Location Estimation and Mobility Prediction Using Neuro-fuzzy Networks In Cellular Networks

Document Sample
Location Estimation and Mobility Prediction Using Neuro-fuzzy Networks In Cellular Networks Powered By Docstoc
					                                                               (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                   Vol. 9, No. 6, June 2011

   Location Estimation and Mobility Prediction Using
                Neuro-fuzzy Networks
                                                      In Cellular Networks

                        Maryam Borna                                                     Mohammad Soleimani
           Department of Electrical Engineering                                   Department of Electrical Engineering
        Iran University of Science and Technology                             Iran University of Science and Technology
                       Tehran, Iran                                                           Tehran, Iran
               maryam.borna@gmail.com                                                     soleimani@iust.ac.ir


Abstract- In this paper an approach is proposed for location                However for managing networks resources consumption
estimation, tracking and mobility prediction in cellular                 and reducing the costs of location update and call delivery
networks in dense urban areas using neural and neuro-fuzzy               procedures, prediction of user's next probable location can be
networks. In urban areas with high buildings, due to the effects         helpful. This is done by analyzing some patterns of his
of multipath fading and Non-Line-of-Sight conditions, the                mobility behavior. Therefore searching for users will be done
accuracy of positioning methods based on direction finding and           in smaller groups of cells avoiding expensive queries to
ranging degrades significantly. Also in these areas, due to high         Home Location Register (HLR). This is also useful in other
user traffic there's a need for network resources management.            wireless networks such as Ad-Hoc networks for efficient
Knowing the next possible position of user would be helpful in
                                                                         bandwidth allocation and uninterrupted hand over between
this case. Here using fingerprint positioning concept, after
choosing appropriate parameters for fingerprinting in GSM
                                                                         access points.
cellular networks, MLP and RBF neural networks were used                     Next sections are as follows: section II describes the
for position estimation. Then by the use of neuro-fuzzy                  problem of positioning in dense urban areas and related
networks a tracking and post-processing method is applied to             studies in the literature. In section III proposed approach of
estimated locations. For mobility prediction purpose the use of          this paper for positioning and mobility prediction is
ANFIS neuro-fuzzy is implemented.                                        explained and contains 3 subsections: fingerprint based
                                                                         positioning in subsection A, post processing of estimated
    Keywords-position    estimation;   neuro-fuzzy;   prediction;
cellular networks.
                                                                         path in subsection B and path prediction in subsection C are
                                                                         discussed. Section IV includes the results of evaluating the
                     I.    INTRODUCTION                                  proposed approach on database collected from GSM mobile
                                                                         phone network in city of Tehran. Results were discussed in
    Positioning in wireless networks is estimating a node's              section V and last section concludes the paper.
distance with reference to a fixed node or locating it by its
geographical coordinates. Positioning is based on parameters                              II. PROBLEM DEFINITION
used by mobile or fixed nodes for communication such as
Received Signal Strength (RSS), Time of Arrival (TOA) and                    With developments in cellular phone networks different
Angle of Arrival (AOA). According to the type of wireless                methods were considered for facilitating user positioning,
network and transmission protocols, different parameters are             such as Cell-ID, Cell-ID+TA, A-GPS, AOA, … the more
used for communication.                                                  accuracy increases the more expensive the deployment
                                                                         would be and the need for hardware and software changes in
    Among several types of wireless networks, cellular phone             both cell phone device and network infrastructure rises.
networks due to increasing usage of cell phones for                      Moreover most of these methods are sensitive to Non Lin of
communications are more distributed with more subscribers                Sight (NLOS) communication between transmitter and
so it can be said that one of the most probable items found in           receiver and multipath fading, conditions that dense urban
everyone's pocket is his cell phone. Having location                     areas are involved with. Although everyday there are more
information in cellular phone networks, various services can             and more mobile phone devices equipped with GPS receivers
be provided based on user's location ranging from                        with positioning accuracy up to few meters, but in urban
commercial and advertising services to routing, navigation               areas with high buildings where it's less likely to have line of
and emergency calls. In cellular phone networks these                    sight communication with at least 3 GPS satellites, or inside
services are referred to as Location Based Services (LBS).               buildings where the signals attenuate significantly passing
Also by locating a user's exact position network's resources             through the walls, positioning accuracy degrades
can be efficiently managed and allocated leading to proper               considerably. In such cases there's a need for an auxiliary
handover between cells and reduced co-channel interference.              method to overcome these problems.




