ZigBee-based intelligent indoor positioning system soft computing2013.6

METHODOLOGIES AND APPLICATION

ZigBee-based intelligent indoor positioning system soft computing

Leh Luoh

óSpringer-Verlag Berlin Heidelberg 2013

Abstract Nowadays positioning system is no longer only for military purpose,while it has been widely applied to various livelihood purposes such as biological information,emergency rescue,public facilities and individual safety.While the most frequently used to identify the coordinates of users is global positioning system (GPS),however,it tends to be interfered by indoor buildings such that it cannot be effectively used in indoor environment.Recently,wireless sensor network has become a trendy research topic,the positioning service of indoor positioning system can be achieved by the measurements of received signal strength (RSS)or link quality indicator (LQI).In this paper,the average RSS is ?rst adopted for reducing the noise interference of LQI,and then the object to be detected will be trained by radial basis function network (RBFN)with the capability of identifying the environment of location.ZigBee module will then be integrated to realize a set of convenient wireless indoor positioning system with low cost.In addition,multiple similar arti?cial neural networks within the same region will be adopted to further improve the positioning accuracy.Experiments shown that this study is capable of effective enhancement of existing IPS accuracy with the average error of indoor positioning at 2.8meters 100%comparing with other approaches.

Keywords Indoor positioning system áIntelligent wireless sensor network áZigBee áRBFN

1Introduction

Recently,the applications of wireless technology for por-table devices and for positioning of objects have been widely discussed (Alsindi et al.2004;Caffery 2000;Cheng and Lin 2009;Fang and Lin 2008;Jennic 2008;Kim 2009;Li and Pahlavan 2002;Papageorgiou et al.2005;Wang 2006;Zhang et al.2007,2008;Zhu and Zhu 2001).The so-called ‘‘wireless sensor network’’is a kind of wireless communication technology for information detection of surrounding environment by a distributed network com-posed of huge amount of wireless sensors with low power consumption and low cost through self-organized com-munication.Wireless sensor network can be applied to various ?elds including biological information,military battle ?eld,industrial control,intelligent building and family care (Han and Lim 2010;Han et al.2011;Li 2007;Pablo Garc ?0a Ansola 2012;Sung and Hsu 2011;Tekinay 1988;Tseng et al.2010;Sung et al.2010;Wang 2006;Xinrong 2007;Zhang et al.2008).However,among numerous application examples of wireless sensor network,sensors must ?rst identify own positions before realizing the positioning and tracking of external targets (Chien 2006;Chuan 2008;Dye 2009;Fang and Lin 2008;Huang et al.2009).Along with the development of positioning application of wireless sensor network,more attentions have been drawn to the indoor positioning system (IPS)of intelligent robot in recent years (Daniel et al.2009).For example,in large public facilities such as museums,schools,or government agencies,service robots have been used for the positioning of navigation and tour guide.Or robots can be sent out for exploration,searching,mapping and rescuing (such as nuclear power plant accident in Fukushima Japan 2011)in unknown indoor environment to avoid direct human encounter of unidenti?ed danger.

Communicated by A.Di Nola.

L.Luoh (&)

Department of Electrical Engineering,Chung Hua University,707,Sec.2,WuFu Rd.,Hsin-Chu 30012,Taiwan,ROC e-mail:lluoh@https://www.360docs.net/doc/8c14976556.html,.tw

Soft Comput

DOI 10.1007/s00500-013-1067-x

The most well known outdoor positioning technology is the global position system(GPS),yet there must be direct line-of-sight between satellite and receiver in an unshielded environment(Alsindi et al.2004;Lee and Chung2006; Papageorgiou et al.2005;Tekinay1988).For the realiza-tion of IPS by wireless sensor network technology,several sensors with?xed known positions will be set up for cal-culating the positions of receiving nodes previously unknown such that the correct reference coordinates can be obtained.As for the mechanisms for estimation of node position,there are two positioning methods as range-based positioning calculation and range-free positioning calcu-lation(Tekinay1988)for frequently used for radio fre-quency identi?cation(RFID),Bluetooth and ZigBee. Among them,RFID is a passive device mainly providing positioning services of shorter distance such as logistic control and tour guide services,where the user must install additional reader on the robot to be read while being close to the target(with tag).Bluetooth is similar to RFID but small region.As for ZigBee,it provides mid-range sensing with low cost,low power,low complexity and high sca-lability mainly for personnel monitoring and tour guide services(Alsindi et al.2004;Caffery2000;Zhang et al. 2008).Many positioning methods have been applied to wireless network including‘‘Time of Arrival’’(TOA) (Alsindi et al.2004)‘‘Time Difference of Arrival’’(TDOA)(Zhu and Zhu2001),‘‘Arrival of Angle’’,(AOA), or‘‘Difference of Arrival’’(DOA)(Caffery2000)and ‘‘Received Signal Strength’’(RSS)(Li and Pahlavan2002; Parker and Valaee2007;Papageorgiou et al.2005;Zhang et al.2007).Since the positioning system following IEEE 802.11standard usually consumes more electricity,it is not the optimal solution for long-term tracking.There are numerous existing challenges and issues such as the han-dling of fading channel,energy loss due to wave refraction and re?ection,and multipath interference leading to inability of accurate position estimation(Li and Pahlavan 2002).However,while ZigBee hardware with low cost and low complexity is capable of effective acquisition of measurement value of RSS(Cheng and Lin2009),this study is mainly for constructing an IPS with certain accu-racy.The focus will be on the utilization of ZigBee with low power consumption and high scalability,and RSS measurement techniques,in conjunction with RBFN net-work as the algorithm to realize a robot IPS.As indicated by the results of multiple experiments,an average posi-tioning error of1.47m100%can be obtained while the position estimation has been conducted based on multiple RBF networks,while the overall average error is within the range of2.8m100%.

