Research Journal of Applied Sciences, Engineering and Technology 6(7): 1233-1239, 2013
ISSN: 2040-7459; e-ISSN: 2040-7467
© Maxwell Scientific Organization, 2013
Submitted: October 30, 2012
Accepted: December 21, 2012
Published: July 05, 2013
Fuzzy Neural Network based RFID Positioning and Navigation Method for Mobile Robots
Bo-Wen Hong, Ying-Jeh Huang, Chu-Yung Chen, Ping-Chou Wu and Wei-Chung Chen
Department of Electrical Engineering, Yuan Ze University, Chungli, Taiwan
Abstract: This study proposes the Radio Frequency Identification (RFID) indoor positioning and navigation method
based on fuzzy neural network. The proposed method is applied to a wheelchair home health care robot with
wireless communication. One reader and four tags are used. Based on the Received Signal Strength Indication
(RSSI) data, the position of the robot can be determined. Further, to overcome the measurement error problem due
to environmental parameter variation, a Fuzzy Neural Network (FNN) is proposed to compensate the measurement
data. The FNN automatically adjust the weight, the variance and the mean value to overcome effectively the
environmental parameter variation. A back-propagation algorithm is developed to achieve self-learning. The
successful experiment results show that the proposed system architecture and positioning system provide
satisfactory accuracy and make home health care wheelchair robot positioning system available for navigation and
guidance.
Keywords: Fuzzy neural network, indoor positioning, RFID, RSSI, wheelchair robot
INTRODUCTION
With the social and elderly population growth,
human medical science and health care technology have
gained much attention in recent years. The home care
robot is particularly focused on the wheelchair robot for
the aged people (Cooper et al., 2008). Consider an
intelligent wheelchair robot that combines manipulation
and mobility assistance with perception and decisionmaking. Some commonly used sensors include infrared
sensors, ultrasonic sensors, gyroscopes, accelerometers,
radio frequency identification systems, laser range
finder, camera and so on Vachhani and Sridharan
(2008), Misono et al. (2007), Celeste et al. (2008) and
Ho et al. (2011). Many researchers have worked on
automatic indoor positioning for years, for example,
RFID indoor positioning system many researchers have
worked on automatic indoor positioning for years, for
example, the RFID indoor positioning systems (Ho
et al., 2004).
Radio frequency identification indoor positioning
system has been made significant progress. The
hardware strategy is divided into active and passive.
The application of passive RFID technology is
extensive in various fields. Recently, the active RFID
technology has been applied to develop indoor
concerning the indoor positioning wheelchair robots.
These robots need a number of sensors for
positioning/navigation and result in complex system
architectures.
In this study, an RFID indoor positioning
technology for wheelchair home health care robots is
Fig. 1: The prototype of the IRW
proposed. Ultrasonic sensors are used to detect and
avoid obstacles while navigating the wheelchair robot.
With four tags and one reader, we measure the signal
strength and calculate the position of the wheelchair
robot. Further, to deal with the disturbed experimental
RFID data leading to incorrect localization, we further
develop the fuzzy neural network technology to
compensate the measurement RSSI data. With the selflearning functions, the FNN technology can be applied
to robotic applications.
Our study intends to support an on-going research
project in the Gerontechnology Research Center in
Yuan Ze University, Taiwan in developing an
Intelligent Robotic Wheelchair (IRW) as shown in
Fig. 1. The goal of this project is to redefine the
Corresponding Author: Bo-Wen Hong, Department of Electrical Engineering, Yuan Ze University, Chungli, Taiwan
1233
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
wheelchair as the center of mobility, everyday living
and healthcare for the senior users. In this study, the
proposed fuzzy neural network effectively overcomes
the environmental parameter variation and compensates
the uncertain RSSI value. Experimental result shows
that the proposed system architecture and localization
technique are indeed effective.
RFID INDOOR POSITIONING METHOD
The proposed positioning system consists of 2.4
GHz active RFID components which are four active
tags and one active reader. The signals transmitted from
four tags are collected by the reader setting on the
wheelchair robot. They are further weighted and
calculated in the world coordinates. Figure 2 shows the
experimental floor on it every 50 cm of distance being
one unit. The room size in experiment is 25 m2. There
are totally 81 measuring points.
Figure 3 is the schematic diagram of the RFID
location sensing. According to the measured RSSI
value from each tag to the reader, the distance between
each tag to the reader can be estimated. There are four
tags, so the distances are defined as the radii R α , R b ,
R c , and R d , respectively. Long dotted line, short dotted
line, point long dotted line and point short dotted line
represent respectively the converted distance of the four
tags according to the measured RSSI values. Moreover,
A-B intersections: W1 and W2, A-C intersection: P1
and P2, A-D intersection: Q1 and Q2, B-C intersection:
Fig. 2: The experimental environment (81 measuring points
on the floor)
S1 and S2, B-D intersection: U1 and U2, C-D
intersection: V1 and V2. These intersectional points can
be calculated by the following formulas:
A-B intersection W:

