IDENTIFICATION OF ASYMMETRIC BOUC-WEN MODEL BASED ON

GA ALGORITHM FOR A PIEZO-ACTUATED STAGE

Xin SHEN Jianguo ZHAO Qing XIAO Quan ZHANG* Yan PENG
School of Mechatronic Engineering and Automation, Shanghai University, Shanghai 200072, China
*Corresponding author, E-mail: lincolnquan@shu.edu.cn;

In recent years, piezoelectric actuators (PEAs) are widely used in precision positioning stage due to its outstanding
advantages of small size, high displacement resolution and fast response. However, in practical applications, PEAs are
also affected by the inherent nonlinear factors such as hysteresis and creep, which further results the positioning accuracy
deceasing of the stage. In this paper, a modified Bouc-Wen model is proposed to identify the hysteresis characteristics of
PEAs. In order to improve the identification accuracy of the model, the modified Bouc-Wen model parameters are
identified by the Genetic Algorithm (GA) which has a good global search capability. The experimental results show that
the range of the absolute error (RAE) of the modified Bouc-Wen model is reduced by 5.87% and the average fitness
value (AFV) is reduced by 4.87% compared to the standard Bouc-Wen model, which further validate the accuracy of the
proposed modified Bouc-Wen model.

Keywords: PEAs; Hysteresis; Bouc-Wen; GA algorithm

1. INTRODUCTION

In recent decades, the fields of semiconductor hysteresis models of differential equations were
lithography, micromachining, precision measurement, Backlash-like model, Bouc-Wen [3, 4] model and
ultra-precision machining, micro-assembly and Duhem [5] model. Preisach [6] model and
nanotechnology have been rapidly developed [1]. Prandtl-Ishlinskii model were typical representation of
Meanwhile, as one of the core technology in such operator hysteresis model. The neural network
fields, the requirements of the precision positioning hysteresis model can also identify the hysteresis
systems with high accuracy and response speed are characteristics by introducing a hysteresis factor,
getting higher and higher. PEAs are widely used in leading the network featuring the hysteresis
precision positioning stage due to its advantages of identification ability. The back propagation (BP)
high displacement resolution, high stiffness, large neural network and radical basis function (RBF)
actuating force and fast response[2]. However, the neural network were employed by many scholars
inherent hysteresis nonlinearity seriously degrades the because of their high precision approximation ability
output characteristics and positioning accuracy of [7]. Among the above models, the Bouc-Wen model
PEAs. Therefore, it is necessary to identify the exact owned the advantages of relatively fewer parameters
hysteresis model of PEAs in order to further predict and easy to obtain the analytical solution.
and compensate the output displacement accurately. However, since the actual output displacement
So far, many models for the behavior of hysteresis loop of the piezo-actuated stage also has
hysteresis nonlinearity of PEAs have been proposed, asymmetry characteristic, when the standard
including Physical hysteresis model, symmetric Bouc-Wen model is used to model the
Phenomenological hysteresis model and Neural output displacement of PEA or piezoelectric driving
network hysteresis model, etc. The physical hysteresis stage, a large fitting error is generated. In this research,
model was represented by the Jiles-Atherton model. in order to accurately characterize the asymmetric
The phenomenological hysteresis model was hysteresis of the output displacement of the
classified as two categories: the differential equation piezo-actuated stage, two parameters of asymmetric
model and the operator model. The common hysteresis factor ΔΕ and input bias coefficient φ are

978-1-7281-5253-0/19/$31.00 ©2019 IEEE

1
introduced to a standard symmetric Bouc-Wen Table 1. The different values of parameters (A, β, γ and n)
hysteresis model, leading the proposed modified Figure A β γ n
Bouc-Wen model featuring the asymmetric hysteresis Fig. 1 (a) {0.5, 1, 1 -0.5 1
identification ability. Then the genetic algorithm is 1.5, 2,
employed to recognize all parameters required by the 2.5}
proposed asymmetric Bouc-Wen hysteresis model. Fig. 1 (b) 1 {0.5, 1, -0.5 1
The experimental results verify the effectiveness of 1.5, 2,
the asymmetric Bouc-Wen hysteresis model and 2.5}
Fig. 1 (c) 1 1 {-0.6, 1
parameter identification method.
-0.3, 0,
The remainder of the paper is organized as
0.5, 1}
follows. In Section 2, an asymmetrical Bouc–Wen Fig. 1(d) 1 0.8 -0.5 {0.5, 1,
model is proposed to establish the hysteresis model of 1.5, 2,
the piezo-actuated stage. In Section 3, the GA 2.5}
algorithm is adopted to identify the parameters of the
asymmetric Bouc-Wen model. In Section 4, parameter As shown in Table 1, when one parameter value is
identification results are given. Finally, the conclusion varied, the other three parameters are fixed, in order to
is drawn in Section 5. observe the relationship between the nonlinear hysteresis
term h(t) and the input u(t) in the hysteresis model. The
2. THE MODIFIED ASYMMETRIC BOUC-WEN hysteresis curve shapes under different parameters is
MODEL shown in Fig. 1.

