Neural Comput & Applic

DOI 10.1007/s00521-015-2098-2

ORIGINAL ARTICLE

Adaptive control of MEMS gyroscope using fully tuned RBF

neural network
Juntao Fei1 • Dan Wu1

Received: 24 November 2014 / Accepted: 4 November 2015

 The Natural Computing Applications Forum 2015

Abstract In this paper, a novel adaptive control 1 Introduction

scheme that incorporates fully tuned radial basis function
(RBF) neural network is proposed for the control of MEMS In recent years, MEMS gyroscopes become more and more
gyroscope with respect to external disturbances and model popular for measuring angular velocity due to their proper-
uncertainties. An adaptive fully tuned RBF neural network ties of low cost, compact size and high sensitivity. However,
controller is used to compensate the effect of external existing fabrication imperfections always result in some
disturbances and model uncertainties, thus improving the damping effects and stiffness and deteriorate the perfor-
dynamic characteristics and robustness of the MEMS mance of the control system. Advanced control methods
gyroscope. The fully tuned RBF neural network compen- seem to be effective way to compensate fabrication imper-
sating controller and the adaptive nominal controller are fections and model uncertainties and improve robustness in
combined in the unified Lyapunov framework to ensure the the presence of external disturbances. Leland [1] described
stability of the control system. By using the proposed two adaptive controllers for a MEMS gyroscope. An adap-
scheme, not only the effect of model uncertainties and tive add-on control algorithm for the conventional mode of
external disturbances can be eliminated, but also satisfac- operation of MEMS z-axis gyroscopes is presented in [2].
tory dynamic characteristics and strong robustness can be Cuong [3] designed a pre-shaped open-loop driving actua-
obtained. Simulation studies are implemented to verify the tion waveform to reduce the bouncing effect while main-
effectiveness of the proposed scheme and demonstrate that taining fast switching of a microelectromechanical system
the fully tuned RBF network control has better robustness contact switch. Farivar [4] proposed modified projective
and dynamic characteristics than traditional RBF network synchronization methods for underactuated unknown heavy
control. symmetric chaotic gyroscope systems via optimal Gaussian
radial basis adaptive variable structure control. Adaptive
Keywords Adaptive control  Fully tuned radial basis controller and adaptive fuzzy controller have been proposed
function  Lyapunov framework  MEMS gyroscope for MEMS gyroscopes in [5, 6]. Recently, because of the
powerful property of neural network to approximate any
nonlinear function, the RBF neural network has been widely
used in the control of time-varying and nonlinear systems to
compensate the effect of time-varying parameters and
external uncertainties. Young et al. [7] presented an online
retrainable RBF network to control the coiling temperature
for a hot coil strip at the iron and steel company. An adaptive
inverse model control system is designed for electronic
& Juntao Fei throttle, and two RBF neural networks are utilized [8]. An
jtfei@[Link]
RBF neural network control is used for the optimum control
1
College of Mechanical and Electrical Engineering, Hohai of a cold storage in [9]. An adaptive output feedback control
University, Changzhou 213022, China scheme incorporating RBF neural network to adaptively

