IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control ,
1, January 2009 101
FPGA-Based Elman Neural Network Control
System for Linear Ultrasonic Motor
Faa-Jeng Lin, Senior Member, IEEE, and Ying-Chih Hung
Abstract—A field-programmable gate array (FPGA)-based
Elman neural network (ENN) control system is proposed to
control the mover position of a linear ultrasonic motor (LUSM)
in this study. First, the structure and operating principle of
the LUSM are introduced. Because the dynamic character-
istics and motor parameters of the LUSM are nonlinear and
time-varying, an ENN control system is designed to achieve
precision position control. The network structure and online
learning algorithm using delta adaptation law of the ENN are
described in detail. Then, a piecewise continuous function is
adopted to replace the sigmoid function in the hidden layer of
the ENN to facilitate hardware implementation. In addition,
an FPGA chip is adopted to implement the developed control
algorithm for possible low-cost and high-performance indus-
trial applications. The effectiveness of the proposed control
scheme is verified by some experimental results.
I. Introduction
B iotechnology, semiconductor and nano-technology,
and other technologies have recently implemented in-
novations in high-precision accuracy. Therefore, precise
positioning technology is required in many fields, espe-
cially for the advance of semiconductor manufacturing
equipment. This type of manufacturing equipment re-
quires highly precise positioning, multidimensional drive,
miniaturization, and lightweight materials. Piezoelectric
ceramic linear ultrasonic motors (LUSMs) are one of the
new kinds of ultrasonic motors (USMs) that are driven by
the ultrasonic vibration force of piezoelectric ceramic ele-
ments and the mechanical friction effect [1]–[3]. Different
constructions and driving principles of LUSMs have been
reported [3], [4]. They permit high precision, fast control
dynamics, and large driving force in small dimensions.
Thus, some precision motion control stages using LUSMs,
which are small and light and have highly precise position-
ing, have been developed for industrial applications [5],
[6]. However, the control accuracy of the LUSMs has many
uncertainties, including parameter variations, external
disturbances, and high-order dynamics. In addition, the
frictional force dynamics between 2 moving bodies should
also be considered in the linear actuator system. Further-
more, the mathematical models of piezoelectric ceramic
LUSMs are complex, and the motor parameters are time-
Manuscript received May 29, 2008; accepted August 15, 2008. The
author would like to acknowledge the financial support of the National
Science Council of Taiwan, ROC, through its grant NSC 95-2221-E-008-
177-MY3.
The authors are with the Department of Electrical Engineering, Na-
tional Central University, Taiwan, Republic of China (e-mail: linfj@
ee.ncu.edu.tw).
Digital Object Identifier 10.1109/TUFFC.2009.1009
0885–3010/$25.00
Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
vol. 56, no.
varying due to increases in temperature and changes in
motor drive operating conditions. In addition, the control
characteristics of the LUSMs are complicated and highly
nonlinear. Therefore, the intelligent control approaches are
good candidates to control the motion control stage using
LUSM for high performance applications due to their abil-
ity to approximate nonlinear systems without using the
mathematical models of the controlled system.
Intelligent control approaches such as neural network
and fuzzy system do not require mathematical models
and have the ability to approximate nonlinear systems.
Therefore, there were many researchers using intelligent
control approaches to represent complex plants and con-
struct advanced controllers [7]. The Elman neural network
(ENN) was first proposed for speech processing [8]. Due
to its distinguished dynamical characteristics, it also has
been widely applied in dynamical systems identification
and control [9]–[12]. Generally, ENN can be considered to
be a special kind of feed-forward neural network with ad-
ditional memory neurons [8]. Due to the context neurons,
it has certain dynamical advantages over static neural
network, such as multilayer perceptrons and radial-basis
function networks (RBFN). Furthermore, the ENN can
approximate high-order systems with high precision, and
its converge speed is fast.
The recurrent neural network has received increasing
attention due to its structural advantage in nonlinear sys-
tem modeling and dynamic system control [5], [13], [14].
The most important characteristic of the recurrent neural
network is its self-connection to memorize feedback infor-
mation of the historical influence in the same neuron. The
adopted ENN can be considered to be a special type of
recurrent neural network with feedback connections from
the hidden layer to the context layer. The context layer
is an additional layer that is used as an extra memory to
memorize previous activations of the hidden neurons and
feed to all the hidden neurons after the one-step time de-
lay. Therefore, compared with the general recurrent neu-
ral networks, ENN has a special explicit memory to store
the temporal information. Moreover, in the general recur-
rent neural networks, the specific self-connection feedback
of the hidden neuron or output neuron is responsible for
memorizing the specific previous activation of the hidden
neuron or output neuron and feed to itself only. Therefore,
the outputs of the other neurons have no ability to affect
the specific neuron. However, in the complicated, time-
varying, and nonlinear dynamic system such as LUSM, the
inherent cross-coupled interference of each state is always
a factor. Hence, if each neuron in the recurrent neural
networks is considered as a state in the nonlinear dynamic
© 2009 IEEE
 102 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

