresis and system dynamics, including the coupling cascading structure, e.g., the system dynamics preceded by the
dynamics are unwanted behaviors, they can cause inaccurate three PI models. Therefore, the identification procedure has two

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DYNAMICS MODELING AND INVERSION-BASED SMPC FOR A FABRY–PEROT SPECTROMETER 1821

Fig. 3. Hysteresis effects and system dynamics with input u 1 = 0.6 sin(2fr πt), u 2 = u 3 = 0. (a) u 1 − y 1 (b) u 1 − y 2 (c) u 1 − y 3 .

TABLE I
THRESHOLDS AND WEIGHTS OF THE THREE PI MODELS

C1 ∈ RN ×(n +1) is defined as the output matrix of the given


play operators Fr i [u](k), and C1 = [Fr 0 , . . . , Fr n ], Fr i =
[Fr i [1], . . . , Fr i [k], . . . , Fr i [N ]]T , N is defined as N = tf /Ts ,
tf denotes the total length of the given input, Ts denotes the sam-
Fig. 4. Diagram of the dynamic model. pling time. In this application, tf = 10 s, and Ts = 0.00005 s.
d1 = [y1 [1], . . . , y1 [k], . . . , y1 [N ]]T with d1 ∈ RN ×1 is the
output values of the PEA 1. The weights W1 can be calculated
steps: identification of the three PI models and identification of as
the nine transfer functions in the transfer function matrix.
min{[C1 W1 − d1 ]T [C1 W1 − d1 ]} (9)
A. Step I: Identification of the Three PI Models with the constraints
P1 , P2 , and P3
First, a sinusoidal input pi ≥ 0, i ∈ {0, 1, . . . , n}. (10)

u1 (t) = Ab (t) sin(2πt) (7) The MATLAB command lsqnonneg is employed to determine
the weights in W1 . The procedure for identifying the parameters
with in P1 applies to P2 and P3 . Table I lists the identified results,
where ri are chosen as ri = 0.1i (i = 0, 1, . . . , n). Note that the
2 t ∈ [0, 1)
Ab (t) = (8) zero weights are not provided in Table I.
2 − 0.2τ t ∈ [τ, τ + 1), (τ = 1, 2...9)
B. Step II: Identification of the Nine Transfer Functions
is applied to the PEA 1 only while keeping u2 = u3 = 0.
Remark: It is noted that the PI model is identified at a fixed The frequency responses of the F–P system can be obtained
frequency. The selected frequency cannot be either too high to by applying a chirp signal to each PEA. For example, a chirp
excite the dynamic behavior of the system or too low to cause signal u1 is applied to PEA 1 only (u2 = u3 = 0), and then the
the creep effect of the PEA. Therefore, a 1 Hz identification corresponding y1 , y2 , and y3 can be obtained. Then, MATLAB
signal is selected in this paper. However, there is no general rule command invfreqs with the derived data set u1 − y1 , u1 − y2 ,
on the selection of the frequency of the identification signal. and u1 − y3 can be utilized to identify the parameters in transfer
Some trial-and-error work need to be done. functions G11 , G21 , and G31 .
Then, the I/O data are collected and will be used to iden- By repeating this procedure, the parameters for the rest trans-
tify the weights pi in P1 [u1 ](t). Define the identified weights fer functions can be identified. Fig. 5 illustrates the comparison
in a vector as W1 = [p0 , . . . , pn ]T , where W1 ∈ R(n +1)×1 . of the magnitudes between the experimental data and the model.

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
1822 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 4, AUGUST 2019

Fig. 5. Frequency response of the F–P system versus identified response.