2.1. The introduction of the standard Bouc-Wen

model

The standard Bouc-Wen model can describe the

hysteresis behavior when the center of the hysteresis
curve is approximately symmetric. The physical system
of hysteresis behavior u(t )  ybw (t ) is expressed as
follows [5]:

ybw (t )   k  u(t )  (1   )k  h(t ) (1)

dh dx dx n 1 dx n
 A    h  h     h (2)
dt dt dt dt
Fig. 1 (a) The change of parameter A
where the hysteresis output ybw(t) described in Eq. (1)
consists of a linear term αk  u(t) and a nonlinear term (1 -
α)k  h(t). Where k (k > 0) is the stiffness coefficient of the
hysteresis system, α (0 < α < 1) is the weight coefficient,
t is time and h is the hysteresis nonlinear term of the
model. Its value is solved by the differential equation
shown in Eq. (2), where A, β and γ are the model
parameters related to the shape of the hysteresis loop; n
is the parameter which determining the smoothness of
the curve.

2.2. Influence of model parameters on the shape of

hysteresis curve
Fig. 1 (b) The change of parameter β
Given a specified input signal u(t) = 2 – 2cos(t),
different values of parameters A, β, γ and n are selected
to analyze the influence of the model parameters on the
shape of hysteresis curve.

2
  ybw (t )  k1u(t )  k2 h(t )

 dh(t )  A. dU (t )    dU (t )  h(t ) n 1  h(t )
 dt dt dt

dU (t ) n 1
    h(t )   (3)
 dt
 du (t )
    u(t )  sign( )
 dt
 U (t )  u(t   )

where δ is asymmetric compensation coefficient, k1 is the

Fig. 1 (c) The change of parameter γ
linear coefficient, and k2 is the nonlinear coefficient.

2.4. Influence of the asymmetric parameters on the

shape of hysteresis curve

As shown in Fig. 2 (a) and Fig. 2 (b), with the change of

the asymmetric compensation coefficient δ and input bias
coefficient φ, the hysteresis loop is also occurred the
subtle rotation, and the combination of the two
parameters finally corrects the effect of the asymmetric
model error to some extent.

Fig. 1 (d) The change of parameter n

Figure 1. The hysteresis curve shapes under different
parameters

As shown in Fig. 1, we notice that the parameters A

and β are used to adjust the amplitude and rotation angle
of the hysteresis loop, the parameter γ can control the
rotation angle of the hysteresis loop, and the parameter n
affects the smoothness of the hysteresis loop.

2.3. The modified asymmetric Bouc-Wen model

Fig. 2 (a) The change of parameter δ
Due to the asymmetry characteristics of the hysteresis
loop of the piezo-actuated stage, the use of the standard
Bouc-Wen model to describe the stage hysteresis loop
will lead relatively large errors. Therefore, it is necessary
to introduce asymmetric hysteresis parameters to modify
the Bouc-wen model, which helps the model have the
ability to describe the asymmetric hysteresis loop of the
piezo-actuated stage. According to Eq. (1-2), an input
bias coefficient φ and an asymmetric hysteresis factor ΔΕ
are simultaneously introduced to compensate the phase
offset and the asymmetric hysteresis error of the model.
Fig. 2 (b) The change of parameter φ
The asymmetric Bouc-Wen model is defined as follows:
Figure 2. The hysteresis curve shapes under different
asymmetric model parameters