systems, the self-connection feedback type is unable to ap- presents the structure and operating principle of the SP
proximate the dynamic systems efficiently. On the other series LUSM briefly. In Section III, an ENN controller is
hand, the feedbacks in ENN not only are self-connecting designed to control the mover position of the LUSM. In
but they also store in the context neurons and feed to all addition, the back-propagation method is derived using
the hidden neurons. Thus, the structure of ENN is more delta adaptation law to train the weights of the ENN on-
powerful than the general recurrent neural networks for line. In Section IV, the circuit design on the FPGA chip
dealing with time-varying, and nonlinear dynamic systems is discussed. In Section V, experiments are performed to
due to the cross-coupled interference of each state can be confirm the effectiveness of the proposed control scheme.
approximated efficiently with the additional context layer Conclusions are drawn in Section VI.
[14], [15].
A field-programmable gate array (FPGA) incorporates
the architecture of gate arrays and the programmability II. Linear Ultrasonic Motor Motion
of a programmable logic device (PLD). It consists of thou- Control Stage
sands of logic gates, some of which are combined together
to form a configurable logic block (CLB) thereby simpli-
The motion control stage can be divided into 2 parts:
fying high-level circuit design. Interconnections between
the LUSM servo drive and a single-axis stage as shown
logic gates using software are externally defined through
in Fig 1. The AB1A-SP-E1 drive manufactured by Nano-
static random access memory (SRAM) and ROM, which
motion Ltd. (Yokneam, Israel) is adopted for the servo
will provide flexibility in modifying the designed circuit
drive. Moreover, the motion control stage comprises one
without altering the hardware. Moreover, concurrent op-
SP series LUSM, the SP1–02 LUSM, which is also manu-
eration, simplicity, programmability, a comparatively low
factured by Nanomotion Ltd. The travel of the moving
cost, and rapid prototyping make it the favorite choice
table of the motion control stage is linear. The moving
for prototyping an application specific integrated circuit
table is directly driven using the spacer of the LUSM. Fur-
(ASIC) [16], [17]. All the internal logic elements and all
thermore, the output of the ENN is the command voltage
the control procedures of the FPGA are executed continu-
(control effort) of the LUSM servo drive. The command
ously and simultaneously. On the other hand, if the con-
voltage with ±10 V is applied to the drive AB1A-SP-E1.
trol algorithm is executed sequentially using software in a
The drive generates a fixed frequency (39.6 kHz) sine wave
digital signal processor (DSP) or personal computer (PC),
voltage whose amplitude (0–280 Vrms) is a function of
the minimum execution time is limited. Therefore, the ex-
the applied command voltage (0–10 V). The generated
ecution time of the FPGA is faster than either a DSP or
sine wave voltage drives the LUSM and then generates
PC. Furthermore, the circuits and algorithms can be de-
thrust force and velocity to rotate the moving table to
veloped in the very high speed integrated circuit (VHSIC)
reach the desired position. Detailed structure and operat-
hardware description language (VHDL) [16], [17]. This
ing principle of LUSM can be referred in the references
method is as flexible as any software solution. Another
[3]–[5], [18].
important advantage of VHDL is that it is technology
independent. The same algorithm can be synthesized into
any FPGA and even has a direct path to an ASIC to open
III. Proposed Control System
interesting possibilities in industrial applications in terms
of performance and cost. However, the major disadvantage
A. Elman Neuron Network
of an FPGA-based system for hardware implementation is
the limited capacity of available cells. In this study, the
The structure of the ENN is shown in Fig. 2(a), includ-
FPGA chip is adopted to implement the proposed ENN
ing the input layer (i layer), the hidden layer (j layer),
controller to allow possible low-cost and high-performance
the context layer (r layer), and the output layer (o layer).
industrial applications. The proposed control algorithms
In the network, the hidden layer neurons are fed by the
are realized on a 24 MHz FPGA (XC2V1000) with 1 mil-
outputs of the context neurons and the input neurons.
lion gate counts and 10,240 flip-flops from Xilinx, Inc.
Context neurons are known as memory units because they
(San Jose, CA) using VHDL with the Q-format arithme-
store the previous output of hidden neurons. Signal propa-
tic representation. The design and implementation of the
gation and the basic function of each layer are introduced
FPGA-based control system will be described in detail.
in the following:
Compared with a PC-based control system, the merits
of the FPGA-based control system are parallel process-
1) Layer 1 (Input Layer): In the input layer, the node
ing and small size in addition to low cost. The developed
input and the node output are represented as
VHDL code also can be easily modified and implemented
to control any type of motors.
e i(k ) = f i(net i) = net i, i = 1, 2. (1)
In this study, an FPGA-based ENN controller is de-
veloped to control the mover position of an LUSM to
track periodical sinusoidal and trapezoidal reference tra- where neti represents the ith input to the input layer. In
jectories with accuracy tracking performance. Section II this study, the inputs of the ENN are e1(k) = e and