                                                                    65                          http://sites.google.com/site/ijcsis/
                                                                                                ISSN 1947-5500
                                                                (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                    Vol. 9, No. 6, June 2011
    The difficulty in these areas is the complexity of                    standards and data collection tool. Data collection and
propagation model of electromagnetic waves caused by                      pattern learning are done offline before online real-time
multipath fading, diffraction and scattering that makes it hard           positioning.
for geometrical and statistical positioning methods relying on
relations between signal parameters and Tx-Rx separation.                     Input parameters of neural networks, fingerprint of a
                                                                          point, must be measurable, collectable and different from
    Fingerprint based positioning methods are better for                  place to place. Here we aim to attain intended fingerprints
mentioned cases [1] [2]. In these methods first a database of             from information provided by mobile phone device without
signal parameters in certain places is collected with no                  software or hardware changes in the device and additional
knowledge of propagation model of the environment and                     signaling between MS and BTS. The data can be obtained
position estimation will be done upon these information and               from mobile phone routing table. This table is used for
possible mapping among them.                                              selecting the best cell to reside and is resorted every few
                                                                          seconds. In GSM900 standard for mobile phone networks,
    One way to find the mapping relations in the fingerprint              this table contains a list of 30 radio channels (ARFCN)
database is to use Artificial Neural Networks. These                      sorted in descending order based on received power. In
networks are able to estimate the complex nonlinear                       addition to received power other parameters of currently
functions like mapping relations by parallel processing of                selected and neighboring cells are available like cell name,
neurons. Position estimation can be considered as a function              absolute radio frequency number for broadcasting cell's
approximation problem in neural networks that aims to find                status, received power level, received signal quality and
the nonlinear mapping between inputs (fingerprints) and                   timing advance (TA). Also the attributes of BTS antennas of
outputs or targets (mobile phone's coordinates).           In             each cell like its height and installation coordinates are
comparison with other database lookup methods like K-                     accessible.
Nearest Neighbor (KNN) that uses fingerprint parameters to
find its nearest Euclidean neighbors, neural networks are                     From the mentioned parameters those were chosen for
better. On the other hand since neural networks approximate               fingerprint that own following properties:
functions and fingerprint details are somewhat related to
delay and power loss of arrived signals and in turn these are                      •    Being sensitive to spatial changes. Therefore
dependent on Tx-Rx separation, it seems that neural                                     fixed parameters within a cell boundary like
networks combine both features of RSS and TOA-TDOA                                      radio channel numbers, cell antenna height and
based systems [3] [4]. Two common models of neural                                      similar parameters are not suitable for
networks in function approximation, multi layer perceptron                              fingerprinting.
(MLP) and Radial Basis Function (RBF) networks, are used                           •    Parameters should be representative of
more [2] [3] [5] [6].                                                                   multipath fading effects in propagation
     After user localization, history of his travelled places can                       environment. Received signal level, received
be considered as a time series that by recognizing his                                  signal quality and Timing Advance (TA) are
mobility pattern, his next location can be predicted. In                                such parameters. However TA is a discrete
literature for user's path prediction in wireless networks with                         value of estimated BTS-MS separation with an
different standards, Recurrent Neural Networks (RNN),                                   accuracy about 550 meters, say if TA=1, MS is
Bayesian Neural Networks (BNN) or neuro-fuzzy networks                                  in a radius of 550 meters from BTS and TA=2
were employed that some used user's behavioral pattern in a                             means MS is in a radius of 550 to 1100 meters
long period of time and different situations to learn his                               from cell antenna. Hence in a cell with radius
mobility pattern and similar users then predict their next                              less than 550 meters-like most urban cells- TA's
location [8] [9] [10].                                                                  value isn't much helpful in positioning.