The structure of this paper is as follows.In Sects.2,3, the related knowledge and the proposed novel system structure have been introduced.Section4discussed the experimental results and comparison analysis,while the conclusion and future outlook of this paper can be found in Sect.5.

2Positioning technologies and ZigBee

While people have been exploring how to introduce robots into daily lives in additional to the military and industrial applications,thus,more focuses have been put on the research on IPS of robot.Among wireless sensor networks, the ZigBee features such as easy installation,ultra low power consumption,and low statistics transmission have made it popular recently after RFID.

2.1Comparisons among IPS technologies

The more frequently used IPS technologies are infra-red (IR),ultra sonic,and wireless local area network(WLAN) etc.,and the comparisons among their characteristics are as shown in the Table1and Fig.1.

There are four frequently used technologies based on the measurement of distances or angles between nodes:‘‘Angle of Arrival’’(AOA),‘‘Time of Arrival’’(TOA),‘‘Time Difference of Arrival’’(TDOA),and‘‘Received Signal Strength Indication’’(RSSI).(Kegen and Jay Guo2009).

2.1.1Angle of arrival

Angle of arrival(AOA)is to determine the position of moving target according to the angle direction of arrived signal as Fig.2.The position of moving terminal can be assured by Eq.(1).

y p?x pátan h1ty1àx1átan h1

eT

y p?x pátan h2ty2àx2átan h2

eT

&

e1T

For multiple receivers,the equation can be extended as (2)for more accurate calculation.

y p?x pátan h1ty1àx1átan h1

eT

y p?x pátan h2ty2àx2átan h2

eT

...

y p?x pátan h nty nàx nátan h n

eT

8

>>>

<

>>>

:

e2T

However,this AOA technique relies heavily on the directional precision of antenna where additional antenna array must be installed leading to greatly increased cost.

2.1.2Time of arrival

Time of arrival(TOA)is calculation of distance between the detected target and the base station according to signal

L.Luoh

transmission time,and using algorithm as Fig.2.The dis-tance between receiving node and signal source can be calculated as:

R i?ct ie3Twhere c is the speed of light,t i is the signal transmission time detected by receiving node i,and R i is the signal transmission distance detected by receiving node i.That is, we can get a circle with signal source on the circumference as:

R i?

????????????????????????????????????????

ex iàxT2àey iàyT2

q

e4Twhereex i;y iTis the position of receiving node i and(x,y)is the position of target signal source.If there are three receiving nodes,then following equations can be obtained: R1?

??????????????????????????????????????????

ex1àxT2àey1àyT2

q

R2?

??????????????????????????????????????????

ex2àxT2àey2àyT2

q

R3?

??????????????????????????????????????????

ex3àxT2àey3àyT2

q

8

>>>

<

>>>

:

e5T

In ideal situation,the three circles obtained from the measurement results of these three receiving nodes will intersect at one point,and this point is the position of target signal source(x,y).For this method to accurately determine the time delay of signal transmission,all nodes must be kept synchronized(Fig.3).

2.1.3Time difference of arrival

Here two signals with different transmission speeds are transmitted from transmitting nodes simultaneously.The calculation of distance between these two nodes is con-ducted according to the transmission speeds and the arrival times of these two signals(T1and T2)as Fig.4,the known two transmission speeds C1and C2,and the distance between these two nodes is D,such that:

Table1Comparisons with other settled methodologies

Principle Advantages/disadvantages

Booth-tooth Distance is measured from

signal strength

NLOS,small size,short

range but low power,

operate on small region,

unstable with low noise

rejection

RFID RF NLOS,short range

(several meter6–10m),

low cost

Ultra-wideband Narrow-pulse(ns)

transmission

High-penetrability,low

power,low complexity,

high multi-path rejection

Ultra sonic After ultra sonic wave is

emitted,the different

cross-sections re?ection

from objects will be

observed Susceptible to environmental and multi-path interference, high cost

IR Distance is measured from

the arrival time of

received re?ection Line-of-sight(LOS) transmission,short range,susceptible to sunlight interference and accuracy due to small emission angle,low penetrability and fail on shelter,high cost