W (xa , y w ) ⇒

W (xb , y w ) ⇒
(xa − xw )2 + ( ya − y w )2 = Ra
(xb − xw )2 + ( yb − y w )2 = Rb
(1)
(x
(x
(2)
A-C intersection P:
 P (x , y ) ⇒
a
p

(
,
P
x
y

c
p )⇒
a
− x p ) 2 + ( y a − y p ) 2 = Ra
2
2
c − x p ) + ( y c − y p ) = Rc
A-D intersection Q:
W2
P2
Rb
RTb
B
V2
A
RTa
Q1
TTa
Ra
Q2
U1
S1
S2
Move
W1
P1
U2
D
C
RTc
RTd
Rd
Rc
V2
RFID Tag
A Level
A-B, A-C, A-D node
RFID Reader
B Level
B-C, B-D node
Tracking Target
C Level
C-D node
RFID Reader After Move
D Level
Fig. 3: Illustration of the operation of RFID localization
1234
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
 Q (x , y ) ⇒
a
q

(
,
Q
x
y

d
q )⇒
(x
(x
a
d
− xq ) 2 + ( ya − yq ) 2 = Ra
(3)
− xq ) 2 + ( yd − yq ) 2 = Rd
fading and the shadowing effect that lead to
measurement error. The maximum and minimum RSSI
values read by the reader are defined as R max and R min
and its scope is defined as:
B-C intersection S:
S (x , y ) ⇒
b
s

(
S
x
y
,
 c s ) ⇒
Rscope = Rmax − Rmin
(xb − xs ) + ( yb − ys ) = Rb
(xc − xs ) 2+ ( yc − ys ) 2 = Rc
2
2
(4)
Let w i be the weight between the reader and Tag i, i.e.:
wi =
B-D intersection U:

 U ( xb , yu ) ⇒


U ( xd , yu ) ⇒
(xb − xu ) 2+ ( yb − yu ) 2 = Rb
(xd − xu ) 2+ ( yd − yu ) 2 = Rd
Rscope
(5)
(
ReaderA = Ax , Ay
(xc − xv ) 2+ ( yc − yv ) 2 = Rc
(xd − xv ) 2+ ( yd − yv ) 2 = Rd
(6)
Consequently, W (x W , y W ), P (x P , y P ), Q (x Q , y Q ),
S (x S , y S ), U (x U , y U ), and V (x V , y V ), are obtained.
Applying the above algorithm, the exact location of the
wheelchair robot can be obtained.
The definition of coordinates for the RFID tag is
Tag A (A x , A y ). Each measured RSSI value recorded
from the reader to the associated tag is expressed as Tag
R i , i = 1, 2, …, 4,. Due to varying RSSI values in the
experimental environment considering the multi-path
)
n
 n

Ri _ x ⋅ wi2 ∑ Ri _ y ⋅ wi2
∑

, i =1
=  i =1
n
n

∑ wi2
∑ wi2

i =1
i =1

(
) (
Error = e x ,e y = Target x − Ax ,Target y -Ay
neto4 = ∑ wko4 xk4
output layer
k
4 4
w25
x25
w14 x14
yk3 = net k3
Layer3
rule layer
net 3 = x3 x3
k
1 j 2l
net 3 = x3 x3
k
2 j 1l
Layer2
(
yk2 = exp netk2
membership layer
input layer
)
(10)
In short, (8) and (9) are computed to get (10). The
FNN on-line tunes the weighting values iteratively.
The calculated coordinates are sent to the Field
Programmable Gate Array (FPGA). Then the FPGA
y14 = net14
Layer1