    In this paper after gathering enough fingerprints with                         •    For less signaling between mobile phone and
appropriate parameters in GSM cellular phone network,                                   BTS it's better to acquire fingerprint in IDLE
positioning of mobile phone device is done by searching for                             mode rather that ACTIVE mode. TA and
the best architecture for MLP and RBF neural networks in a                              Received signal quality are determined in
dense urban area. Afterwards using tracking and prediction                              ACTIVE mode while RSS is monitored
feature of ANFIS neuro-fuzzy network, the estimated path is                             periodically even in IDLE mode.
post processed and user's upcoming path is predicted.                         We chose parameters that fulfill mentioned requirements
                                                                          namely Received signal strength from cell antenna beside its
                  III.   PROPOSED APPROACH
                                                                          coordinates for serving cell and two of neighboring cells. So
A. Fingerprint based positioning                                          there are 9 parameters to be recorded in a single fingerprint
    In fingerprint based positioning methods first a database             beside the coordinates of data gathering location. By
of fingerprints in a certain area is collected. A fingerprint of a        collecting these fingerprints in sufficient data points of the
certain point includes particular information like                        designated area, we have a suitable database for further
geographical coordinates of the point which is a specification            analysis by fingerprint based positioning methods. As
of that point. This information includes estimated signal                 discussed in previous section, for database processing, neural
parameters that are different depending on wireless network               networks predominates other methods so is employed here.
                                                                          Training set tuples are mentioned 9 parameters as input and




                                                                     66                          http://sites.google.com/site/ijcsis/
                                                                                                 ISSN 1947-5500
                                                              (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                  Vol. 9, No. 6, June 2011
latitude, longitude of respective data point as target or output
for neural networks.
    One of the problems in neural networks design,
particularly MLP networks is the lack of certain equations
for determining the perfect architecture of the network and
number of neurons in hidden layers. In the training phase of
a NN while evaluating its ability to learn, its response to new
untrained data should also be considered for the network to
generalize well. In order that 84% of data set were used for                    Figure 2. Proposed architecture for RBF neural network
training and the remaining for testing the trained network.
     For finding the best architecture for MLP NN first an
upper limit regarding members of training set is considered
for maximum number of network parameters i.e. total
number of weights and biases in neural network then by
modifying the number of hidden layers and neurons in each
layer in the defined range, the architecture yielding less
positioning error for training, testing and the whole data set
is chosen. For available database the best architecture for
MLP was one hidden layer with 23 neurons. Input layer
neurons were set to 9 and output layer neurons for estimated
latitude and longitude of mobile phone's location are 2. Fig. 1
    In standard Radial Basis Function NN a ruling parameter               Figure 3. Neuro-fuzzy network for Sugeno's fuzzy inference structure
in the design is the radial neuron's spread that determines its
sensitivity to the resemblance between network's inputs and                 Inputs of ANFIS network were estimated path with 5
weights. Searching for spread parameter resulted in value of            delays and the same path with no delays as output. For initial
196 leading to less error for testing data set. The number of           FIS generation fed to ANFIS, we used subtractive clustering
neurons has been set to its maximum i.e. the same as training           that the influence radius of every cluster for all 11
set members. Fig.2                                                      dimensions of data was set to 0.5.

    Another type of RBF networks employed here is                          ANFIS is used here to estimate the user's movement
Generalized Regression Neural Network (GRNN) that has a                 function and smooth the NN estimated path so it can be
fixed structure with little difference to standard RBF                  specified which road the user is moving on, useful in map
networks. Spread parameter for radial neuron has been                   routing and navigation purpose.
obtained like former case and set to 2.                                 C. Predicting user's next location
    After position estimation with designed neural networks,                Subsequent locations travelled by mobile phone user can
there were rather big errors in few points probably caused by           be assumed as a time series. Here we used the prediction
inadequate members of training set and unavailability of RSS            ability of ANFIS network. The structure is the same as
in some points. To lessen this error we applied a post                  before. For training the network, 20% of the beginning of the
procession on estimated path coming in next subsection.                 travelled path with 2 delays was selected as input and the
                                                                        same path with one precession as output. The remaining 80%
B. Post-processing the estimated path                                   of the path was used for testing. In this way trained network
    In this section we use ANFIS (Adaptive Neuro-Fuzzy                  would be able to predict next location by knowing the
Inference Structure) for processing the previously NN                   present and one previous location of user. Here we've used
estimated path of user's travelled places. Employed neuro-              the estimated path by neural network in section III and
fuzzy network is the neural network equivalent for Sugeno               calculated the error with respect to real path.
FIS (Fuzzy Inference Structure). In comparison with MLP,
ANFIS has a fixed architecture and no searching for best                       IV. EVALUATION OF THE PROPOSED APPROACH
structure is needed. It responds faster with less computational             For data collection we've used TEMS® drive tester tool
resource consumption. Fig. 3                                            that is used for optimization and troubleshooting of mobile
                                                                        phone network by monitoring its status. It represents
                                                                        network's data intercepted by mobile phone in a computer
                                                                        interface for further processing and also able to record data
                                                                        collection point coordinates via GPS receiver.
                                                                           We've used this tool for fingerprint database collection in
                                                                        GSM communication network in city of Tehran for about
                                                                        250 data points. For network training and simulation we used
                                                                        MATLAB® neural network and fuzzy logic toolboxes.
Figure 1. Proposed architecture for MLP neural network