TABLE II
IDENTIFIED TRANSFER FUNCTIONS

Since the proposed model shows a cascading structure and its It can be observed from Fig. 7 that the submodels P3 + G13
identification is separated, one of the components in the structure and P3 + G23 show relatively large modeling error than other
ought to be normalized. The transfer functions are normalized coupling dynamics submodels: P1 + G21 and P1 + G31 . The
in this application. The normalized transfer functions Gj k are reason might be due to the un-modeled system dynamics in the
given in Table II. F–P system.
In order to verify the developed model, sinusoidal signals To further validate the proposed model, two other input
u(t) = 0.6 sin(2fr πt) with fr = 1, 10, 50, 100, are applied. As signals, u(t) = sin(40πt) + 0.2 sin(20πt) and a triangular sig-
an illustration, Fig. 6 shows the comparison between the exper- nal with the frequency of 50 Hz under the input condition
imental data and the model at 100 Hz for the input condition u1 = u2 = u3 = u, are applied. Fig. 8 shows the comparison
u1 (t) = u(t), u2 (t) = u3 (t) = 0. To give a quantified analysis, between the experimental data and the model for PEA 3, which
a normalized root mean square error (NRMSE) En is defined as also shows a fair prediction behavior.

1 1 N
En = e2 (k) (11)
ȳ N k =1 IV. INVERSION-BASED SYNCHRONIZED MODEL
where N is the number of the error data, e(k) = ye (k) − ym (k) PREDICTIVE CONTROL
is the modeling error, ye and ym denote the output of the As indicated in the introduction, two control tasks: control
experimental data and the model, respectively. ȳ = ymax − ymin , of each PEA precisely at the nanometer level and control of
where ymax and ymin denote the maximum and minimum value the three PEAs to achieve a synchronized motion. In order to
of the PEAs under the given input frequency. The calculated meet both requirements, an inversion-based SMPC is proposed.
NRMSEs under different input frequencies are shown in Fig. 7. To this end, an inverse compensator is built to compensate

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DYNAMICS MODELING AND INVERSION-BASED SMPC FOR A FABRY–PEROT SPECTROMETER 1823

Fig. 6. Model validation with the desired input u(t) = 0.6 sin(200πt) for PEA 1. (a) u 1 − y 1 (b) u 1 − y 2 (c) u 1 − y 3 .

Fig. 7. NRMSE with different input frequencies. (a) u 1 = u ∗ , u 2 = u 3 = 0 (b) u 2 = u ∗ , u 1 = u 3 = 0 (c) u 3 = u ∗ , u 1 = u 2 = 0.

For MPC design, a discrete formulation of the model of the


F–P system should be derived first. Based on the developed
dynamic model in (1) and (5), the discrete model is written as

xm (k + 1) = Am xm (k) + Bm P [u](k)
y(k) = Cm xm (k) (12)

where xm (k) ∈ Rn 1 ×1 , y(k) ∈ Rq ×1 , u(k) ∈ Rd×1 , Am ∈


Rn 1 ×n 1 , Bm ∈ Rn 1 ×d , Cm ∈ Rq ×n 1 , n1 , d, and q denote the
number of the states, inputs and outputs, respectively. The MAT-
LAB commands ssdata and c2dm are utilized to find the matrices
Am , Bm , and Cm , with the sampling time Ts = 0.00005 s.
Before conducting the SMPC design, an inverse feedforward
compensator is developed to mitigate the hysteresis effects de-
scribed by the PI models in (1). The numerical implementation
of the inverse PI model [7] is expressed as


n
P −1 [v](k) = p̄0 v(k) + p̄i Fr̄ i [v](k) (13)
i=1

where v(k) denotes the output of the PI model, the thresholds


r̄i , and the weights p̄i can be calculated as
Fig. 8. Model validation with a complex sinusoidal signal and a triangu-
lar signal for PEA 3. (a) Complex sinusoidal signal as input (b) Triangular
signal as input. 
i−1
r̄i = p0 ri + pj (ri − rj ) (14)
j =1
for the hysteresis nonlinearity, and the MPC approach with
a synchronization strategy by considering the cross-coupling p̄0 = 1/p0 (15)
error concept in the MPC design is developed. Note that the pi
MPC design employed in this paper follows the main steps p̄i = −  i  i−1 . (16)
p0 + j =1 pj p0 + j =1 pj
reported in [25] and [26].