123
 Neural Comput & Applic

compensate for nonlinearities is presented for the output 2. The updating laws for weights, centers and widths of
tracking of a class of continuous-time nonlinear plants in the fully tuned RBF network are designed under the
[10]. Behera [11] investigated the application of inversion of Lyapunov framework, thus ensuring the stability and
a RBF network to nonlinear control problems for which the convergence of the control system. Comparing the
structure of the nonlinearity is unknown. The RBF network- proposed control scheme with the traditional RBF
based control methods have been studied to control the robot neural network control, the fully tuned RBF network
manipulators [12, 13]. Meanwhile, people extended the fully controller scheme has three adjustable parameters;
tuned RBF network control scheme to many fields [14–16]. therefore, the tracking error using the proposed
The RBF network control method has been used to control scheme converges to zero more quickly than the
MEMS gyroscopes [17], but the fully tuned RBF network exiting one and the control system behaves with better
control has never been used in this field. Hence, it is nec- dynamic performance and robustness.
essary to incorporate the fully tuned RBF network control
The paper is organized as follows. In Sect. 2, dynamics
into the control of MEMS gyroscopes to get satisfactory
of MEMS gyroscope is derived. In Sect. 3, a design of the
tracking performance and strong robustness with respect to
adaptive control based on the nominal model is proposed.
model uncertainties and external disturbances.
In Sect. 4, a fully tuned RBF network compensation is
In this paper, a novel adaptive fully tuned RBF network
proposed and stability of the system is discussed in detail.
control scheme which has not been applied to MEMS
In Sect. 5, a simulation using the proposed scheme is
gyroscopes is proposed. The novelty of this paper is that it
implemented and a comparative study between the adap-
firstly extends the fully tuned RBF network scheme to the
tive RBF network control and the adaptive fully tuned RBF
control of MEMS gyroscopes. The integration of the
network control is performed.
adaptive nominal control and nonlinear approximation of
fully tuned RBF network control could compensate for the
effects of model uncertainties and improve the tracking 2 Dynamics of MEMS gyroscope
performance and robustness of the system in the presence
of model uncertainties and external disturbances. Using The dynamics of MEMS gyroscope is described in this
Lyapunov analysis, the asymptotically stability of the section. Dynamics of a MEMS gyroscope is derived from
MEMS gyroscope can be guaranteed and the tracking Newton’s law in the rotating frame.
errors between the reference output and the plant output In a z-axis gyroscope, by supposing the stiffness of
can converge to zero asymptotically. spring in z direction much larger than that in the x, y di-
The paper designs a fully tuned RBF network-based rections, motion of poof mass is constrained to only along
adaptive control scheme to compensate the model uncer- the x–y plane as shown in Fig. 1. Assuming that the mea-
tainties and external disturbance and improve the tracking sured angular velocity is almost constant over a long
performance of the system. Due to the powerful property of enough time interval, the equation of motion of a gyro-
approximating nonlinearities, the fully tuned RBF network scope is simplified as follows.
compensation is able to update the estimates of the control Taking fabrication imperfections, which cause extra
system online and improve the dynamic characteristics. In coupling between x- and y-axes, and external disturbances
addition, to verify the effectiveness of the proposed control into account, the governing equation for MEMS gyroscope
scheme, we compare it with the adaptive neural compensa- is:
tion scheme, thus evaluating the system performance such as
the convergence, stability and the robustness. The advan- Sensing Axis
tages of the proposed control methods are listed as follows: y
Y
1. Utilizing the powerful approximation property of RBF
kyy
neural network, a fully tuned RBF network compen- O X
sation is incorporated into adaptive nominal control dxx m k xx Driving Axis
scheme to compensate nonlinear functions and obtain
x
satisfactory transient characteristics and strong robust-
ness against model uncertainties and external distur-
bances. After combining nominal control and Ωz
dyy
nonlinear approximation of fully tuned RBF network
control, the convergence of the tracking errors can be
Supporting Frame
guaranteed and the modeling error can be reduced
effectively. Fig. 1 Schematic model of a z-axis MEMS gyroscope