                                                                   67                            http://sites.google.com/site/ijcsis/
                                                                                                 ISSN 1947-5500
                                                                                                                                               (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                                                                                                   Vol. 9, No. 6, June 2011
    Simulation results of trained neural network showed that                                                                                                               51.48
                                                                                                                                                                                                           Tajrish-Zarrabkhane

designed MLP network performs better than RBF networks,                                                                                                                                  Real track
however RBF networks are faster and easier to design but are                                                                                                              51.475         Trained Track
                                                                                                                                                                                         Predicted Track
suitable when training set members are very high.
                                                                                                                                                                           51.47


                                               51.49                                                                                                                      51.465
                                                                         Real track




                                                                                                                                                              Longitude
                                                                         Neural Network estimated                                                                          51.46
                                                                         ANFIS post-processing
                                               51.48
                                                                         BTS position                                                                                     51.455


                                                                                                                                                                           51.45
                                               51.47
      longitude




                                                                                                                                                                          51.445

                                               51.46
                                                                                                                                                                           51.44
                                                                                                                                                                              35.74   35.75   35.76    35.77     35.78     35.79   35.8   35.81   35.82
                                                                                                                                                                                                                Latitude

                                               51.45
                                                                                                                                                                          Figure 7. Proposed user's mobility prediction with ANFIS

                                               51.44                                                                                                         Fig.4 displays estimated location after post processing
                                                                                                                                                         with ANFIS indicating mitigation of high errors. Fig.5 is
                                                                35.75      35.76     35.77          35.78       35.79       35.8       35.81             Cumulative Error Probability after and before post-
                                                                                                   latitude
                                                                                                                                                         processing showing alleviation of high errors. In Fig.6 the
    Figure 4. Position estimation and post-processing result in a road
                                                                                                                                                         same path is displayed on the map by Google Earth®. It can
                                                                                                                                                         be seen after ANFIS post-processing the road the user is
                                                                                                                                                         travelling can be defined more accurately.
                                                       1

                                                      0.9
                                                                                                                                                             Fig.7 displays real, trained and predicted path by ANFIS.
                                                                                                                                                         CEP-60% is less than 115 meters which makes this
             Cumulative Distribution Function (CDF)




                                                      0.8
                                                                                                                                                         prediction useful in determination of user's next probable cell
                                                      0.7
                                                                                                                                                         to reside leading to better management of network resources
                                                      0.6
                                                                                                                                                         and successful cell reselection.
                                                      0.5

                                                      0.4                                                                                                                       V. CONCLUSION
                                                      0.3                                                                                                    In proposed approach for location estimation in cellular
                                                      0.2                                                                                                networks by neural networks in dense urban areas, mean
                                                      0.1
                                                                                                              NN estimated                               positioning error less than 80 meters and CEP-60% of 65m
                                                                                                              ANFIS post-processing