 (9)




where, R i_x and R i_y are the X and Y coordinates of ith
tag, w i is the weighting value between ith tag and the
reader. Suppose the target coordinates is e = (Target x ,
Target y ) and the error is Error = Target A -Tag A . Then:
Layer4
netij2 =
(8)
Rmax − ReaderRi
C-D intersection V:

 V ( xc , yv ) ⇒


V ( xd , yv ) ⇒
(7)
(x
y11
x11
2
ij
− mij
(σ ij )
)
2
2
y12
x12
Fig. 4: The structure of the fuzzy neural network
1235
)
i =1~ 2
j =1~ 5
k = 1 ~ 25
o =1
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
determines the necessary control commands for the
navigation and obstacle avoidance of the wheelchair
robot.
Next, to determine whether the FNN finishes
learning, define a cost function E as:
E=
FUZZY NEURAL NETWORK
The structure of the FNN is shown in Fig. 4. The
purpose here is to improve the accuracy of the
localization9. Define Layer 1, 2, 3 and 4 be the input
layer, the membership layer, the rule layer and the
output layer, respectively. Let the number of neurons be
f (x) = x. Define the X-axis error be e x and Y-axis error
be e y . The error difference 𝑒𝑒𝑥𝑥̇ and 𝑒𝑒𝑦𝑦̇ are selected to be
input variables. For every neuron in the input layer, the
net input and the net output can be represented as:
x11 = ec , x12 = ec , c = x, y
y1i
=
f i1
( )=
neti1
(11)
i = 1, 2
x1i ,
net
(x
=
2
i
− mij
(σ )
)
2
i = 1, 2 , j = 1,,5
(
)
yij2 = f ij2 netij2 = exp netij2 , i = 1, 2 , j = 1,,5, (14)
where, m ij and σ ij are the mean and the standard
deviation of the Gaussian function, respectively.
In the rule layer, each neuron multiplies the input
signals and sends out the product, i.e.:
net k3
=
x13j x23l ,
j , l = 1,...,5, k = 1,...,25
( )
y k3 = f k3 net k3 = net k3 , k = 1,...,25
(
=
( )
neto4
= neto4 ,
o = 1,
(20)
)
(21)
After the weights are updated, the means and
variances of the Gaussian functions in the membership
layer will be tuned with the outputs of the rule layer.
The error term to be propagated is given by:
δ k3 = −
(15)
∂E
∂net k3
δ12j = −
(16)
δ 22 j = −
(17)
∂E
∂net 2j
= δ14 wk
=
(22)
∑ δ k3 y13j
(23)
j
=
∑ δ k3 y23 j
(24)
j
(
)
(25)
)
(26)
2 xi2 − mij
∂E
= η mδ ij2
∂mij
σ ij 2
(
( )
2 xi2 − mij
∂E
∆σ ij = −ησ
= ησ δ ij2
∂σ ij
σ ij 3
(18)
where,
x4 k : The kth input to the neuron of Layer 4 and the
connecting weight
x4 ko : The output action strength of the oth output
associated with the kth rule.
∂E
∂net 2j
∆mij = −η m
k
f o4
)
∆wk = η w Y1 − y14 xk4
The overall output is the summation of all incoming
signals, i.e.:
yo4
(
 ∂E ∂y14 
 = Y1 − y14
 ∂y 4 ∂net 4 
1 
 1
δ14 = −
where,
x3 lj : The jth input to the neuron of Layer 3
The lth input to the neuron of Layer 3
x3 2l :
4 4 k = 1,...,25, o = 1,
net o4 = ∑ wko
xk ,
(19)
where,
w k = The weight between the rule layer and the output
layer
η w = The learning rate of the connected weight
(13)
2
)
2
In order to minimize the cost function, the gradient
decent method is adopted to tune those parameters and
weights. The learning rate affects the variation of the
parameters and the weights. The learning algorithm
based on the back-propagation method is represented