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
1824 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 4, AUGUST 2019

By substituting (13) into (12), one has samples [26]. Nc is a tuning parameter, which determines the
number of future control inputs involved in the optimization.
xm (k + 1) = Am xm (k) + Bm v(k)
J = E T QE + ΔV T RΔV (21)
y(k) = Cm xm (k). (17)
where E = Y − Rd with Rd = R̄d r(k) = [Iq ×q , Iq ×q ,
Since (17) is a standard linear system, the MPC control tech-
. . . , Iq ×q ]T r(k) and r(k) ∈ Rq ×1 denote the reference signal
nique can be directly applied. In order to facilitate the controller
and is assumed to have constant entries during one prediction
design, an augmented system formulation of (17) with the aug-
horizon, Q ∈ Rq N p ×q N p and R ∈ RdN c ×dN c are diagonal
mented system matrices A ∈ R(n 1 +q )×(n 1 +q ) , B ∈ R(n 1 +q )×d ,
weighting matrices.
and C ∈ Rq ×(n 1 +q ) can be written as [25]
It should be noted that in [26] the cost function (21) does
x(k + 1) = Ax(k) + BΔv(k) not involve the synchronization strategy. By only including
the tracking error in the cost function cannot guarantee a
y(k) = Cx(k) (18) synchronized motion. In order to involve the synchronization,
where x(k + 1) ∈ R(n 1 +q )×1 and a synchronization error should be introduced into the cost
⎡ ⎤ function. The synchronization error Ξs is defined as [32]
 
Δxm (k + 1) Am 0q ×n 1
x(k + 1) = , A=⎣ ⎦ Ξs = Th Etr (22)
y(k + 1) Cm Am Iq ×q
where Ξs = [ξ1 , ξ2 , ..., ξn ]T ∈ Rn ×1 , Etr = [e1 , e2 , ..., en ]T ∈
  Rn ×1 is the tracking error and Th ∈ Rn ×n denotes the syn-
Bm  
B= , C = 0q ×n 1 Iq ×q chronization matrix. Th can be chosen in different forms to
Cm Bm characterize the synchronization error. For example,
⎡ ⎤
with 2 −1 0 · · · −1
⎢ ⎥
Δv(k) = v(k) − v(k − 1) ⎢−1 2 −1 · · · 0⎥
⎢ ⎥
Th = ⎢⎢ ..
⎥.
Δxm (k + 1) = xm (k + 1) − xm (k). (19) ⎢ . . .. . .. . .. .. ⎥
.

⎣ ⎦
The predicted Np output sequence can be expressed as
−1 0 · · · −1 2
Y = F x(k) + ΦΔV (20)
Remark: The selection of the synchronization matrix is not
where Y ∈ Rq N p ×1 , F ∈ Rq N p ×(n 1 +q ) , ΔV ∈ RdN c ×1 , and unique. The discussion on the sensitivity of system performance
Φ ∈ Rq N p ×dN c to the form of the matrix selected can be found in [33].
⎡ ⎤ ⎡ ⎤ Thus, the synchronization error is included in the cost func-
y(k + 1|k) CA tion in (21) as
⎢ ⎥ ⎢ ⎥
⎢ y(k + 2|k) ⎥ ⎢ CA2 ⎥
⎢ ⎥ ⎢ ⎥ J = E T QE + ΔV T RΔV + ΞT Qs Ξ (23)
⎢ ⎥ ⎢ ⎥
Y =⎢ .. ⎥ , F = ⎢ .. ⎥ ,
⎢ . ⎥ ⎢ . ⎥ where Ξ ∈ Rq N p ×1 and Ξ = [ΞTs ΞTs ... ΞTs ]T , and Qs ∈
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ Rq N p ×q N p is a diagonal weighting matrix. By substituting (22)
y(k + Np |k) CA Np
into (23), one has
⎡ ⎤
Δv(k) J = E T QE + ΔV T RΔV + E T T T Qs T E
⎢ ⎥
⎢ Δv(k + 1) ⎥ = E T Q∗ E + ΔV T RΔV
⎢ ⎥ (24)
ΔV = ⎢
⎢ ..