123
Neural Comput & Applic

x þ dxx x_ þ dxy y_ þ kxx x þ kxy y ¼ ux þ 2mXz y_

m€ matrices which can ensure the stability of the system and
ð1Þ improve the robustness against model uncertainties and
y þ dxy x_ þ dyy y_ þ kxy x þ kyy y ¼ uy  2mXz x_
m€
external disturbances by choosing suitable values. Here we
 2   
In Eq. (1), dxx and dyy are damping; kxx and kyy are a 0 2a 0
choose kv and kp as kp ¼ and kv ¼ to
spring coefficients; dxy and kxy called quadrature errors are 0 a2 0 2a
coupled damping and spring terms, respectively, due in obtain satisfactory dynamic performance and strong
large part to the asymmetries in suspension structure and robustness.
misalignment of sensors and actuators. However, it is hard for us to get the precise MEMS
Dividing both sides of Eq. 1 by m, q0 and x20, which are gyroscope model without external disturbances. In prac-
a reference mass, length and natural or resonant frequency, tice, we can only obtain the nominal parameters of the
respectively, and considering lumped external distur- MEMS gyroscope. Define the nominal parameters D, Kb, X
bances, we get the form of the nondimensional equation of as D0, K0, X0, and then, the control is designed as:
motion as u ¼ q€d  kv e_  kp e þ D0 q_ þ K0 q þ 2X0 q_ ð8Þ
x€ þ dxx x_ þ dxy y_ þ x2x x þ xxy y ¼ ux þ 2Xz y_
ð2Þ we can get
y€ þ dxy x_ þ dyy y_ þ xxy x þ x2y y ¼ uy  2Xz x_
dxy dyy
q€ þ Dq_ þ Kq ¼ q€d  kv e_  kp e þ D0 q_ þ K0 q þ 2X0 q_
dxx
where ! dxx , ! dxy , ! dyy ,  2Xq_ þ d ð9Þ
mx
qffiffiffiffiffiffi
ffi0 qffiffiffiffiffiffiffi mx0 mx0
Xz kxx k k
x0 ! Xz , mx2 ! xx , mxyy2 ! xy , mxxy2 ! xxy ~ ¼ D0  D; K~ ¼ K0  K; X ~ ¼ X0  X,
0 0 0 Denoting D
The vector form of MEMS gyroscope dynamics model the closed-loop system in Eq. (9) can be rewritten as:
can be written as ~q_ þ Kq ~ q_ þ d
~ þ 2X
e€ þ kv e_ þ kp e ¼ D ð10Þ
q€ þ Dq_ þ Kb q ¼ u  2Xq_ þ d ð3Þ
      We can see from Eq. (10) that the scheme can still
x u 0 Xz ensure the stability of the system, but the dynamic char-
where q ¼ , u¼ x , X¼ , D¼
y uy Xz 0 acteristic and the robustness will be degraded by the
   2   
dxx dxy xx xxy d imperfect modeling. Therefore, it is necessary for us to
, Kb ¼ , d ¼ 1 represents an
dxy dyy xxy x2y d2 design a compensating controller to improve the tracking
uncertain extraneous disturbance, and d1 and d2 are the performance and guarantee the stability of the control
disturbances in the x-axis and y-axis, respectively. system with respect to model uncertainties and external
disturbances.
The modeling error is designed as:
3 Adaptive controller
~q_ þ Kq
f ¼D ~q_ þ d
~ þ 2X ð11Þ
The objective of the proposed scheme is to keep the proof Then, we modify the control law as
mass to track a designed reference trajectory generated by
an ideal oscillation: u ¼ q€d  kv e_  kp e þ D0 q_ þ K0 q þ 2X0 q_  f ð12Þ
xd ¼ A1 cosðx1 tÞ; yd ¼ A2 cosðx2 tÞ ð4Þ Hence, the modified computed force controller leads to
The reference model can be redefined in vector terms as: the closed-loop system described in (7) and minimize the
modeling uncertainty. However, we do not know the
q€d þ kd qd ¼ 0 ð5Þ modeling error item f ð  Þ so that the modified scheme in
where kd ¼ diagfw21 ; w22 g; qd ¼ ½ xd yd T . (12) cannot be taken in practice. Therefore, it is necessary
Suppose that there are no model uncertainties and for us to design a satisfied estimated function f^ðÞ of f ðÞ
external disturbances, i.e., d = 0, then the parameters of to improve the dynamic characteristic of the MEMS
the MEMS gyroscope can be used to calculate the elec- gyroscope.
trostatic drive force of proof mass to obtain ideal tracking Equation (10) can be rewritten in state space form as
performance. The control law is designed as: follows:
u ¼ q€d  kv e_  kp e þ Dq_ þ Kb q þ 2Xq_ ð6Þ x_ ¼ Ax þ Bf       ð13Þ
e 0 I 0
so that the closed-loop system becomes as follows: where x ¼ ; A¼ ; B¼ ; x2
e_ kp kv I
e€ þ kv e_ þ kp e ¼ 0 ð7Þ R41 ; A 2 R44 ; B 2 R42 ; I is an identity matrix.
where qd is the desired reference trajectory and e = q - qd Because kv, kp and I are constant, A and B must be also
is the tracking error. kv and kp are positive definite constant constant. From Eq. (11), it is obviously seen that f can be

123
 Neural Comput & Applic

identified through x. Therefore, we can denote f(•) as f(x).