WLAN Position estimation based

on the T/R of wireless

signals,IEEE802.11Small range1–20m,high power,wireless signal is susceptible to interference from obstructions

ZigBee Wireless sensor network NCLO,700m for

hardware using in this

paper,high ef?ciency,

low power,low cost

ZigBee-based intelligent indoor positioning system

D ?eT 1àT 2T

C 1C 2C 1àC 2

e6T

Time synchronization is necessary for both TDOA and TOA.However,the advantage of TDOA is the use of relative time of arrival instead of the absolute TOA for error reduction.

2.1.4Received signal strength indication

Received signal strength by multiple receiving nodes from transmission nodes are used for determination of the dis-tance between receiving node and transmitting node according to the corresponding relationship between transmission distance and signal attenuation established by transmission model in real environment.In addition,it also relies on several factors such as the accuracy of distance estimation,quantity,density and locations of receivers,and variables of onsite environment.2.2IEEE 802.15.4and ZigBee

Power supply for sensor installation is a crucial concern for wireless network technology especially for the demand for installations of large amount of sensors in outdoor envi-ronment.This is how the ZigBee technology was born as a wireless network technology.ZigBee Alliance (2009)and IEEE have jointly formulated the IEEE 802.15.4standard protocol with low power consumption,low data transmis-sion and low cost as shown in Fig.5.The IEEE 802.15.4standard protocol is shown in Fig.4(Lu et al.2008).2.2.1Network topology

In the physical layer of IEEE 802.15.4,there are two kinds of hardware nodes:Full Function Device (FFD)and Reduced Function Device (RFD).In terms of software function,network nodes can be categorized as PAN Coordinator,Coordinator,and Device (also known as End Device in ZigBee protocol)as described in Table 2.

In terms of network connection,IEEE 802.15.4provides three basic types of network topologies:

2.2.1.1Star topology As shown in Fig.6,star topology is a network formed by a PAN Coordinator at the center position connected to surrounding nodes,where these surrounding nodes are called end device,which can only communicate with the PAN Coordinator at the center.2.2.1.2Tree topology Tree topology structure is a par-ent–child relationship.There is a mother node for each node other than PAN Coordinator,and there can be mul-tiple child nodes for each node.Each node in the network can only communicate with its mother node or child node,while each mother node can be the Local Coordinator for its child nodes as shown in Fig.7.

2.2.1.3Mesh topology In mesh topology,in addition to one node capable of becoming the PAN Coordinator,mesh-like interconnection can be found among all nodes.Some but not all of the nodes can be directly used for information transmission as shown in Fig.8.

Application

Application Layers

Security

32-/64-/128-bit encryption Network

Star/Mesh/Cluster-Tree

MAC Layer

PHY Layer

868MHz/915MHz/2.4GHz

Silicon Zigbee Stack Application

User

ZigBee Alliance

IEEE 802.15.4

Fig.5IEEE 802.15.4/ZigBee protocall and structure Table 2Node function

Software function of node

Node type PAN coordinator

Assign the PAN ID to network,and manage

FFD

Frequency searching and assignment Connect ability with other device Routing ability

Package reception and transmission

Coordinator

Connect ability with other device FFD Routing ability

Package reception and transmission

End device

Connect with PAN

FFD/RFD Package reception and transmission

L.Luoh

2.2.2Software structure

There are two layers in the software con?guration within the node of IEEE802.15.4as shown in Fig.9.

Interaction among applications are conducted through the Stack API of802.15.4and the Stack Layers of IEEE 802.15.4.The interactive access to hardware registers by applications is conducted through integrated peripheral equipment API and integrated on-chip peripheral equip-ment.In hardware layer all kinds of interruptions can be generated and sent to every software module through interruption control codes.In addition,the following fea-tures can be used for software design including(1)‘‘Clear Channel Assessment’’(CCA)for detection of current status of communication channel;(2)Link Quality Indicator (LQI)is a kind of measurement of RSS Yet the variation range of LQI value is larger than RSS value such that it has higher resolution.

2.3Arti?cial neural network algorithm

Arti?cial neural network(ANN)has the information pro-cessing capability similar to biological neurons with basic model as shown in Fig.10.There are multiple inputs x0x1;...;x p and one output y for each arti?cial neuron,and the relationship can be described by the equation below: y j?u

X p

i?1

w jpáx pàh j

!

e7T

where w jp represents the weight from the p th dimension to the j th arti?cial neuron,h j?w j0represents the threshold value of this neuron,

P p

i?1

w jiáx i represents the overall input obtained by the j th arti?cial neuron with physical meaning as the membrane potential at axon hillock,and ueáTrepresents activation function.

u veT?expà

v2

2r2

e8Twhere r is the real number positive constant.