below. The error term to be propagated is given by:
(12)
ij
(
)
where,
Y i = The target
y i = The output of FNN controller
In the membership layer, each neuron performs a
membership function. The Gaussian function is adopted
to be the membership function. The net outputs are:
2
ij
(
1
Yi − yi4
2
( )
2
where,
m ij = The mean of the Gaussian function in the
membership layer
σ ij = The variance of the Gaussian function in the
membership layer
1236
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
η m = The learning rate of the mean
η σ = The learning rate of the variance
setup on the wheelchair robot. The space size of the
experiment is 5 m in length and width. Indoor signal
transmission quality affects positioning precision. In
fact, indoor signals could be easily interfered by signal
reflection, scattering, diffraction, multi-path fading and
shadowing effect. The weighting algorithm is used for
the calculation to get the real measurement data and
display them on the map coordinate. The measurement
of RSSI values in 81 test points from Tag 1, 2, 3 and 4
is shown in Fig. 5.
The angle of the RFID Tag placement affects the
accuracy remarkably. Hence, we add bar antenna on
RFID tags to reduce the directional influence. After
adding the antenna, the average measurement error
reduces to below 0.7 m, while it was 1 to 1.5 m without
using the antenna. Then, the Fuzzy neural network is
applied to further reduce positioning errors. The
adjusted weight, variance and the mean are shown in
Fig. 6.
Every weight, mean and variance is updated as:
wk = wk + ∆wk
(27)
mij = mij + ∆mij
(28)
σ ij = σ ij + ∆σ ij
(29)
Note that FNN off-line learns the environment.
Then the updated parameters are used for the indoor
localization.
EXPERIMENTAL RESULTS
In the experiment, the RFID components include
one 2.4 GHz active reader and four tags. The reader is
Tag 3
Tag 4
81
72
63
54
45
36
27
18
09
80
79
78
77
76
75
74
73
01: -76.8
02: -98.7
03: -65.8
04: -111.5
01: -108.1
02: -88.6
03: -70.2
04: -118.2
01: -102.8
02: -111.6
03: -96
04: -106.1
01: -106.2
02: -101.3
03: -97.1
04: -95.3
01: -106.7
02: -94.8
03: -105.8
04: -84.1
01: -95.6
02: -105.5
03: -106
04: -87.1
01: -92.8
02: -103.8
03: -106.9
04: -84.7
01: -73.5
02: -76.6
03: -107
04: -69.8
01: -73.9
02: -101.7
03: -65.8
04: -114.6
01: -105.2
02: -105.8
03: -99.1
04: -101
01: -99
02: -104.7
03: -95.8
04: -107.3
01: -96.7
02: -99.3
03: -89.5
04: -95.8
01: -93
02: -91.2
03: -101.1
04: -76.9
01: -107.2
02: -96
03: -82.5
04: -84.3
01: -103
02: -106.5
03: -104
04: -84
01: -103.4
02: -121.7
03: -88
04: -85.7
01: -92.6
02: -95.1
03: -90.4
04: -85.9
01: -101
02: -105.8
03: -70.8
04: -95.2
01: -101.7
02: -103.9
03: -97.2
04: -109.2
01: -84.4
02: -104.2
03: -85.2
04: -106.5
01: -95.7
02: -106
03: -76.5
04: -106.2
01: -77.4
02: -103.9
03: -76.7
04: -79.9
01: -118.6
02: -104.7
03: -101.4
04: -90.5
01: -93.1
02: -106.6
03: -101.8
04: -86.9
02: -91.7
03: -90
04: -73.9
02: -109.3
03: -118.1
04: -77.4
01: -99.9
02: -89.9
03: -99.6
04: -115.7
01: -103.9
02: -85
03: -73.7
04: -85
01: -103
02: -128.9
03: -91.8
04: -96.7
01: -90.1
02: -83.8
03: -77.7
04: -78