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor 103

Fig. 1. Photo of motion control stage.

e2(k) = e, which are the tracking error and its derivative, x rc(k ) = x j(k - 1). (4)
respectively.

2) Layer 2 (Hidden Layer): In the hidden layer, the

4) Layer 4 (Output Layer): In the output layer, the
receptive field function is usually a sigmoid function or
node input and the node output are represented as
a logistic function [8], [11]. To reduce the computational
requirements and facilitate the hardware implementation
y o(k ) = f (net o(k )) = net o(k ) (5)
using FPGA, a piecewise continuous function shown in
Fig. 2(b) is selected as the receptive field function to ap-
proximate sigmoid function in the following:
net o(k) = åW jo2 ´ x j(k) (6)
j

x j (k ) =
ìï1 where W jo2 are the connective weights of hidden neurons
ïï if net j > 0.5/a,
í 0 if net j < -0.5/a , j = 1, 2, , 9, to output neurons; yo(k) is output of the ENN and also is
ïï the control effort denoted by V*. The control block dia-
ïî a ´ net j + 0 . 5 otherwise.
gram of ENN control system is shown in Fig. 3, where dm
(2) and d represent the reference trajectory and mover posi-
net j = å x rc(k) + åW ij1 ´ net i(k), r = 1, 2, , 9, tion of the LUSM, respectively.
r i
(3) B. Online Learning Algorithm Using Delta Adaptation
Law
where netj and xj(k) are the input and the output of the
hidden layer, respectively; x rc(k ) is output of the context The central part of the learning algorithm for an ENN
concerns how to recursively obtain a gradient vector in
layer; W ij1 are the connective weights of input neurons to which each element in the learning algorithm is defined
hidden neurons; and a is the slope of linear function. as the derivative of an energy function with respect to
a parameter of the network. This is done by means of
3) Layer 3 (Context Layer): In the context layer, the the chain rule, and the method is generally referred to as
node input and the node output are represented as the back-propagation learning rule. To describe the online

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 104 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

1) Layer 4: The error term to be propagated is given

by

¶E ¶E ¶e ¶E ¶e ¶d
do = - =- =- (8)
¶y o(k) ¶e ¶y o(k) ¶e ¶d ¶y o(k)

and the weight is updated by the amount

¶E ¶E ¶y o(k) ¶net o(k)

DW jo2 = -h 1 = -h 1
¶W jo2 ¶y o(k) ¶net o(k) ¶W jo2 (9)
= h 1d ox j(k).

The connective weight W jo2 is updated according to the

following equation:

W jo2 (k + 1) = W jo2 (k ) + DW jo2 (10)

where the factor η1 is the learning rate.

2) Layer 2: By using the chain rule, the update law of

W ij1 is

¶E ¶E ¶y o(k ) ¶net o(k ) ¶x j(k )

DW ij1 = -h 2 = -h 2
¶W ij1 ¶y o(k ) ¶net o(k ) ¶x j(k ) ¶W ij
1

2
= h 2d oW joae i(k ).
(11)
Fig. 2. Network structure of ENN: (a) network structure and (b) piece-
wise continuous function in hidden layer.
The connective weight W ij1 is updated according to the
following equation:

learning algorithm of the ENN using supervised gradient
decent method, first the energy function E is defined as
1 1
E = (d m - d) 2 = e 2. (7)
2 2
Then the learning algorithm is described as follows:
W ij1(k + 1) = W ij1(k ) + DW ij1 (12)
where the factor η2 is the learning rate.
The exact calculation of the Jacobian of the controlled
plant, ∂d/∂yo(k), cannot be determined due to the un-
known dynamics of the LUSM drive system. To overcome
the problem and increase the online learning speed of the

Fig. 3. Block diagram of ENN control system.