2.3.1Competitive learning algorithm

The single layer neural network based on competitive learning algorithm(Su and Chang2007),which is also known as Kohonen rule or‘‘Winner-take-all learning rule’’, is as shown in Fig.11.This learning algorithm is usually used for cluster analysis,which is to?gure out the structure and cluster relationship of data itself with no prior classi-?cation information.If we assume that the N th dimension input can be represented by X:

X?x1;x2;...;x N

? Te9TThe weight vector of the j th arti?cial neuron is as shown in(10):

w j?w j1;w j2;...;w jN

??T

j?1;2;...;Ke10Twhere K is the number of selected arti?cial neurons.The steps are as shown below:

ZigBee-based intelligent indoor positioning system

Step 1:competitive phase :winner is selected.

The arti?cial neuron with the smallest Euclidean dis-tance between input vector X and weight vector w j is the winner as shown in (11):

d X eT?min j

X àw j

;j ?1;2;...;K

e11TStep 2:reward phase :the weight vector of winner is

adjusted.

w j n t1eT?w j n eTtg X àw j n eTàá

e12T

where g and n are the number of iterations of learning speed and learning process.

Step 3:iteration phase:If the number of iterations has reached,the training must be stopped.Otherwise the training will be resumed from competitive phase.2.3.2Radial basis function network

The RBFN is a typical multi-layer feed-forward network with rapid learning capability (Su and Chang 2007).Fig-ure 12shown a three-layer RFBN.If we assume the input dimension as p and the number of neurons in hidden layer as J ,then the output can be shown as:F x eT?

X J j ?1

w j u j x eTth ?

X J j ?0

w j u j x eT

e13T

where x represents input vector,w j represents the weight

value from the j th arti?cial neuron in hidden layer to the arti?cial neuron in output layer,h ?w 0represents adjustable offset,and u j x eTrepresents the basis function of output value of the j th arti?cial neuron in hidden layer,which is a Gaussian function with de?nition as (14

):

Fig.9Basic software

con?guration of IEEE802.15.4

L.Luoh

u j x eT?exp à

x àm j 2

2r 2j

!

e14T

where J represents the number of arti?cial neurons in hidden layer,m j represents the center position of the j th Gaussian function,and r j represents the standard error of the j th Gaussian function.

Two phases of learning processes are included in the RBFN.The learning of nodes in hidden layer is mainly about using non-supervision method for adjusting the adjustable parameters m j and r j ,while output nodes are adjusted by supervision method adoption of ‘‘Last Mean Square Algorithm’’(LMS)for network parameters.The LMS is by the gradient descent method.E n eT?1

y n àF x n eTeT2

e15T

where F x n eTis the n th output value,and y n represents the expected output of the n th input data.According to the selected basis function as (6),the adjustment equations for w ,m j and r j are derived as:w n t1eT?w n eTàg

o E n eTo w n eT

?w n eTtg y n àF x n eTeTáu x n eT

e16T

m j n t1eT?m j n eTàg

o E n eT

o m j n eT

?m j n eTtg y n àF x n eTeTáw j n eTáu x n eT

á1r 2j

x n àm j n eTà

áe17T

r j n t1eT?r j n eTàg

o E n eTo r j n eT

?r j n eTtg y n àF x n eTeTáw j n eTáu x n eT

á1r 2j

x n àm j n eT

2e18T

where g is learning rate,meanwhile

u x n eT?u 0x n eT;u 1x n eT;...;u J x n eTh i T

e19Tu 0x n eT?1e20T

and

w n eT?h n eT;w 1n eT;...;w J n eT? T

e21T

The key to the technology of RBFN is to conduct dynamic adjustment of cluster center according to the actual varia-tion of sample set,and its learning rule can be regarded as the best solution for optimization problem.

3The proposed IPS

3.1Design concept

Currently the research on context-awareness system has

become more and more popular with robot playing an important role.The IPS proposed is designed by integrating the application of ZigBee hardware,and the RBFN esti-mation algorithm is used for achieving the required accu-racy.First,RBFN is used for conducting model of environment characteristics.It is divided into of?ine phase and real time phase.Of?ine phase includes setting of experimental environment,selection of sampling points,and measurement.Real time phase is all about the follow-up veri?cation and analysis of experiments.The architec-ture is as shown in Fig.13.

The positioning method as shown in Fig.14by ?rst placing three sensors with known positions in indoor envi-ronment,and then the strength of signal sent from sensor to target (robot)is used to estimate the position of

robot.