⎢ . ⎥ where Q∗ = Q + T T Qs T , and
⎣ ⎦ ⎡ ⎤
Δv(k + Nc − 1) Th
⎡ ⎤ ⎢ ⎥
⎢ ⎥
CB 0q ×d ··· 0q ×d T = ⎢ ... ⎥.
⎢ ⎥ ⎣ ⎦
⎢ CAB CB ··· 0q ×d ⎥
⎢ ⎥ Th
⎢ ⎥
⎢ 2
··· ⎥
Φ = ⎢ CA B CAB 0q ×d ⎥ By minimizing the cost function in (24), the optimal control
⎢ ⎥
⎢ .. .. .. .. ⎥ sequence is derived as
⎢ . . . . ⎥
⎣ ⎦
ΔV = (ΦT Q∗ Φ + R)−1 ΦT Q∗ (R̄d r(k) − F x(k)). (25)
CAN p −1 B CAN p −2 B · · · CAN p −N c B
In order to consider the input constraints problem in the
Np and Nc denote the prediction and control horizons, respec- controller design, the constraints are given by
tively. It is noted that Nc ≤ Np is selected and it is assumed that
the incremental control has reached the steady state after Nc C1 vm in ≤ C1 v(k − 1) + C2 ΔV ≤ C1 vm ax (26)

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DYNAMICS MODELING AND INVERSION-BASED SMPC FOR A FABRY–PEROT SPECTROMETER 1825

where vm in ∈ Rd×1 and vm ax ∈ Rd×1 are the input constraints, TABLE III
CALCULATION RESULTS FOR THE INVERSE PI MODELS
and C1 ∈ RdN m ×d and C2 ∈ RdN m ×dN c are defined as
⎡ ⎤ ⎡ ⎤
I I 0 ··· 0
⎢ ⎥ ⎢ ⎥
⎢I ⎥ ⎢I I · · · 0⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
C1 = ⎢ . ⎥ , C2 = ⎢
⎢ ..
⎥.

⎢ .. ⎥ ⎢
.. . .
⎢ ⎥ ⎣. . . . .⎥
.

⎣ ⎦
I I I ··· I

Nm (≤ Nc ) denotes the number of future sampling instants


imposed within the control horizon Nc . Nm is a tuning pa-
rameter, which is selected depending on the performance of the
computational capacity of the microcontroller. If Nm = Nc , it performance of the three PEAs is examined with the input
means the input constraints are imposed on all future control signals at different frequencies. A comparison between the MPC
commands, which may cause a computational burden. It should and the inversion-based SMPC will be made to further verify
be noted that another two typical constraints (control incremen- the proposed controller.
tal variation and output constraints) can also be applied, see
examples reported in [25].
A. Controller Implementation
The input constraints problem can be expressed in a compact
form as follows: To implement the SMPC design, the feedforward inverse
compensator in (13) should be constructed first. Based on the
M1 ΔV ≤ N1 (27) identified results shown in Table I and (14), the thresholds r̄i
with and weights p̄i of the inverse PI models are listed in Table III.
⎡ ⎤   The transfer functions in (5) are second-order systems but the
−C2 −C1 vm in + C1 v(k − 1) measurements are only available for the displacement, i.e., not
M1 = ⎣ ⎦ , N1 = .
C2 C1 vm ax − C1 v(k − 1) all the states in (17) are measurable. Thus, a Kalman observer is
employed to estimate these states with unwanted system noise
The above input constraint problem can be solved using the and disturbances. For the detailed implementation of the Kalman
Hildreth’s quadratic programming algorithm. This approach ap- observer, the interested readers may refer to [34]–[36].
plies Lagrange multipliers with λ ≥ 0, as the decision variables The control parameters for the SMPC are selected as q = 3,
and the constrained optimization problem is written as d = 3, Q = blkdiag(Q1 , . . . , Q1 ), where blkdiag is a function
  defined in MATLAB to construct the block diagonal ma-
1 T
min λ Hλ + λT f + constant (28) trix, Q1 = 1.5eye(3), where eye(·) denotes the identity ma-
λ≥0 2 trix. Qs = blkdiag(Qs1 , . . . , Qs1 ), with Qs1 = eye(3). R =
where H = M1 (ΦT Q∗ Φ + R)M1T and f = −ΦT Q∗ (R̄d r(k) blkdiag(R1 , . . . , R1 ), with R1 = 5eye(3). The input constraints
− F x(k)). Using the reported algorithm in [25], the Lagrange are set as vm in = [−10 − 10 − 10]T and vm ax = [10 10 10]T .
multiplier vector λ can be computed. Then, the controller
considering the input constraints can be expressed as B. Controller Validation