01: -106.3
02: -92.9
03: -89.1
04: -106.3
01: -101.1
02: -84.7
03: -89.5
04: -79.8
01: -98.4
02: -105.8
03: -107.3
04: -71.4
01: -76.8
02: -91.5
03: -99
04: -78.6
01: -87.8
02: -97.3
03: -77
04: -85.9
02: -77.3
03: -72.4
04: -91.7
02: -73.4
03: -76.7
04: -105
01: -100
02: -83.9
03: -87.9
04: -93.5
01: -116.5
02: -76.5
03: -95.1
04: -95.9
01: -69.8
02: -105.9
03: -106.3
04: -77.7
01: -69.1
02: -100
03: -89.4
04: -76.4
01: -77.4
02: -86.7
03: -95.7
04: -95.3
01: -78.1
02: -114.7
03: -100.6
04: -91.6
01: -88.8
02: -99.5
03: -102.9
04: -95.9
01: -77.9
02: -107
03: -105.8
04: -96.5
01: -74
02: -99.7
03: -95
04: -91.8
71
62
53
70
61
52
69
60
51
44 01: -106 43 01: -87.1 42
35
34
33
68
59
50
41
32
67
58
49
40
31
66
57
48
39
30
65
56 01: -77.2 55 01: -114.1
47
01: -83.6
02: -78
03: -96.6
04: -104.5
01: -86
02: -85.1
03: -85.6
04: -87.6
01: -86.8
02: -76.8
03: -91.8
04: -101.1
01: -78.2
02: -71.7
03: -106.4
04: -87.3
01: -91.9
02: -84.9
03: -106.8
04: -95.9
01: -104.5
02: -90.3
03: -114.4
04: -93
01: -69.9
02: -77.2
03: -106.8
04: -95.1
01: -85
02: -107.1
03: -95.1
04: -89
01: -88.5
02: -97.4
03: -91.8
04: -94.8
01: -103.5
02: -71.7
03: -99.8
04: -76.5
02: -77
03: -104.2
04: -78.5
01: -85.4
02: -83.1
03: -105.7
04: -85
01: -100
02: -78.4
03: -101.1
04: -105.4
01: -73.2
02: -87.7
03: -89
04: -78.1
01: -85
02: -103
03: -100.3
04: -96.7
01: -69.8
02: -91
03: -97.1
04: -79.5
01: -100.1
02: -83.3
03: -97.8
04: -86.3
01: -102.4
02: -67.1
03: -99.9
04: -86.8
01: -90.4
02: -76.7
03: -92.1
04: -91.3
01: -85.1
02: -79.9
03: -95.9
04: -83.3
01: -94.5
02: -71.4
03: -104.3
04: -87
01: -91.8
02: -94.9
03: -103
04: -92
01: -74.9
02: -97.7
03: -84.4
04: -88.5
01: -88.3
02: -77.4
03: -92.2
04: -94.6
26
25
17 01: -89 16
08
07
24
15
06
23
14
05
22
13
04
21
12
03
64
01: -91.4
02: -121.2
03: -90
04: -67
38
29
20
11
02
01: -84.4
02: -117
03: -91.1
04: -99.2
46
37
28
19
01: -85.5
02: -76.9
03: -94.5
04: -73.2
01: -92.4
02: -83.6
03: -106
04: -76.3
01: -73.8
02: -96.7
03: -99.2
04: -102.8
01: -77.3
02: -102.8
03: -89.7
04: -76.3
01: -89
02: -106.2
03: -107.2
04: -107.2
01: -86
02: -91.5
03: -102.7
04: -95
01: -84.2
02: -112.6
03: -123.3
04: -88.1
01: -77.9
02: -90.9
03: -84.7
04: -86
01: -58.1
02: -113.1
03: -103.3
04: -73.9
10
01
Tag 2
Tag 1
RFID Tag1
RFID Tag2
RFID Tag3
RFID Tag4
Fig. 5: The measured RSSI values and the testing points
1237
n
RFID Reader Coordinate
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
30
W1
W2
W3
W4
20
W5
W6
W7
10
W8
W9
W 10
weights
W 11
0
W 12
W 13
W 14
W 15
-10
W 16
W 17
W 18
W 19
-20
W 20
W 21
W 22
W 23
-30
W 24
W 25
-40
0
0.1
0.2
0.3
0.4
0.5
0.6
time (sec)
0.7
0.8
0.9
1
(a)
5
x 10
500
σ 11
8
σ 12
7
σ 13
6
σ 14
0
σ 21
4
σ 22
3
σ 23
M15
-500
M21
M22
-1000
M23
M24
σ 24
2
M13
M14
σ 15
5
M11
M12
M
SIGMA
9
-1500
σ 25
M25
1
-2000
0
-1
0
0.1
0.2
0.3
0.4
0.5
0.6
time (sec)
0.7
0.8
0.9
1
-2500
(b)
0
0.1
0.2
0.3
0.4
0.6
0.5
time (sec)
(c)
Fig. 6: The weight, variance and mean diagram of the FNN
Fig. 7: The tracks of the wheelchair robot
1238
0.7
0.8
0.9
1
Res. J. Appl. Sci. Eng. Technol., 6(7): 1233-1239, 2013
Next, the wheelchair robot is driven to move