3. MODEL PARAMETER IDENTIFICATION

BASED ON GENETIC ALGORITHM

3
3.1. Genetic algorithm 1 N
fitness(k1 , k2 , A,  ,  , n,  ,  )  [  ( y  ybw )2 ]1/ 2 (4)
N i 1
Genetic algorithm is a computational method for
simulating the biological evolution process of natural
where y is the actual experimental displacement, ybw is
selection and genetics, which is also an approach for
the output displacement of the proposed model, and N
searching optimal solutions in an iterative process.
represents the number of the experiment.
Genetic algorithm has the advantages of simple
algorithm, process adaptation and parallel processing. It
4. EXPERIMENTAL VERIFICATION OF THE
is widely used in parameter estimation, especially for
MODIFIED MODEL
nonlinear, multi-model and multi-objective function
optimization problems. 4.1. The construction of the experimental stage
3.2. Parameter Identification of Asymmetric The experimental stage is constructed to research the
Hysteresis Model input and output relationship of the piezo-actuated stage.
In the genetic algorithm, after the initial group is In particular, the XY stage is a novel decoupled XY
composed by coding, a certain operation is applied to the compliant micro-positioning stage based on bridge
individual fitness value in the group, and finally the amplification mechanism and parallelogram mechanism.
evolution process of the fittest individual is realized. In this stage, the bridge amplification mechanism is used
Three basic genetic operators, including selection, for displacement amplification, and the parallelogram
crossover and variation constitute genetic manipulation. mechanism is used to reduce coupling displacement from
The hysteresis model parameter schematic of GA the other direction [8]. The systems consist of a PC, the
algorithm is shown in Fig. 3, in this paper, the amplifier and controller, PEA, XY stage and laser sensor.
asymmetric Bouc-Wen model is identified by GA The experimental setup is shown in Fig. 4. First, the
algorithm. The parameter identification process of input signal is generated from PC and amplified by the
hysteresis model can be transformed into the amplifier and controller, the input signal of PEA is a
optimization process of searching the optimal solution cosine voltage with the frequency of 2 Hz with u(t) = 10
for the set of parameters {k1, k2, A, β, γ, n,δ, φ}. – 10cos(4πt) to drive the XY stage. Then the output
displacements of the XY stage in X direction are
gathered by the laser sensor.

The amplifier and PC

controller of PEA
XY stage

Laser Sensor

Figure 3. The hysteresis model parameter schematic

of GA algorithm
PEA

The parameter identification process consists of the

following processes: First, encode the parameter set, Figure 4. The construction of the experimental stage
initialize the population and assign the appropriate range
4.2. Simulation result and discussion
values to the parameters in the population; Second,
determine the fitness function which is derived as Eq. (4) Since the Bouc-Wen model is a phenomenological
to assess individual quality and evaluate the fitness of the hysteresis model, the accuracy of the model depends on
group; Thirdly, conduct the genetic manipulation which the number of samples, and hence the number of periodic
includes selection, crossover and mutation to enhanced samples is required as large as possible. The samples set
search ability of the population; Finally, loop the last is N=10000 in this paper. From experience, although the
three steps until the maximum number of iterations of the convergence value will become more accurate as the
GA algorithm is met, or the fitness function reaches the number of populations and the number of iterations
set value. increased, the running time will increase greatly. Hence,

4
the number of populations is taken as P= 30 and the calculated each time, and then theaverage value is
maximum number of iterations is set to 100 adopted.
compromisingly. The initial mutation probability and
crossover probability of GA algorithm are taken as Mu =
0.2 and Cr = 0.7 respectively. The stopping threshold
value of the proposed algorithm is set to 10-7. According
the experience gained from a large amount of work, the
six basic parameter {k1, k2, A, β, γ, n} ranges ([xmin, xmax])
of the Bouc-Wen model are set to [-0.001, 0.001], [-1, 1],
[-1, 1], [-10, 10], [0, 10] and [2, 5] respectively. The
variation limit of the two asymmetric parameters {δ, φ}
are set to [-1, 1] and [-0.0001, 0.0001] respectively.
During the identification process, the six basic
parameters of the Bouc-Wen model are first identified, Fig. 5(a) The comparisons of the hysteresis loop