ΔV = (ΦT Q∗ Φ + R)−1 ΦT Q∗ (R̄d r(k) − F x(k)) To validate the proposed controller in terms of the hysteresis
compensation and the synchronization performance, different
− (ΦT Q∗ Φ + R)−1 M1T λ. (29) tracking signals are applied. Fig. 9(a) shows the output re-
sponses of the three PEAs and the desired tracking signal
The incremental control input for the next sampling time is
u = 0.5 sin(2πt), and Fig. 9(b) illustrates the I/O responses of
determined as
the three PEAs. As shown in Fig. 9(b), the I/O responses of the
Δv(k) = Ky r(k) − Kx x(k) − Kc (30) three PEAs have a very good linear relationship, which means
that the hysteresis effect has been mitigated. Then, a multistep
where Ky , Kx , and Kc are the first d rows of (ΦT Q∗ Φ +
signal is applied and Fig. 10 shows the results. It can be seen
R)−1 ΦT Q∗ R̄d , (ΦT Q∗ Φ + R)−1 ΦT Q∗ F , and (ΦT Q∗ Φ +
from Fig. 10 that the motion of the three PEAs is synchronized.
R)−1 M1T λ.
To further validate the proposed controller, the comparisons
between the MPC (without considering the inverse compen-
V. EXPERIMENTAL RESULTS sation and the synchronization strategy) and the inversion-
In this section, the developed controller will be implemented based SMPC are made. The desired input signals are u(t) =
on the experimental F–P system. To validate the proposed 0.07 sin(2fr πt) with fr = 1, 30, 50, 100. Fig. 11(a) and (b)
inversion-based SMPC, different input conditions will be tested. show the tracking performance of the three PEAs with the MPC
First, a low frequency input signal is applied to validate the and the inversion-based SMPC at 100 Hz. For the MPC as
elimination of the hysteresis effects. Then, the synchronization shown in Fig. 11(b), the outputs of the three PEAs are not

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
1826 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 4, AUGUST 2019

Fig. 9. Hysteresis compensation using the proposed control approach.


(a) Output response. (b) I/O response.

Fig. 11. Model validation with different input frequencies and the
RMSE. (a) Inversion-based SMPC at 100 Hz. (b) MPC at 100 Hz.
(c) RMSE of the synchronization error.

Fig. 10. Synchronized motion validation in multistep input condition.


synchronization error is given by
 3
1 M
well synchronized, which show different responses and cannot Er = (δi (k))2 (31)
M k =1
i=1
accurately track the desired trajectories. The main reasons for
the MPC to cause the nonsynchronized motion of the three where δi (k) = Thi Etr , Etr = [e1 (k) e2 (k) e3 (k)]T denotes the
PEAs are due to the hysteresis nonlinearity of the PEA and not tracking error for the three PEAs, Thi denotes the ith row in the
to consider the synchronization errors among the three PEAs. synchronization transform matrix Th , M is the total number
Both reasons cause the outputs of the three PEAs with the MPC of the error data. It can be seen from Fig. 11(c) that with
approach to show lag behind and nonsynchronized tracking higher input frequencies, the RMSE increases dramatically for
performance in Fig. 11(b). the MPC approach, while for the inversion-based SMPC the
To have a quantified analysis, the root mean square error RMSE remains at the acceptable level. The comparison results
(RMSE) of the synchronization error is calculated for the indicate that the inversion-based SMPC works very well for the
input signals at 1, 30, 50, and 100 Hz. The RMSE of the input conditions at higher input frequencies.