Fig.12RBFN structure

Senser Node1

Senser Node2

Senser Node3

Target Node

RS-232

Fig.13The proposed localization con?guration of ZigBee

ZigBee-based intelligent indoor positioning system

3.2Pre-processing for received signal

Basically,signals transmitted from sensor nodes must be received for a period of time.In other words,signal mea-surement is susceptible to environmental effects such as people walking,etc.The effects of these interference fac-tors can be reduced by adopting pre-processing of received

signal(value in the box)as Fig.15.

3.3Overall system

The ZigBee development kit(model number:FT-6250/FT-6251)developed by Fontal has been adopted for measuring LQI,and the estimated algorithm is RBFN.The complete system is Fig.16.

Figure16is the proposed function block diagram of robot IPS,where the positioning procedure can be divided into two phases.(1)The averaging LQI values of envi-ronmental will be collected.Then it can be input into RBFN for training.The cluster classi?cation is conducted for all LQI values in the database through competitive learning to de?ne the parameters of RBFN.(2)The real time LQI values will be thrown into the trained RBFN for veri?cation as shown in Fig.17.

3.4RBFN-based localization algorithm

3.4.1Competitive learning algorithm

The main task of competitive learning algorithm here is the cluster analysis of stored training sample sets for identifying the cluster center vector of the training sample sets.The inputs in Fig.17are(x,y|LQI1,LQI2,LQI3),where(x,y)represents the coordinates of the point to be measured.Follow the similar steps as Subsect.2.3.The number of neurons in set in hidden layer and the sample is selected in Fig.12,where

X?LQI1;LQI2;LQI3

? Te22TIn addition,the cluster center vector of initial con?guration is selected from input training for the?rst time.The winning neuron in hidden layer is calculated by (11).The weight vector of the winning neuron is adjusted by(12).The calculated cluster center vector m j and standard deviation r j are then saved for the con?guration of RBFN,and the learning procedure is shown as Fig.18..

3.4.2Radial basis function network

We?rst use the competitive learning algorithm for deter-mination of initial cluster center vector m j and standard deviation r j,then the RBFN in Fig.12is used for training of environmental positioning model.

Step1:the result of competitive learning algorithm is used for initiating the cluster center vector m j and standard deviation r j of RBFN.The weight vector w,learning rate is g,and convergence conditions are also set.A proper ini-tialization of RBFN are given as6,000iteration,one out-put,three hidden neuron,three input vector,and initial weights with-1–1.

Step2:this is the feed-forward phase where the input training sample is sent into the network during the n th learning such that the output of this learning can be calculated.

Step3:this is the back propagation phase for calculation of mean square error of expected outputs of network out-puts and training samples.

Step4:the parameters of RBFN such as w,m j,r j and h are adjusted simultaneously in the opposite direction of the gradient of mean square function.

Predefined Environment

Model

Localization Algorithm

Estimated Position of Target

RSS

Sensor

Fig.14RSSI localization con?guration

Neural network

training

RS-232

Sensor Node 1Sensor Node 2

Sensor Node 3

Target Node

Collected

LQI datum

Build map via

LQI datum

Trained neural

model Training Stage

Positioning Stage

Fig.16System function block diagram

L.Luoh

Step5:checking of whether or not the convergence conditions have been met.Otherwise it will go back to Step 2for continued training.Number of iterations will be assigned here as the set convergence condition.

Step6:output of the estimated coordinates of the position of measured target.

Here the number of vector centers is determined based on

the number of inputs.For example,if the number of inputs is three and the determined number of vector centers is three or four,the trained arti?cial neural network will perform better.The whole procedure of RBFN is shown as Fig.19.

3.5Establishment of experimental statistics

The selected experimental environment is a7.26916.5m rectangular space on the third?oor of Engineering Build-ing I of Chung Hua University as Fig.14,where S1,S2,S3 are the transmission nodes,and white and red nodes are the nodes of target measurement and testing(Figs.20,21).In this experiment the?eld measurement is adopted for constructing the required database,while random distri-bution sampling approach has been conducted,and49 target nodes are selected for LQI.5LQI values from three transmission nodes will be recorded for each node to establish the database of this site,where38white nodes selected from the49nodes of moving target are used as sampling nodes for the training of RBFN.Two strategies are proposed.Strategy1:a single network is given for whole environment.Strategy2:the experimental environ-mental will be divided into four independent areas and multiple neural networks will be used.After the estab-lishment of network,the remaining11red nodes from database will be used as the tested nodes.

Refine neural network number

Input training

samples

Samples clusting by competition algorithm

Comfirm

cluster center

Y

N

Converge

Start

A

End Fig.18The learning of competition algorithm

Initialization

Forward step

Back propagation

step

Define cluster

center and divition

Parameters

adjust

Converge

Target position

estimation and

localization

Y

N Input training

samples

Start

A

A

End

Fig.19The training procedure of RBFN

ZigBee-based intelligent indoor positioning system

3.6Software/hardware implementation

FT-6200(Jenic,Surewin2010)kit includes FT-6250and FT-6251is as Fig.15with32-bit RISC compatible, 2.4GHz IEEE802.15.4communication protocol,and 64kB ROM and96kB RAM.Development platform developed by Jennic is used for program development.