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor
Fig. 4. FPGA-based control system for motion control stage using LUSM drive.
weights, the delta adaptation law is adopted as follows
[19]:
d o @ (d m - d) + (d m - d) = e + e (13)
where d m and d represent the first derivative of the
reference trajectory and the mover position of the LUSM,
respectively.
IV. Circuit Design on FPGA Chip
The FPGA-based control system for a motion control
stage is shown in Fig. 4. A linear scale with resolution
Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
105
1 μm is adopted to detect the position of the moving table.
To implement the ENN control system, the timing con-
trol module, encoder interface module, and the ENN con-
trol module are realized on the FPGA chip, respectively.
Moreover, to show the effectiveness of the control system
with a small number of neurons, the ENN has 2, 9, 9, and
1 neurons at the input, hidden, context, and output layers,
respectively. The sampling interval, i.e., the execution in-
terval, of the control algorithm is 1.366 ms (732 Hz). Two
D/A converters are used to display the reference trajec-
tory dm, the mover position d, the control effort V*, and
the tracking error e alternately on the oscilloscope. The
entire I/O port for this chip includes 2 pins for the input
ports and 36 pins for the output port. Furthermore, 3256
 106 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

Fig. 5. Circuits design on FPGA: (a) encoder interface module, (b) input layer in ENN control module, (c) hidden and context layer in ENN control
module, (d) output layer in ENN control module, and (e) online learning algorithm in ENN control module.

out of 10,240 flip-flops (31%) have been used in the FPGA
chip. In addition, the circuits and control algorithm in
the FPGA are developed using VHDL by a personal com-
puter as the development system. In the development of
the VHDL codes for a 12-bits FPGA processor, by using
fractional numbers between –1 and 1 with the Q-format
representation, the results of multiplication can be easily
handled because the product of 2 numbers is always in
the same range. Besides, all the sines and multipliers are
implemented using available intellectual property (IP).
A. Encoder Interface Module
The block diagram of the encoder interface module
is shown in Fig. 5(a), which consists of timing control,
2 digital filters, a decoder, an up-down counter, 2 clock

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor 107

Fig. 6. Experimental results of RBFN control system for periodical sinusoidal reference trajectory: (a) tracking response at nominal condition, (b)
control effort at nominal condition, (c) tracking error at nominal condition, (d) tracking response at parameter variation condition, (e) control effort
at parameter variation condition, and (f) tracking error at parameter variation condition.

(CLK) generators, a register, a command generator, and
one adder. The function of the encoder interface is to ob-
tain the position and speed values of the mover. The pulse
count signal PLS and the rotating direction signal DIR
are obtained using the A, B pulse input signals from the
decoder through 2 digital filters. The position signal d
can be obtained using the PLS and DIR signals through
up-down counter. In addition, the command generator in-
cludes periodical sinusoidal IP and periodical trapezoidal
IP to generate the reference trajectory dm. Also, v is the
velocity signal and results from the difference between the
position signal d and the time delay of d, ddelay. In addi-
tion, the resolution of the encoder is 1 mm = 1000 digital
value at the sampling frequency 732 Hz.
B. ENN Control Module
To implement the control law effectively, the ENN con-
trol module is divided into 4 parts that are shown in Figs.
5(b) to 5(e) with various control signals to enable the
respective circuit:
1) Fig. 5(b) shows the block diagram of input layer, us-
ing dm, d, e, and e to construct.
2) Fig. 5(c) shows the block diagram of the hidden and
context layers used to obtain the output of the hid-
den layer x1(k) ~ x9(k) using e, e , and x rc(k ) based
on (2), (3), and (4). First, 2 multipliers are used to
multiply the outputs of the input layer, e and e ,
with the weights between input layer and hidden
layer W 111 ~ W 191 and W 21 1
~ W 29 1
individually.
Then, add the summation of the output of the con-
text layer to result in net1 ~ net9 based on (3). Also,
the ouputs of each neuron in hidden layer x1(k) ~
x9(k) are generated using linear functions based on
(2) where all the slopes are 1.0.
3) Fig. 5(d) shows the block diagram of the output lay-
er used to obtain the output of the ENN, V*, using