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
LI et al.: DYNAMICS MODELING AND INVERSION-BASED SMPC FOR A FABRY–PEROT SPECTROMETER 1827

VI. CONCLUSION [14] M. Goldfarb and N. Celanovic, “Modeling piezoelectric stack actuators
for control of micromanipulation,” IEEE Control Syst. Mag., vol. 17,
Control of FPS actuated by three PEA requires: first, indi- no. 3, pp. 69–79, Jun. 1997.
vidual control of PEA at nanometer level; second, guaranteed [15] H. Adriaens, W. L. De Koning, and R. Banning, “Modeling piezoelectric
actuators,” IEEE/ASME Trans. Mechatronics, vol. 5, no. 4, pp. 331–341,
synchronized motion of the three PEAs. Toward these control Dec. 2000.
requirements, a dynamic model is first developed to represent [16] G.-Y. Gu, L.-M. Zhu, C.-Y. Su, and H. Ding, “Motion control of
the inherent hysteresis effects and system dynamics of the F–P piezoelectric positioning stages: Modeling, controller design, and exper-
imental evaluation,” IEEE/ASME Trans. Mechatronics, vol. 18, no. 5,
system. Based on the proposed dynamic model, an inversion- pp. 1459–1471, Oct. 2013.
based SMPC is developed. The proposed model is validated [17] J. Yi, S. Chang, and Y. Shen, “Disturbance-observer-based hysteresis com-
experimentally using different input signals and shows a good pensation for piezoelectric actuators,” IEEE/ASME Trans. Mechatronics,
vol. 14, no. 4, pp. 456–464, Aug. 2009.
prediction of the hysteresis effects and the system dynamics. For [18] X. Zhang, C.-Y. Su, Y. Lin, L. Ma, and J. Wang, “Adaptive neural
the inversion-based SMPC controller, the experimental results network dynamic surface control for a class of time-delay nonlin-
show that the developed controller can compensate for the ear systems with hysteresis inputs and dynamic uncertainties,” IEEE
Trans. Neural Netw. Learn Syst., vol. 26, no. 11, pp. 2844–2860, Nov.
hysteretic behaviors and at the same time achieve a synchronized 2015.
performance of the three driven PEAs. To further validate [19] J. Zhou, C. Wen, and T. Li, “Adaptive output feedback control of uncertain
the synchronization performance, the comparisons between the nonlinear systems with hysteresis nonlinearity,” IEEE Trans. Autom.
Control, vol. 57, no. 10, pp. 2627–2633, Oct. 2012.
MPC and the inversion-based SMPC are made. In comparison to [20] Y. Cao, L. Cheng, X. Chen, and J. Peng, “An inversion-based model
the MPC approach, the inversion-based SMPC approach shows predictive control with an integral-of-error state variable for piezoelectric
excellent synchronization performance, especially at higher actuators,” IEEE/ASME Trans. Mechatronics, vol. 18, no. 3, pp. 895–904,
Jun. 2013.
input frequencies. [21] U. Bhagat, B. Shirinzadeh, L. Clark, Y. Qin, Y. Tian, and D. Zhang,
“Experimental system identification, feed-forward control, and hysteresis
compensation of a 2-dof mechanism,” Int. J. Intell. Mechatronics Robot.,
REFERENCES vol. 3, no. 3, pp. 1–21, Sep. 2013.
[22] D. Habineza, M. Rakotondrabe, and Y. Le Gorrec, “Bouc-Wen model-
[1] Q. Xu, “Continuous integral terminal third-order sliding mode motion ing and feedforward control of multivariable hysteresis in piezoelectric
control for piezoelectric nanopositioning system,” IEEE/ASME Trans. systems: Application to a 3-dof piezotube scanner,” IEEE Trans. Control
Mechatronics, vol. 22, no. 4, pp. 1828–1838, Aug. 2017. Syst. Technol., vol. 23, no. 5, pp. 1797–1806, Sep. 2015.
[2] Z. Li, J. Shan, and U. Gabbert, “Inverse compensation of hysteresis us- [23] J. Shan, H. T. Liu, and S. Nowotny, “Synchronised trajectory-
ing Krasnoselskii-Pokrovskii model,” IEEE/ASME Trans. Mechatronics, tracking control of multiple 3-dof experimental helicopters,” Proc.
vol. 23, no. 2, pp. 966–971, Apr. 2018. Inst. Elect. Eng. Control Theory Appl., vol. 152, no. 6, pp. 683–692,
[3] Z. Li, X. Zhang, C. Y. Su, and T. Chai, “Nonlinear control of systems 2005.
preceded by Preisach hysteresis description: A prescribed adaptive control [24] D. Zhao, S. Li, and Q. Zhu, “Adaptive synchronised tracking control for
approach,” IEEE Trans. Control Syst. Technol., vol. 24, no. 2, pp. 451–460, multiple robotic manipulators with uncertain kinematics and dynamics,”