For FT-6250/FT-6251,data is not transmitted to each other at any time among all nodes.Instead,the polling approach is adopted.When the Coordinator as an unknown node has assigned certain End Device(as a known node)to transmit data,this End Device will start transmitting own data.It can avoid signal congestion caused by simultaneous data transmission of every End Device leading to the so-

Antenna

I/O port Ext ension

Button

Serial Port

LED

Power(b).FT-6251

Tempreture

FT-6250/FT-6251hardware(Surewin2010)

L.Luoh

called‘‘Signal Collision’’.The software is using C/C?? language.Moreover,the speci?cation of FT-6250/FT-6251 is shown as Table3.

Remark1According to the FT-6250/FT-6251speci?ca-tion(Jennicy2008),the time T s for a package to be transmitted is0.008s,the time T r or listen is0.2s.Given 1h duty cycle T the normal number of transmission N s nd reception N r are8and4,respectively.Thus,the idle time T i is about3599.136se?TàeN s T stN r T rTT.Also the amp consumption of transmission A s sleep A i and listen A l are100,0.03,and70mA/sec,respectively,the total Amp consumption for Jenic’s FT-6250/FT-6251can be calcu-lated byeN s A s T stA i TtN r A l T riT.That is the total Amp consumption is about170.37408mA,and the average consumption is0.047326133mA/sec(170.37408/3,600). In other words,with the mounting Li battery1,000mAh (For a general3A battery,the capacity is about 1,700–2,200mAh)inside FT-6250/FT-6251,it can be operated almost2.4years.

Remark2In fact,for attaining larger operate range (700m,see Table3),FT-6250/FT-6251given larger Amp consumption for transmission.For general,the ZigBee module can be operated longer than2.4years.4Experiment results and comparisons

Here,the effects of several kinds of arti?cial neural net-works and all parameters on positioning accuracy have been tested such as the effect of number of arti?cial neu-rons in network hidden layer on positioning accuracy,and the effect of size of training area on positioning accuracy.

4.1Effect of different network

The comparisons among(1)Back Propagation Network (BPN),(2)RBF,(3)RBF integrated with Kohonen learning rules,and(4)other approaches are given,where two hid-den layers have been adopted by BP,are as Table4.The observation meets the expectation where the RBFN inte-grated with Kohonen learning rules has best performance.

4.2Effect of neuron number

Based on the comparisons among three kinds of arti?cial neuron numbers in hidden layer,network training error analysis is conducted by using3,4,and5arti?cial neurons in hidden layer.The trained RBF network will be input into testing nodes for testing,and the error results of testing nodes are as shown in Table5.

It is shown in Table5that the minimum average error takes place when there are four arti?cial neurons in hidden layer,while Fig.22is about the conditions of errors with different numbers of arti?cial neurons in hidden layer.It is indicated that when there are four arti?cial neurons in hidden layer,the error within90%of test sample is less than4m.However,the error result is not decreasing with increasing number of arti?cial neurons in hidden layer.

4.3Effect of training area size

As shown in Figs.23and24,a great improvement is achieved,the use of multiple RBF networks in different

Table3FT-6250/6251speci?cation Item Speci?cation

RAM size96KB Flash size128KB Data rate250kb/s

Frequency band and operating channels 2.405–2.480GHz Channel11–26

Receive sensitivity-88dBm

Transmit range Indoor(open of?ce

environment)

100m

Outdoor(open

space)

700m Available transmit-16to14dBm

Power setting6steps of control

Antenna type Omni dipole antenna with IPEX connector Power draw1W maximum

Input and output Serial RS-232

ADC4input12-bits

resolution

DAC292input

11-bits

resolution

Comparator0/5/10/20mV

hysteresis level Sensor range Temperature0–70°C

Humidity5–90% Dimension6191369234mm Table4Error comparison for different network

Average error(m)RMS error(m)

BP 4.87 5.07

RBF 3.96 4.48

RBF/Kohonen 3.35 3.44

Table5Error comparison for different neuron number

Average error(m)RMS error(m)

n=3 3.35 3.44

n=4 3.01 3.10

n=5 4.04 4.40

ZigBee-based intelligent indoor positioning system

partitions of the same environment is going to affect the accuracy largely.There are four small partitions in this environment,each with one RBF network assigned for network training.The errors of testing results will be compared to the error of single RBFN prior to the partition,and the other settled methodologies for indoor positioning as shown in Table 6.