Due to the powerful property of approximating nonlinear
Output
functions, the fully tuned RBF network controller will be y1 y2
layer
applied to estimate the modeling error item in Sect. 4.
W
4 Fully tuned RBF network control Hidden
1 2 3 4 5
layer

The block diagram of the adaptive nominal controller using

fully tuned RBF network for a MEMS gyroscope is shown
in Fig. 2.
The most useful characteristic of RBF networks is their x11 x2 x3 x4 Input
ability to approximate nonlinear function to represent layer
complex plant through learning. Due to above properties,
neural network controllers have been used in many fields Fig. 3 Architecture of RBF network
for the compensation for the external disturbances, model  
nonlinearity and system uncertainties. Therefore, we apply maxf^ðx; h Þ  f ðxÞ  e0
the fully tuned RBF network compensation to the control
of the MEMS gyroscope. Applying the fully tuned RBF network, we use f^ðx; hÞ to

The architecture of RBF network is shown in Fig. 3. We parameterize the unknown function f(x), where h 2 Rn 2
*
can see that the RBF network has a three-layer feedback and n is the number of hidden nodes.
network. The output of the neural network can be written as: Let us rewrite Eq. (13) as
n
X x_ ¼ Ax þ Bff^ðx; h Þ þ ½f ðxÞ  f^ðx; h Þg ð14Þ
yi ¼ hTij /j ; i ¼ 1; 2; . . .; L
i¼1 where h* is the optimal weight value.
  2  The modeling error is designed as g after using the fully
/j ðxÞ ¼ exp x  cj  =2  b2j ; j ¼ 1; 2; . . .; n
tuned RBF network compensation:
where L and n*are the numbers of output nodes and hidden g ¼ f ðxÞ  f ðx; h Þ ð15Þ
node. hij is the weight from the jth hidden node to the ith
The upper bound of g is g0, where
output node, and cj denotes the jth center vector. bj denotes
the width of the jth hidden node. In most times, we choose g0 ¼ supkf ðxÞ  f ðx; h Þk ð16Þ
Gaussian function g(a) = exp (-a2) as the activation
then rewritten (14) as:
function g.
For the convenience of applying the fully tuned RBF x_ ¼ Ax þ B½f^ðx; h Þ þ g ð17Þ
network to the MEMS gyroscope, the following assump-
Applying the fully tuned RBF network control to
tion is made.
Given an arbitrary positive constants e0, there exist a MEMS gyroscope, the function f^ðx; h Þ can be written as:
positive integer n* and a weight vector h = h* so that the f^ðx; h Þ ¼ hT / ðxÞ ð18Þ
output of f^ðx; h Þ satisfies following equation:
where h* is the n* 9 2 optimal weight matrix and /*
Mechanical and
determined by the ideal widths b* and centers c* is a
RBF-NN environmental n* 9 1 vector of Gaussian functions.
noise Equation (17) can be rewritten as:
d 
+ u2 + x_ ¼ Ax þ B hT / þ g ð19Þ
qd e u1 u
Reference Nominal MEMS
Then, after incorporating the fully tuned RBF network
model - + controller
+ + gyroscope
control into the nominal control, we redesign as follows:
u ¼ u1 þ u2 ð20Þ
Signal with
detection
u1 ¼ q€d  kv e_  kp e þ D0 q_ þ K0 q þ 2X0 q_ ð21Þ
Fig. 2 Block diagram of the adaptive nominal control scheme based
on the fully tuned RBF network for a MEMS gyroscope
^
u2 ¼ h^T /ðxÞ ð22Þ