Mar. 2016. Int. J. Syst. Sci., vol. 47, no. 4, pp. 791–804, 2016.
[4] Z. Zhu, S. To, K. F. Ehmann, and X. Zhou, “Design, analysis, and [25] L. Wang, Model Predictive Control System Design and Implementation
realization of a novel piezoelectrically actuated rotary spatial vibration using MATLAB. New York, NY, USA: Springer, 2009.
system for micro-/nano-machining,” IEEE/ASME Trans. Mechatronics, [26] S. Chai, L. Wang, and E. Rogers, “Model predictive control of a permanent
vol. 22, no. 3, pp. 1227–1237, Jun. 2017. magnet synchronous motor with experimental validation,” Control Eng.
[5] Z. Li and J. Shan, “Inverse compensation based synchronization control of Pract., vol. 21, no. 11, pp. 1584–1593, 2013.
the piezo-actuated fabrycperot spectrometer,” IEEE Trans. Ind. Electron., [27] C.-Y. Lin and Y.-C. Liu, “Precision tracking control and constraint
vol. 64, no. 11, pp. 8588–8597, Nov. 2017. handling of mechatronic servo systems using model predictive control,”
[6] Z. Li and J. Shan, “Modeling and inverse compensation for coupled hys- IEEE/ASME Trans. Mechatronics, vol. 17, no. 4, pp. 593–605, Aug.
teresis in piezo-actuated Fabry-Perot spectrometer,” IEEE/ASME Trans. 2012.
Mechatronics, vol. 22, no. 4, pp. 1903–1913, Aug. 2017. [28] M. Rana, H. Pota, and I. R. Petersen, “High-speed AFM image scanning
[7] P. Krejci and K. Kuhnen, “Inverse control of systems with hysteresis and using observer-based mpc-notch control,” IEEE Trans. Nanotechnol.,
creep,” Proc. Inst. Elect. Eng.—Control. Theory Appl., vol. 148, no. 3, vol. 12, no. 2, pp. 246–254, Mar. 2013.
pp. 185–192, 2001. [29] M. Rana, H. Pota, and I. Petersen, “Nonlinearity effects reduction of an
[8] M. Rakotondrabe, “Bouc-Wen modeling and inverse multiplicative struc- AFM piezoelectric tube scanner using MIMO MPC,” IEEE/ASME Trans.
ture to compensate hysteresis nonlinearity in piezoelectric actuators,” Mechatronics, vol. 20, no. 3, pp. 1458–1469, Jun. 2015.
IEEE Trans. Automat. Sci. Eng., vol. 8, no. 2, pp. 428–431, Apr. 2011. [30] J. S. Milne, J. M. Dell, A. J. Keating, and L. Faraone, “Widely tunable
[9] Q. Xu, “Identification and compensation of piezoelectric hysteresis with- mems-based Fabry–Perot filter,” J. Microelectromech. Syst., vol. 18,
out modeling hysteresis inverse,” IEEE Trans. Ind. Electron, vol. 60, no. 4, pp. 905–913, 2009.
no. 9, pp. 3927–3937, Sep. 2013. [31] F. Preisach, “Über die magnetische nachwirkung,” Zeitschrift für Physik,
[10] G. Song, J. Zhao, X. Zhou, and J. A. De Abreu-Garcı́a, “Tracking vol. 94, pp. 277–302, 1935.
control of a piezoceramic actuator with hysteresis compensation using [32] D. Sun, X. Shao, and G. Feng, “A model-free cross-coupled control for
inverse Preisach model,” IEEE/ASME Trans. Mechatronics, vol. 10, no. 2, position synchronization of multi-axis motions: Theory and experiments,”
pp. 198–209, Apr. 2005. IEEE Trans. Control Syst. Technol., vol. 15, no. 2, pp. 306–314, Mar.
[11] Z. Li, O. Aljanaideh, S. Rakheja, C.-Y. Su, and M. Al Janaideh, “Compen- 2007.
sation of play operator-based Prandtl–Ishlinskii hysteresis model using a [33] Z. Li and J. Shan, “LQG-based synchronization control of Fabry-Perot
stop operator with application to piezoelectric actuators,” Int. J. Adv. Mech. spectrometer using multiple piezoelectric actuators,” in Proc. IEEE Int.
Sys., vol. 4, no. 1, pp. 25–31, 2012. Conf. Inf. Automat, 2016, pp. 448–453.
[12] X. Tan and J. S. Baras, “Adaptive identification and control of hysteresis in [34] M. S. Grewal, Kalman Filtering. Berlin, Germany: Springer, 2011.
smart materials,” IEEE Trans. Autom. Control, vol. 50, no. 6, pp. 827–839, [35] C. K. Chui and G. Chen, Kalman Filtering: With Real-Time Applications.
Jun. 2005. Berlin, Germany: Springer, 2008.
[13] H. Jiang, H. Ji, J. Qiu, and Y. Chen, “A modified Prandtl-Ishlinskii model [36] P. Mercorelli, “A switching Kalman filter for sensorless control of
for modeling asymmetric hysteresis of piezoelectric actuators,” IEEE a hybrid hydraulic piezo actuator using MPC for camless inter-
Trans. Ultrason., Ferroelect., Freq. Control, vol. 57, no. 5, pp. 1200–1210, nal combustion engines,” in Proc. IEEE Conf. Control Appl., 2012,
May 2010. pp. 980–985.