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 108 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

Fig. 7. Experimental results of RBFN control system for periodical trapezoidal reference trajectory: (a) tracking response at nominal condition, (b)
control effort at nominal condition, (c) tracking error at nominal condition, (d) tracking response at parameter variation condition, (e) control effort
at parameter variation condition, and (f) tracking error at parameter variation condition.

x1(k) ~ x9(k) based on (5) and (6). First, 2 multiplex-

W 111(k + 1) ~ W 191(k + 1), and 1
W 21 (k + 1) ~
ers with a multiplier are used to multiply the out-
puts of the hidden layer x1(k) ~ x9(k) with the weights 1
W 29 (k + 1) are obtained by adding the weights of
between the hidden layer and the output layer
current iteration W 112(k ) ~ W 91
2
(k) with DW 112 ~
W 112 ~ W 91
2
individually. Then, the output V* is
2
DW 91 , W 111(k ) ~ W 191(k ) with DW 111 ~ DW 191,
obtained through a register and an adder.
4) Fig. 5(e) shows the block diagram of the online learn- 1
and W 21 1
(k ) ~ W 29 1
(k ) with DW 21 1
~ DW 29 indi-
ing algorithm used to achieve the adaptation laws of
the weights between the hidden layer and the output vidually.
layer based on (9), (10), and the weights between the
input layer and the hidden layer based on (11), (12).
First, the variations of the weights DW 112 ~ DW 91 2 V. Experimental Results
are obtained by an adder, a multiplexer, 2 multipli- The control objective is to control the mover of the
ers and a register; the variations of the weights LUSM to move periodically according to the reference tra-
DW 111 ~ DW 191 and DW 21 1
~ DW 29 1
are obtained jectories including sinusoidal and trapezoidal commands
using the proposed ENN controller. A RBFN controller
by 2 multiplexers, 4 multipliers, and 2 registers. [20]–[22] is also implemented in the experimentation for
Then, the new weights W 112(k + 1) ~ W 91
2
(k + 1), the comparison of the control performance. In the adopted

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor 109

Fig. 8. Experimental results of ENN control system for periodical sinusoidal reference trajectory at nominal condition: (a) tracking response, (b)

control effort, (c) tracking error, (d) connected weight W 512 between hidden layer and output layer, (e) connected weight W 151 between input layer
1
and hidden layer, and (f) connected weight W 25 between input layer and hidden layer.

RBFN, only the connected weights between the hidden n

layer and the output layer are trained on line. Moreover, a
triangular function is also adopted to replace the Gaussian
TS = å (T e(k) - m) 2 n. (16)
k =1
function in the hidden layer of the RBFN to facilitate the
hardware implementation. Furthermore, the RBFN has 2, In addition, 2 test conditions are provided in the ex-
9, and 1 neurons at the input, hidden, and output layers, perimentation, which are the nominal condition and the
respectively. To measure the control performance of the parameter variation condition. The parameter variation
propose control systems, the maximum tracking error TM, condition is the addition of one iron disk with the mass
the average tracking error m, and the standard deviation 3.66 kg on the moving table.
of tracking error TS for the trajectory tracking are defined Some experimental results are provided to demonstrate
as follows [13]: the effectiveness of the proposed FPGA-based control sys-
tems. The control objective is to control the mover to
move ±1 mm periodically for sinusoidal reference trajec-
T M = max(T e(k)), where T e(k) = d m(k) - d(k) (14) tory and 1 mm periodically for trapezoidal reference tra-
k
jectory. Fig. 6 and 7 depict the experimental results of the
command tracking due to the periodical sinusoidal and
æ n ö÷
çç ÷÷ trapezoidal reference trajectories of the RBFN control
çç å T
( e(k ))
÷÷ system. The tracking responses of the mover position due
m = ççè k = 1 ÷ø (15)
n to the periodical sinusoidal reference trajectory at nomi-

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 110 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

Fig. 9. Experimental results of ENN control system for periodical trapezoidal reference trajectory at nominal condition: (a) tracking response, (b)
2 1
control effort, (c) tracking error, (d) connected weight W 51 between hidden layer and output layer, (e) connected weight W 15 between input layer
1
and hidden layer, and (f) connected weight W 25 between input layer and hidden layer.