123
Neural Comput & Applic

h^ is the estimator of the optimal weights h*, /

^ instead of ^_ ¼ K1 hB
~_ ¼ /
/ ^ T Px ð30Þ
^ * ^ c^ are
/ðb; c^; xÞ is the estimator of the optimal / where b;
the estimated width and center. In the fully tuned RBF Rewriting Eqs. (29) and (30) in a discrete-time frame-
network control method, h, b and c are considered to be work, the updating laws for weights, centers and widths of
adjustable parameters different from the traditional RBF the fully tuned RBF network become:
network control method whose b and c are fixed values. o^
g
h^i ðn þ 1Þ ¼ h^i ðnÞ  g1 T xT ðnÞPðnÞBðnÞ ð31Þ
The new control leads to the following closed-loop ohi
system:
h i ^
o/ðnÞ
c^i ðn þ 1Þ ¼ c^i ðnÞ þ g2 ^ T PðnÞxðnÞ
hB ð32Þ
x_ ¼ Ax þ B hT / ð xÞ þ g  h^T /^ð x Þ ð23Þ cTi
o^

Consider the following Lyapunov function ^

o/ðnÞ
b^i ðn þ 1Þ ¼ b^i ðnÞ þ g3 ^ T PðnÞxðnÞ
hB ð33Þ
1 1 ~ ~T 1 ~T ~ ^T
obi
V ¼ xT Px þ trðhc h Þþ / K/ ð24Þ
2 2 2 where g1, g2 and g3 are positive design constants.
where c is a positive number and K and P are symmetric Equation (28) becomes
and positive definite matrixes. P satisfies the following 1
equation V_ ¼  xT Qx þ gT BT Px ð34Þ
2
PA þ AT P ¼ Q ð25Þ Since kgT k  g0 ; kBT Pk  kPk, we can get
where Q is positive definite matrix. 1
V_   kmin ðQÞk xk2 þg0 kmax ðPÞkxk
Denoting h~ ¼ h^  h ; /~¼/ ^  / , the derivative of 2 ð35Þ
Vis given by: 1
¼  kxk½kmin ðQÞk xk  2g0 kmax ðPÞ
1 1 2
V_ ¼ xT Px_ þ x_T Px þ trðhc~_ h~T Þ þ /
~_ T K/
~
2 2 where kmin(Q) and kmax(P) are the minimum eigenvalue of
1 matrix Q and the maximum eigenvalue of matrix P. In
^T ðxÞ
¼  xT Qx þ gT BT Px þ xT PBhT /T ðxÞ  xT PBh^T / 2kmax ðPÞ
2 order to make v_  0, kmin ðxÞ k xk g 0 is required.
~_ h~T Þ þ /
þ trðhc ~_ T K/
~ Therefore, from Eq. (35), we can know that the tracking
1 error matrix x and all parameters are bounded and x will
¼  xT Qx þ gT BT Px þ xT PBðh^T / ^  h^T /
~  h~T /
^ þ h~T /Þ
~
2 fall inside a residue of radius R0 = bg0 where
 xT PBh^T / ~_ h~T Þ þ /
^T ðxÞ þ trðhc ~_ T K/
~ b = 2kmax(P)/kmin(Q). Only when g0 is zero, the output
tracking error between plant output and reference output
1 ~  xT PBh~T /
¼  xT Qx þ gT BT Px  xT PBh^T / ^ converges to zero.
2
~_ h~T Þ þ /
þ trðhc ~_ T K/
~
1 ~_ T K  xT PBh^T Þ/
~
¼  xT Qx þ gT BT Px þ ð/ 5 Simulation study
2
~_ h~T Þ  xT PBh~T /
þ ½trðhc ^ ð26Þ To illustrate the effectiveness of the proposed adaptive fully
tuned RBF network compensation scheme, a simulation
where PA ? ATP = -Q.
example with application to the MEMS gyroscope is con-
Noting in Eq. (26),
ducted. The control target is to make the mass proof track the
^ ¼ trðxT PBh~T /Þ
xT PBh~T / ^ ¼ trð/x
^ T PBh~T Þ ð27Þ desired reference trajectory and utilize the fully tuned RBF
network to estimate and compensate the unknown parameter
Equation (26) becomes
uncertainties and external disturbances. The parameters of
1 ~_ T K  xT PBh^T Þ/
~ the MEMS gyroscope sensors are as follows:
V_ ¼  xT Qx þ gT BT Px þ ð/
2 ð28Þ m ¼ 1:8  107 kg; kxx ¼ 63:955 N=m; kyy ¼ 95:92 N=m;
~_  /x
þ tr½ðhc ^ T PBÞh~T 
kxy ¼ 12:779 N=m dxx ¼ 1:8  106 N s=m;
If the parameter update laws of h; ~_ are selected as:
~_ / dyy ¼ 1:8  106 N s=m; dxy ¼ 3:6  107 N s=m
_ _
h~ ¼ h^ ¼ c1 /x
^ T PB ð29Þ The unknown angular velocity is set to be Xz = 100 -
rad/s. Selecting 1 lm as the reference length and 1 Khz as