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.
1828 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 4, AUGUST 2019

Zhi Li received the Ph.D. degree in me- Ulrich Gabbert received the Dr.-Ing. degree in
chanical engineering from Concordia University, engineering and the Dr.-Ing. Habilitation degree
Montreal, QC, Canada, in 2015. in computational mechanics from the University
He was a Postdoctoral Research Fellow at of Magdeburg, Magdeburg, Germany, in 1974
Eindhoven University of Technology, the Nether- and 1988, respectively.
lands, and York University, Toronto, ON, Canada. He worked for a few years in industry, respon-
He was an Alexander von Humboldt Research sible for strength of materials, safety design,
Fellow at Otto-von-Guericke Universität Magde- and software development. At the beginning
burg, Magdeburg, Germany. He is currently a of the 1980s, he returned to the University of
Full Professor of the State Key Laboratory of Magdeburg and became the Head of a group
Synthetical Automation for Process Industries, developing finite-element software for industrial
Northeastern University, Shenyang, China. His research interests in- applications. Since 1992, he has been a Full Professor of computational
clude dynamics and control of smart actuators, and hysteresis modeling mechanics. His research interests include finite-element methods, smart
and compensation. structures, active vibration and noise control, and structural health
monitoring.

Jinjun Shan (SM’08) received the Ph.D. degree


in spacecraft design from Harbin Institute of
Technology, Harbin, China, in 2002.
He is currently Full Professor of Space
Engineering and Chair of the Department of
Earth and Space Science and Engineering,
York University, Toronto, ON, Canada. He was
a Postdoctoral Fellow with the University of
Toronto Institute for Aerospace Studies. His re-
search interests include dynamics, control, and
navigation.
Dr. Shan was a recipient of the Alexander von Humboldt (AvH)
Research Fellowship and JSPS Invitation Fellowship in 2012. Since
2012, he has been a Professional Engineer in the Province of Ontario.

Authorized licensed use limited to: NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS. Downloaded on October 28,2020 at 09:09:08 UTC from IEEE Xplore. Restrictions apply.

You might also like