forward. The tracking error between the actual distance
and the measured distance is larger than while the
wheelchair robot stands still. With the parameter
adjustment by using the proposed FNN, the average
error is reduced. It is also found that when the
wheelchair robot moves faster, the error becomes
larger. A worst case typical experimental result is
shown in Fig. 7, where the track of the wheelchair robot
can be determined by applying the proposed method.
The red lines are the actual tracks of the robot and the
continuous green dots are the locations obtained
through FNN positioning. Experimental result shows
that the localization error can be smaller than 2 m as
long as the moving speed of the robots is not more than
20 m/min. In short, the proposed system is realized and
verified to have relatively accurate indoor localization.
The use of FNN further increases the accuracy on the
real measurement.
CONCLUSION
In this study, an RFID technology with the use of
fuzzy neural network for indoor positioning of
wheelchair home health care robots is proposed. The
proposed RFID positioning system includes one 2.4
GHz RFID reader and four tags. The computational
algorithm is somewhat similar to that with four readers
and one tag. Yet the cost is much lower. Further, the
FNN technology for adjusting the positioning system
parameters increases the environmental adaptability and
positioning accuracy. Experimental result shows that
the localization error can be smaller than 2 m as long as
the moving speed of the robots is not more than 20
m/min. The successful result can be applied to most of
the wheelchair robots. Therefore, the method is indeed
practical in many applications.
ACKNOWLEDGMENT
The authors would like to thank National Science
Council, Taiwan, for supporting this study under
Contracts NSC98-2218-E155-007, NSC99-2218-E155002 and NSC100-2221-E155-005.
REFERENCES
Celeste, W.C., T. Filho, M.S. Filho and R. Carelli,
2008. Dynamic model and control structure for an
autonomous wheelchair. Proceeding of IEEE
International Symposium on Industrial Electronics,
Cambridge, pp: 1359-1364.
Cooper, R.A., B.E. Dicianno, B. Brewer, E. LoPresti,
D. Ding, R. Simpson, G. Grindle and H. Wang,
2008. A perspective on intelligent devices and
environments in medical rehabilitation. Med. Eng.
Phys., 30(10): 1387-1398.
Ho, C.C., M.C. Chen and T.T. Tsai, 2011. Development
of sensory fusion-based intelligent fire fighting
robot. Adv. Sci. Lett., 4(6-7): 2037-2042.
Ho, C.C., M.C. Chen, T.T. Tsai and P.P. Abhishek,
2004. LANDMARC: Indoor location sensing using
active RFID. Proceedings of the 1st IEEE
International Conference on Pervasive Computing
and Communications, pp: 407-415.
Misono, Y., Y. Goto, Y. Tarutoko, K. Kobayashi and
K. Watanabe, 2007. Development of laser
rangefinder-based SLAM algorithm for mobile
robot navigation. Proceeding of SICE Annual
Conference, pp: 392-396.
Vachhani, L. and K. Sridharan, 2008. Hardwareefficient prediction-correction-based generalizedvoronoi-diagram
construction
and
FPGA
implementation. IEEE T. Ind. Electron., 55(4):
1558-1569.
1239