123
 Neural Comput & Applic

Fig. 4 Tracking errors of each Tracking Errors

2

x-axis tracking error

axis
1

-1

-2
0 10 20 30 40 50 60 70 80 90 100

y-axis tracking error

1

-1

-2
0 10 20 30 40 50 60 70 80 90 100
Time(s)

Position Tracking for X Axis Approximation of X Axis

2 80

1.5 60
position tracking for x-axis

1 40

0.5 20
f1 and fn1

0 0

-0.5 -20

-1 -40

-1.5 -60

-2 -80
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time(s) time(s)

Fig. 5 Position tracking of x-axis Fig. 7 Fully tuned RBF network approximation of x-axis

Position Tracking for Y Axis the reference natural frequency of each axis of the MEMS
1.5
gyroscope, the nondimensional parameters are listed as
follows:
1
w2x ¼ 355:3; w2y ¼ 532:9; wxy ¼ 70:99; dxx ¼ 0:01;
position tracking for y-axis

0.5 dyy ¼ 0:01; dxy ¼ 0:002; Xz ¼ 0:1

0 The desired reference trajectories are designed as qdx =

cos (6.17t) and qdy = cos (5.11t). The nominal values of D, K
-0.5 and X are set to be D0 = 0.9 9 D, K0 = 0.9 9 K, X0 = 0.
The a in kv and kp is chosen as a = 2. The g1, g2 and g3 in
-1
Eqs. (31–33) are chosen as g1 = g2 = g3 = 0.03. The initial
value of width is set to zero, and each element of the initial
center is set to two. Assuming the fully tuned RBF network has
-1.5
0 10 20 30 40 50 60 70 80 90 100 11 nodes, the initial value of the weight vector of Gaussian
time(s)
function is chosen as
Fig. 6 Position tracking of y-axis

123
Neural Comput & Applic

½1:0  0:8  0:6  0:4  0:2 0 0:2 0:4 0:6 0:8 1:0T Fig. 4, it can be seen that the tracking errors all con-
verge to zero correctly and quickly. Incorporating the
We assume Q = 10 9 I, and all the initial conditions of fully tuned RBF network for compensation for model
the MEMS gyroscope are chosen as zero. The external uncertainties, the unknown modeling error f(x) can be
disturbance is taken as d ¼ ½5  randn(1, 1); 5 approximated by f^ðxÞ effectively. In Fig. 9, we can see
randn(1, 1)]. that the tracking errors converge to zero more quickly
Therefore, it can be seen that ideal tracking perfor- using the fully tuned RBF network control scheme than
mance and strong robustness have been obtained. Fig- the traditional RBF network control. The proposed
ure 4 shows the tracking errors. Figures 5 and 6 show adaptive fully tuned RBF network control scheme be-
the proof mass of each axis tracking trajectories. Fig- haves satisfactory dynamic characteristics and strong
ures 7 and 8 depict the approximation of fully tuned robustness in the presence of model uncertainties and
RBF network to modeling error f(x). Figure 9 compares external disturbances.
the tracking errors between fully tuned RBF network
control and the traditional RBF network control. From

6 Conclusion
Approximation of Y Axis
80
In this paper, an adaptive nominal control based on fully
60
tuned RBF network with application to MEMS gyroscopes
40 is proposed. Utilizing the powerful ability of approxima-
tion of nonlinear systems, the fully tuned RBF network is
20
applied to the adaptive control scheme based on nominal
f2 and fn2

0 model to eliminate the effect of model uncertainties and

-20
external disturbances. The updating laws of the fully tuned
RBF network are derived based on the Lyapunov theory,
-40 thus ensuring the stability of the MEMS gyroscope. In
-60 contrast with the traditional RBF network control, the fully
tuned RBF network control has better performance such as
-80
dynamic characteristics and robustness. Simulation results
-100 show that the proposed control scheme effectively
0 10 20 30 40 50 60 70 80 90 100
time(s) improves the tracking performance of the overall system
and the robustness against model uncertainties and external
Fig. 8 Fully tuned RBF network approximation of y-axis disturbances.