nal condition and parameter variation condition are shown
in Fig. 6(a) and (d); the associated control efforts are
shown in Fig. 6(b) and (e); the tracking errors are shown
in Fig. 6(c) and (f). The tracking responses of the mover
position due to the periodical trapezoidal reference trajec-
tory at nominal condition and parameter variation condi-
tion are shown in Fig. 7(a) and (d); the associated control
efforts are shown in Fig. 7(b) and (e); the tracking errors
are shown in Fig. 7(c) and (f). From the experimental re-
sults, favorable tracking responses of the mover can be
obtained at both the nominal and the parameter variation
conditions for the RBFN controller as shown in Fig. 6(a),
6(d), 7(a), and 7(d). However, the chattering phenomena
in the control efforts are serious and affect the tracking
accuracy. In addition, Fig. 8 to 11 depict the experimental
results of the command tracking due to the periodical si-
nusoidal and trapezoidal reference trajectories of the pro-
posed ENN control system. The tracking responses of the
mover position due to the periodical sinusoidal reference
trajectory at nominal condition and parameter variation
condition are shown in Fig. 8(a) and 10(a); the associated
control efforts are shown in Fig. 8(b) and 10(b); the track-
ing errors are shown in Fig. 8(c) and 10(c); the responses
of one of the connected weights between the hidden layer
and the output layer of the ENN, W 51 2
, are shown in Fig.
8(d) and 10(d); the responses of 2 of the connected weights
between the input layer and the hidden layer of the ENN,
W 151 and W 251
, are shown in Fig. 8(e) and 10(e), and Fig.
8(f) and 10(f), respectively. Furthermore, the tracking re-
sponses of the mover position due to the periodical trap-
ezoidal reference trajectory at nominal condition and pa-
rameter variation condition are shown in Fig. 9(a) and
11(a); the associated control efforts are shown in Fig. 9(b)
and 11(b); the tracking errors are shown in Fig. 9(c) and
11(c); the responses of one of the connected weights be-
tween the hidden layer and the output layer of the ENN,
2
W 51 , are shown in Fig. 9(d) and 11(d); the responses of 2

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor 111

Fig. 10. Experimental results of ENN control system for periodical sinusoidal reference trajectory at parameter variation condition: (a) tracking
2 1
response, (b) control effort, (c) tracking error, (d) connected weight W 51 between hidden layer and output layer, (e) connected weight W 15 between
1
input layer and hidden layer, and (f) connected weight W 25 between input layer and hidden layer.

of the connected weights between the input layer and the
hidden layer of the ENN W 151 and W 25 1
, are shown in Fig.
9(e) and 11(e), and Fig. 9(f) and 11(f), respectively. From
the experimental results, good tracking responses of the
mover can be obtained at both the nominal and the pa-
rameter variation conditions for the proposed ENN con-
troller as shown in Fig. 8(a), 9(a), 10(a), and 11(a). It also
shows that the proposed FPGA-based ENN control sys-
tem has effective hardware online learning ability of the
network parameters and is robust with regard to the pa-
rameter variation.
The performance measures of the RBFN control sys-
tem and the ENN control system for the tracking of the
sinusoidal and trapezoidal reference trajectories are shown
in Table I and Table II, respectively. Comparing the per-
formance measures of the RBFN control system and ENN
control system, the tracking errors are reduced and the
chattering phenomena of the control efforts are also de-
creased in the proposed ENN control system. Therefore,
the FPGA-based ENN control system possesses robust
and accurate control performance.
VI. Conclusions
Because the dynamic characteristics of the LUSM are
nonlinear and time-varying and the precise dynamic model
is difficult to obtain, an FPGA-based ENN control system
has been proposed to control the position of the mov-
ing table of the LUSM to achieve high-accuracy position
control in this study. First, the LUSM drive system was
introduced. Then, the network structure and online learn-
ing algorithm of the proposed ENN control system were
described in detail. According to the online learning algo-
rithm, the connective weights of the ENN are trained on-
line using hardware circuits to obtain the control effort. In
addition, no constrained conditions and prior knowledge
of the controlled system are required in the design process.