In Table 6it is indicated that the positioning by using multiple RBF networks in partitioned environment has better accuracy than the use of single RBF and other approaches.The main reason and signi?cance of this paper is,during training process RBFN it retains memory of certain signal interference factors caused on reception of LQI in individual region such that the environmental interference factors can be reduced leading to reduced errors.The positioning precision can be attained to 3.01m 100%and 2.8m 100%for single/multi-network,respectively.Moreover,it is 1.47m at 90%in some excellent condition.However,although the use of more RBFN will enhance accuracy,the analysis of variation of LQI value received in indoor environment is still necessary for determination of most appropriate number of RBFN to

be used.The use of excessive number of RBFN will occupy too much memory if it failed to enhance the posi-tioning accuracy.

Remark 3For general communication problem,owing to the lots of interference from environment,it is always an inevitable and critical challenge especially in wireless network.For building the electrical magnetic model of positioning environment,the RBFN arti?cial neural net-work specially is proposed in this paper to describe the nonlinear relationships and environment characteristics (including interference and obstacles model)between tar-get and each node or APs (access point).That is,the target position can be computed (need not to consider the infer-ence characteristic again)by RBFN built dynamical model (data base)after of?ine training procedure with inference characteristics in hand.

Remark 4For the follow-up veri?cation and experiment procedure of RBFN,of course,the estimated position error might arose in real word.For example,wireless network is everywhere now.Fortunately,the channel of WIFI is not continuously,the ZigBee model can be operated in dif-ferent channel to avoid the inference situation as much as possible.Surely,the in?uence is worst when there is an inference source such as generator,transformer,etc.,exists in the harsh environment.

5Conclusion

In this paper the development of the use of arti?cial neural network for robot indoor positioning algorithm has

been

Fig.22Cumulative error for different neuron number

Fig.24Cumulative error for single and multi-RBFN

L.Luoh

proposed.Based on the integration with ZigBee module,an intelligent IPS with low power consumption,high scala-bility,and construction simplicity has been designed.The veri?cation of IP algorithm with multi-RBFN has given better accuracy comparing with other approaches.The accuracy can be also enhanced by adjusting the number of hidden layers within certain range.It is applicable not only for robot IPS,but also more indoor information and ser-vices,such as personnel monitoring and logistics man-agement.It also provides various features such as greatly enhanced convenience for onsite provisioning,high feasi-bility,and reduced cost.

Acknowledgments This research was originally creative works and supported by the National Science Council,ROC,under Grant NSC 100-2221-E-216-007.The authors thank the reviewers and editor for their valuable comments and opinions on this paper. References

Alsindi N,Li X,Pahlavan K(2004)Performance of TOA estimation algorithms in different indoor multipath conditions.WCNC 1:495–500

Bahl P,Padmanabhan VN(2000)RADAR:an in-building RF-based user location and tracking system.In:Proceedings of IEEE INFOCOM,2000,pp775–784

Brunato M,Battiti R(2005)Statistical learning theory for location ?ngerprinting in wireless https://www.360docs.net/doc/8c14976556.html,put Netw ISDN Syst 47(6):825–845

Caffery J(2000)Wireless location in CDMA cellular radio systems.

Kluwer Academic Publishers,Dordrecht

Cheng Y-Y,Lin Y-Y(2009)A new received signal strength based location estimation scheme for wireless sensor network.IEEE Trans Consum Electron55(3):1295–1299

Chuan CS(2008)Realization of U campus service with room-based indoor positioning by using ZigBee.Master thesis,Department

of Computer Science,Shu-Te University of Science and Technology,Kaohsiung

Chien TW(2006)Indoor positioning algorithm for wireless radio frequency identi?cation system.Master’s Thesis,Department of Electrical Engineering,National Taiwan Ocean University, Taiwan

Dye HC(2009)Design and implementation of indoor positioning system based on Gaussian mixture model.Master Thesis, Institute of Communications Engineering,Tatung University, Taipei

Fang S-H,Lin T-N(2008)Indoor location system based on discriminant-adaptive neural network in IEEE802.11environ-ments.IEEE Trans on Neural Networks19(11):1973–1978 Garc?0a Ansola P,Morenas J,Garc?0a A,Otamendi J(2012) Distributed decision support system for airport ground handling management using WSN and MAS.Eng Appl Artif Intell 25:544–553

Haeberlen A,Flannery E,Ladd AM,Rudys A,Wallach DS,Kavraki LE(2004)Practical robust localization over large-scale802.11 wireless networks.In:Proceedings of the10th annual interna-tional conference on mobile computing and networking,Sep-tember26–October01,2004.Philadelphia,PA,USA

Han D-M,Lim J-H(2010)Smart home energy management system using IEEE802.15.4and ZigBee.IEEE Trans Consum Electron 56(3):1403–1410

Han J,Choi C-S,Lee I(2011)More ef?cient home energy management system based on ZigBee communication and infrared remote controls.IEEE Trans Consum Electron57(1): 85–89