Tracking Errors
2
x-axis tracking error

tracking error of x-axis using RBF NN compensation

tracking error of x-axis using RBF neural compensation
1

-1

-2
0 10 20 30 40 50 60 70 80 90 100

2
y-axis tracking error

tracking error of y-axis using RBF NN compensation

tracking error of y-axis using RBF neural compensation
1

-1

-2
0 10 20 30 40 50 60 70 80 90 100
Time(s)

Fig. 9 Comparison of tracking errors between the fully tuned RBF control (blue line) and the traditional RBF network control (dash line) (color
figure online)

123
 Neural Comput & Applic

Acknowledgments The authors thank the anonymous reviewers for 7. Young J (2002) Improving lookup table control of a hot coil strip
their useful comments that improved the quality of the paper. This process with online retrainable RBF network. IEEE Trans Ind
work is partially supported by National Science Foundation of China Electron 47(3):679–686
under Grant No. 61374100; Natural Science Foundation of Jiangsu 8. Fang Y (2009) RBF networks-based adaptive inverse model
Province under Grant No. BK20131136; the University Graduate control system for electronic throttle. IEEE Trans Control Syst
Research and Innovation Projects of Jiangsu Province under Grant Technol 18(3):750–756
No. CXLX12_0235; and the Fundamental Research Funds for the 9. Dong S (2000) Using RBF neural network for optimum control of
Central Universities under Grant No. 2014B32814. a cold storage. J Syst Eng Electron 11(4):30–36
10. Seshagiri S (2000) Output feedback control of nonlinear systems
using RBF neural networks. IEEE Trans Neural Netw
11(1):69–79
References 11. Behera L (1995) Inversion of RBF networks and applications to
adaptive control of nonlinear systems. IEE Proc Control Theory
1. Leland P (2006) Adaptive control of a MEMS gyroscope using Appl 142(6):617–624
Lyapunov methods. IEEE Trans Control Syst Technol 12. Huan W (2003) Study on the robot robust adaptive control based
14(2):278–283 on neural networks. J Syst Eng Electron 14(4):55–58
2. Park S, Horowitz R (2003) Adaptive control for the conventional 13. Feng G (1997) A new stable tracking control scheme for robotic
mode of operation of MEMS gyroscopes. J Microelectromech manipulators. IEEE Trans Syst Man Cybern Part B: Cybern
Syst 12(1):101–108 27(3):510–516
3. Cuong D (2012) Energy-based approach to adaptive pulse shap- 14. Zhou L (2010) Adaptive integral dynamic surface control based
ing for control of RF-MEMS DC-contact switches. J Microelec- on fully tuned radial basis function neural network. J Syst Eng
tromech Syst 21(6):1382–1391 Electron 21(6):1072–1078
4. Farivar F (2013) Synchronization of underactuated unknown 15. An H (2007) Remote network controller design based on fully
heavy symmetric chaotic gyroscopes via optimal Gaussian radial tuned RBF neural network. Int Conf Nat Comput 2:24–27
basis adaptive variable structure control. IEEE Trans Control Syst 16. Corradimi L (2010) Discrete time variable structure control of
Technol 21(6):2374–2379 robotic manipulators based on fully tuned RBF neural networks.
5. Fei J, Zhou J (2012) Robust adaptive control of MEMS triaxial IEEE International Symposium on Industrial Electronics,
gyroscope using fuzzy compensator. IEEE Trans Syst Man pp 1840–1845
Cybern Part B 42(6):1599–1607 17. Fei J, Yang Y (2012) Adaptive neural compensation scheme for
6. Fei J (2010) Robust adaptive vibration tracking control for a robust tracking of MEMS gyroscope. IEEE International Con-
micro-electro-mechanical systems vibratory gyroscope with ference on Systems, Man, and Cybernetics, pp 1546–1551
bound estimation. IET Control Theory Appl 4(6):1019–1026

123