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 112 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 56, no. 1, January 2009

Fig. 11. Experimental results of ENN control system for periodical trapezoidal reference trajectory at parameter variation condition: (a) tracking
2 1
response, (b) control effort, (c) tracking error, (d) connected weight W 51 between hidden layer and output layer, (e) connected weight W 15 between
1
input layer and hidden layer, and (f) connected weight W 25 between input layer and hidden layer.

Furthermore, to verify the effectiveness of the proposed

TABLE I. Performance Measures at Nominal Condition. control scheme, the ENN control system was implemented
Controller and Trajectory (μm) in a FPGA-based control system. Finally, the effectiveness
RBFN ENN of the proposed low-cost high-performance FPGA-based
Tracking Errors (μm) Sinusoid Trapezoidal Sinusoid Trapezoidal LUSM drive has been confirmed by some experimental
results.
Maximum 264.58 162.50 43.75 73.75
Average 3.19 6.17 2.82 1.78
Standard Deviation 32.11 32.84 12.62 12.46
References

[1] T. Sashida and T. Kenjo, An Introduction to Ultrasonic Motors.

TABLE II. Performance Measures at Parameter Variation Oxford: Clarendon Press, 1993.
Condition. [2] M. Zhu, “Contact analysis and mathematical modeling of traveling
wave ultrasonic motors,” IEEE Trans. Ultrason. Ferroelectr. Freq.
Controller and Trajectory (μm) Control, vol. 51, no. 6, pp. 668–679, 2004.
RBFN ENN [3] J. Zumeris, Ceramic Motor, U.S. Patent 5777423, July 7, 1998.
[4] F. J. Lin, R. J. Wai, and M. P. Chen, “Wavelet neural network
Tracking Errors (μm) Sinusoid Trapezoidal Sinusoid Trapezoidal control for linear ultrasonic motor drive via adaptive sliding-mode
Maximum 264.58 227.08 50.00 45.83 technique,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 50,
no. 6, pp. 686–698, 2003.
Average 5.15 4.76 3.23 1.08
[5] F. J. Lin and P. H. Shieh, “Recurrent RBFN-based fuzzy neural
Standard Deviation 31.43 35.27 15.53 14.75 network control for X-Y-Theta motion control stage using linear

Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
 Lin and Hung : fpga-based elman neural network control system for linear ultrasonic motor
ultrasonic motors,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control,
vol. 53, no. 12, pp. 2450–2464, 2006.
[6] Y. F. Peng and C. M. Lin, “Intelligent motion control of linear ul-
trasonic motor with H/spl infin/tracking performance,” IET Control
Theory Appl., vol. 1, no. 1, pp. 9–17, Jan. 2007.
[7] J. S. R. Jang, C. T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft
Computing: A Computational Approach to Learning and Machine
Intelligence. Upper Saddle River, NJ: Prentice-Hall, 1997.
[8] J. Elman, “Finding structure in time,” Cogn. Sci., vol. 2, no. 14, pp.
179–211, 1990.
[9] X. Li, G. Chen, Z. Chen, and Z. Yuan, “Chaotifying linear Elman
networks,” IEEE Trans. Neural Netw., vol. 13, no. 5, pp. 1193–1199,
2002.
[10] T. Sawaragi and T. Kudoh, “Self-reflective segmentation of human
bodily motions using recurrent neural networks,” IEEE Trans. Ind.
Electron., vol. 50, no. 5, pp. 903–911, 2003.
[11] M. Wlas, Z. Krzeminski, J. Guzinski, H. Abu-Rub, and H. A. Toli-
yat, “Artificial-neural- network-based sensorless nonlinear control of
induction motors,” IEEE Trans. Energ. Convers., vol. 20, no. 3, pp.
520–528, 2005.
[12] F. Xia, S. Li, and Y. Sun, “Neural network based feedback scheduler
for networked control system with flexible workload,” in Proc. Ad-
vances in Natural Computation, vol. 3611, pp. 242–251, 2005.
[13] F. J. Lin, H. J. Shieh, P. H. Shieh, and P. H. Shen, “An adaptive
recurrent-neural-network motion controller for X-Y table in CNC
machine,” IEEE Trans. Sys. Man. and Cybernetics Pt. B, vol. 36, no.
2, pp. 286–299, 2006.
[14] X. D. Li, J. K. L. Ho, and T. W. S. Chow, “Approximation of dy-
namical time-variant systems by continuous-time recurrent neural
networks,” IEEE Trans. Circuits Syst., II Express Briefs, vol. 52, no.
10, pp. 656–660, Oct. 2005.
[15] S. C. Kremer, “On the computational power of Elman-style recur-
rent networks,” IEEE Trans. Neural Networks, vol. 6, no. 4, pp.
1000–1004, July 1995.
[16] L. P. Douglas, VHDL: Programming By Example, 4th Ed. Columbus,
OH: McGraw-Hill, 2002.
[17] H. C. Roth, Circuit Design with VHDL. Cambridge, MA: MIT Press,
2004.
[18] F. J. Lin, R. J. Wai, K. K. Shyu, and T. M. Liu, “Recurrent fuzzy
neural network control for piezoelectric ceramic linear ultrasonic
motor drive,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol.
48, no. 4, pp. 900–913, 2001.
[19] F. J. Lin, R. J. Wai, and C. C. Lee, “Fuzzy neural network position
controller for ultrasonic motor drive using push-pull DC-DC con-
verter,” IET Proc. Control Theory Appl., vol. 146, no. l, pp. 99–107,
1999.
[20] S. Seshagiri and H. K. Khail, “Output feedback control of nonlinear
systems using RBF neural networks,” IEEE Trans. Neural Netw.,
vol. 11, no. 1, pp. 69–79, Jan. 2000.
[21] M. J. Lee and Y. K. Choi, “An adaptive neurocontroller using RBFN
for robot manipulators,” IEEE Trans. Ind. Electron., vol. 51, no. 3,
pp. 711–717, June 2004.
[22] F. J. Lin, L. T. Teng, P. H. Shieh, and Y. F. Li, “Intelligent con-
trolled-wind-turbine emulator and induction-generator system using
Authorized licensed use limited to: St. Xavier's Catholic College of Engineering. Downloaded on March 12, 2009 at 00:33 from IEEE Xplore. Restrictions apply.
113
RBFN,” IEE Proc. Electr. Power Appl., vol. 153, no. 4, pp. 608–618,
2006.
Faa-Jeng Lin (M’93–SM’99) received the B.S.
and M.S. degrees in electrical engineering from
the National Cheng Kung University, Tainan, Tai-
wan, and the Ph.D. degree in electrical engineer-
ing from the National Tsing Hua University, Hsin-
chu, Taiwan, in 1983, 1985, and 1993, respectively.
From 1993 to 2001, he was an associate professor
and then a professor in the Department of Electri-
cal Engineering, Chung Yuan Christian Universi-
ty, Chungli, Taiwan. From 2001 to 2003, he was
chairperson and professor in the Department of
Electrical Engineering, National Dong Hwa University, Hualien, Taiwan.
He was dean of research and development from 2003 to 2005 and dean of
academic affairs from 2006 to 2007 at the same university. Currently, he
is professor in the Department of Electrical Engineering, National Cen-
tral University, Chungli, Taiwan. In 2007, he was appointed the chair,
Power Engineering Division, National Science Council, Taiwan, a posi-
tion he will hold through 2009. His research interests include AC and
ultrasonic motor drives, DSP-based computer control systems, fuzzy and
neural network control theories, nonlinear control theories, power elec-
tronics, and micro mechatronics.
Dr. Lin received the Outstanding Research Professor Award from
the Chung Yuan Christian University, Taiwan, in 2000; the Crompton
Premium Best Paper Award from the Institution of Electrical Engineers
(IEE), United Kingdom, in 2002; the Outstanding Research Award from
the National Science Council, Taiwan, in 2004; the Outstanding Research
Professor Award from the National Dong Hwa University, Taiwan, in
2004; and the Outstanding Professor of Electrical Engineering Award
in 2005 from the Chinese Electrical Engineering Association, Taiwan.
Moreover, he was the recipient of the Distinguished Professor Award
from the National Central University in 2007. Furthermore, he is Fellow
of the Institution of Engineering and Technology (IET, former IEE).
Ying-Chih Hung was born in Taipei, Taiwan, in
1984. He received the B.S. degree in electrical en-
gineering from the Chung Yuan Christian Univer-
sity, Chungli, Taiwan, and M.S. degree in electri-
cal engineering from the National Dong Hwa
University, Hualien, Taiwan in 2006 and 2008, re-
spectively. He is currently working toward the
Ph.D. degree in electrical engineering at the Na-
tional Central University, Chungli, Taiwan. His
research interests include field-programmable gate
array, linear ultrasonic motor, motion control, and
intelligence control theories.