and in order to verify the correction effect of the

modified model, the GA algorithm is used to identify the
asymmetric compensation coefficient δ and the input bias
coefficient φ again while ensuring that the six basic
parameters are unchanged.
In order to obtain numerical solutions, fourth-order
Runge-Kutta method is adopted in Matlab R2016a. Then,
the parameter identification is designed for three cases,
namely Case Ⅰ (identified parameters: {k1, k2, A, β, γ, n}),
Case Ⅱ (identified parameters: {k1, k2, A, β, γ, n,δ} ) and
Case Ⅲ (identified parameters: {k1, k2, A, β, γ, n,δ, φ}).
The results of the identified parameters are shown in
Table 2.
Fig. 5(b) The comparisons of the output tracking displacement
Table 2. All of the identified model parameters information
Figure 5. The chart of result comparison
Model data Case Ⅰ Case Ⅱ Case Ⅲ
-4 -4 -4
k1 -3.51711×10 -3.51711×10 -3.51711×10 From Table 3, it can be calculated that the range of
k2 -1.94566×10-3 -1.94566×10-3 -1.94566×10-3 the RAE in Case Ⅱ is reduced by 5.6% and the AFV is
A -7.02182×10-2 -7.02182×10-2 -7.02182×10-2 reduced by 4.48% compared to Case Ⅰ. At the same time,
β 4.45753 4.45753 4.45753 the RAE in Case Ⅲ is reduced by 5.87% and the AFV is
γ 3.49259 3.49259 3.49259 reduced by 4.87% compared to Case Ⅰ. These results
n 3.05869 3.05869 3.05869 verify that the modified model can better describe the
-2
δ 0 -1.95×10 -1.95×10-2 asymmetry hysteresis characteristics of the
φ 0 0 5.16×10-7 piezo-actuated stage than the standard Bouc-Wen model.
fitness(μm) 0.361431 0.345347 0.343865
Table 3. the model performance comparison results
From Table 2, the fitness value of Case Ⅲ is Situation Case Ⅰ Case Ⅱ Case Ⅲ
3.43865×10-4, which is similar to the fitness values of
RAE(μm) [-1.0972,1.118] [-1.0803,1.0096] [-1.0764,1.0088]
Case Ⅱ (3.45347×10-4) and much better than Case Ⅰ
(3.61431×10-4). The comparisons of the hysteresis loop AFV(μm) 0.361547 0.345347 0.343951
between the experimental results and the Bouc-Wen
model based predicted outputs are shown in Fig. 5(a),
and the comparisons of the output tracking displacements 5. CONCLUSION
between the experimental results and the Bouc-Wen In this paper, a modified Bouc-Wen model is adopted to
model based predicted output are shown in Fig. 5(b). In model the hysteresis behavior of PEAs. Compared with
order to avoid the accidental situation, the experiment is the standard Bouc-Wen model, the modified model
conducted ten times, the fitness value of the population is introduce an asymmetric compensation coefficient δ and

5
an input offset coefficient φ to describe the asymmetric [3] Qin H, Bu N, Chen W, Yin Z. An Asymmetric Hysteresis
hysteresis characteristics of the piezo-actuated stage. The Model and Parameter Identification Method for
GA algorithm is used to identify the parameters of the
Piezoelectric Actuator. Mathematical Problems in
modified Bouc-Wen model and the results show that the
Engineering 2014: 1-14, 2014.
GA algorithm can effectively identify the parameters.
The experimental results show the RAE in the proposed [4] Zhang Q, Dong Y, Peng Y, Luo J, Xie S, Pu H.
model is reduced by 5.87% and the AFV is reduced by Asymmetric Bouc–Wen hysteresis modeling and inverse
4.87% compared to the standard Bouc-Wen model. compensation for piezoelectric actuator via a genetic
These results verify that the modified model can better
algorithm–based particle swarm optimization
describe the asymmetry hysteresis characteristics of the
identification algorithm. Journal of Intelligent Material
piezo-actuated stage than the standard Bouc-Wen model.
Systems and Structures 30(8): 1263-1275, 2019.
ACKNOWLEDGEMENTS [5] Wang G and Chen G. Identification of piezoelectric
hysteresis by a novel Duhem model based neural network.
This work was supported by the National Natural
Sensors and Actuators A: Physical 264: 282-288, 2017.
Science Foundation of China under Grant 61973207,
91648119, 51605271, the Natural Science Foundation of [6] Dong Y, Hu H, Wang H. Identification and experimental
Shanghai 19ZR1474000 and 17DZ1205000, and the assessment of two-input Preisach model for coupling
equipment pre-research foundation for "the 13th hysteresis in piezoelectric stack actuators. Sensors and
Five-Year Plan" 61403120306.
Actuators A: Physical 220: 92-100, 2014.
[7] Cheng L, Liu W, Hou ZG, Yu J, Tan M.
REFERENCES
Neural-Network-Based Nonlinear Model Predictive
[1] Gu GY, Zhu LM. Motion control of piezoceramic Control for Piezoelectric Actuators. IEEE Transactions on
actuators with creep, hysteresis and vibration Industrial Electronics 62(12): 7717-7727, 2015.
compensation. Sensors and Actuators A: Physical 197: [8] Zhang Q, Zhao J, Shen X, Xiao Q, Huang J, Wang Y.
76-87, 2013. Design, Modeling, and Testing of a Novel XY
[2] Li W, Chen X. Compensation of hysteresis in piezoelectric Piezo-Actuated Compliant Micro-Positioning Stage.
actuators without dynamics modeling, Sensors and Micromachines (Basel) 10(9), Aug 31 2019.
Actuators A: Physical 199: 89-97, 2013.