Huang C-T,Wu C-H,Lee Y-N,Chen J-T(2009)A novel indoor RSS-based position location algorithm using factor graphs.IEEE Trans Wireless Commun8(6):3050–3058

Jennic(2008)Available online https://www.360docs.net/doc/8c14976556.html,/support/user_ guides.Data accessed,2008

Kegen Y,Jay Guo Y(2009)Statistical NLOS identi?cation based on AOA,TOA and signal strength.IEEE Trans Veh Technol 58(1):274–286

Kim M,Chong NY(2009)Direction sensing RFID reader for mobile navigation.IEEE Trans Autom Sci Eng6(1):44–54

Lee D,Chung W(2006)Discrete-status-based localization for indoor service robots.IEEE Trans Ind Electron53(5):1737–1746

Li CH(2007)Design and implementation of intelligent lighting environment based on ZigBee wireless sensor network.Master thesis,Department of Electrical Engineering,National Cheng Kung University,Tainan

Li X(2007b)Collaborative localization with received-signal strength in wireless sensor networks.IEEE Trans Veh Technol56(6): 3807–3817

Lu CH(2008)ZigBee principle and its application.Beijing Univer-sity,Beijing

Li K,Pahlavan X,Makela JP(2002)Indoor geolocation science and technology.IEEE Commun Mag40:112–118

Parker R,Valaee S(2007)Vehicular node localization using received-signal-strength indicator.IEEE Trans Veh Technol 56(61):3371–3380

Papageorgiou IG et al(2005)An enhanced received signal level cellular location determination method via maximum likelihood and Kalman?ltering.WCNC4:2524–2529

Pizarro D,Mazo Manuel S,Enrique MM,Fernandez I(2009) Localization and geometric reconstruction of mobile robots using a camera ring.IEEE Trans Instrum Meas58(8):2396–2409

Prasithsangaree P,Krishnamurthy P,Chrysanthis P(2002)On indoor position location with wireless LANs.In:The13th IEEE international symposium on personal,indoor and mobile radio communications,vol2.pp720–724

Table6Comparisons of average error for existing systems System Algorithm Precision

ENRADAR(Bahl and Padmanabhan 2000)Nearest neighbor with

environment pro?ling

3.16m at90%

RADAR(Bahl and

Padmanabhan2000)

Nearest neighbor 2.1m at38%

Prasithsangaree et al. (2002)Weighted k-nearest

neighbor

7.2m at75%,

12.2m at95%

Youssef et al.(2003)Bayesian 2.1m at90% Haeberlen et al.

(2004)

Bayesian3m at97%

Brunato and Battiti (2005)Neural network and

weighted k-nearest

neighbor

4.9–

5.2m at

90%

Single-network RBF/Kohonen 3.01m at100% Multi-network RBF/Kohonen 1.47m at90%,

2.8m at100% ZigBee-based intelligent indoor positioning system

Su MC,Chang SD(2007)Machine learning:arti?cial neural network, fuzzy system,and genetic algorithm.Chung Hwa Publishing Ltd.,Taiwan

Sung W-T,Hsu Y-C(2011)Designing an industrial real-time measurement and monitoring system based on embedded system and ZigBee.Expert Syst Appl38:4522–4529

Sung W-T,Chen J-H,Wag M-H,Hsu Y-C(2010)Remote medical care system design based on RFID and ZigBee technology via wireless sensors network.Int J Innov Comput Inform Control 6(11):5203–5220

Surewin.Available online https://www.360docs.net/doc/8c14976556.html,/.Accessed data Apr2010

Tekinay S(1988)Wireless geolocation systems and services.IEEE Commun Mag36(4):28

Tseng P-C,Chen C-Y,Hwang W-S,Pan J-S,Liao B-Y(2010)Qos-aware residential gateway supporting ZigBee-related services based on a service-oriented architecture.Int J Innovat Comput Inform Control6(6):2803–2816Wang CC(2006)Wireless sensor network-based localization imple-ment.Master Thesis Chung Cheng University,Taiwan Youssef MA,Agrawala A,Shankar AU(2003)WLAN location determination via lustering and probability distributions.In: IEEE international conference on pervasive computing and communications(PERCOM-2003),Fort Worth

Zhang J,Sun M,Wang X(2007)Dynamic distance estimation method based on RSSI and LQI.Electron Meas Technol 33(2):142–145

Zhang L,Cheng Q,Wang Y,Zeadally S(2008)A novel distributed sensor positioning system using the dual of target tracking.IEEE Trans Comput57(2):246–260

ZigBee alliance.Available online https://www.360docs.net/doc/8c14976556.html,/en/resources/ Accessed data Jan2009

Zhu L,Zhu J(2001)A new model and its performance for TDOA estimation.VTC4:2750–2753

L.Luoh