                                                       0
                                                                                                                                                         were obtained in a 3 by 4 km area that in comparison with
                                                            0      100     200     300      400       500
                                                                                         Positioning Error(m)
                                                                                                              600    700    800       900                most commercial positioning methods implemented in
                                                                                                                                                         cellular networks like E-CGI, E-OTD and AOA with 200 m
 Figure 5. Cumulative Error Probability before and after post-processing                                                                                 positioning error in such conditions, 50% improvement was
                                                                                                                                                         achieved. Meanwhile this method can be a complement to
                                                                                                                                                         GPS positioning in cases GPS signals are weak. In this
                                                                                                                                                         method there is no additional signaling or extra hardware-
                                                                                                                                                         software installation in both phone device and network.
                                                                                                                                                         We've used a fingerprint database of RSS parameters which
                                                                                                                                                         is available in most wireless networks. In comparison with
                                                                                                                                                         other positioning methods based on neural networks, we've
                                                                                                                                                         avoided a fixed structure for MLP NNs by searching for the
                                                                                                                                                         best one that suits certain database with a simple script.
                                                                                                                                                         Applying ANFIS post-processing by approximating user's
                                                                                                                                                         movement function, decreased high errors. The accuracy of
                                                                                                                                                         proposed mobility prediction by ANFIS with respect to
                                                                                                                                                         radius of cells in most cities that are about 100 to 150 m
                                                                                                                                                         makes it useful in anticipation of user's next cell to be
                                                                                                                                                         causing decrement in costs of location update and paging
                                                                                                                                                         procedure.




                                   Figure 6. Part of the path of Fig. 4 on the map of Tehran




                                                                                                                                                 68                                            http://sites.google.com/site/ijcsis/
                                                                                                                                                                                               ISSN 1947-5500
                                                                       (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                           Vol. 9, No. 6, June 2011


                             REFERENCES                                                                    AUTHORS PROFILE
[1]   Ines Ahriz, Yacine Oussar, Bruce Denby, and Gerard Dreyfus, "Full-              Maryam Borna has received her Bachelor of Science in Electrical
      Band GSM Fingerprints for Indoor Localization Using a Machine              Engineering with major of Telecommunications from Shahed University,
      Learning Approach," International Journal of Navigation and                Tehran, Iran and her Master of Science in IT Engineering with major of
      Observation, 2010.                                                         Secure Communications from Dept. of Electrical Engineering of Iran
[2]   Claude Takenga and Kyandoghere Kyamakya, "A Low-cost                       University of Science and Technology,Tehran,Iran. Her research interests
      Fingerprint Positioning System in Cellular Networks," in Second            include mobile phone networks, neural networks, microstrip antennas.
      International Conference on Communications and Networking in                    Mohammad Soleimani received the B.S. degree in electrical
      China,CHINACOM '07. , 2007.                                                engineering from the University of Shiraz, Shiraz, Iran, in 1978 and the
[3]   Anthony Taok, Nahi Kandil, and Sofiene Affes, "Neural Network for          M.S. and Ph.D. degrees from Pierre and Marie Curio University, Paris,
      Fingerprinting-Based Indoor Localization Using Ultra-Wideband,"            France, in 1981 and 1983, respectively. He is working as a Professor with
      Journal of Communications, vol. 4, no. 4, 2009.                            the Iran University of Sciences and Technology, Tehran, Iran. His research
[4]   M.H Hung, Shi-Shung Lin, Jui-Yu Cheng, and Wu-Lung Chien, "A               interests are in antennas, small satellites, electromagnetic, and radars. He
      ZigBee Indoor Positioning Scheme using Signal-Index-Pair Data              has served in many executive and research positions including: Minister of
      Preprocess Method to Enhance Precision," in IEEE International             ICT, Student Deputy of Ministry of Science, Research and Technology,
      Conference on Robotics and Automation (ICRA), 2010.                        Head of Iran Research Organization for Science and Technology, Head of
[5]   Aylin Aksu, Joseph Kabara, and Michael B.Spring, "Reduction of             Center for Advanced Electronics Research Center; and Technology
      Location Estimation Error using Neural Networks," in Proceedings of        Director for Space Systems in Iran Telecommunication Industries.
      the first ACM international workshop on Mobile entity localization
      and tracking in GPS-less environments,MELT'08 , 2008.
[6]   C Laoudias et al., "Ubiquitous Terminal Assisted Positioning
      Prototype," in IEEE Wireless Communications and Networking
      Conference, WCNC , 2008.
[7]   Hani Kaaniche and Farouk Kamoun, "Mobility Prediction in Wireless
      Ad Hoc Networks Using Neural Networks," Journal of
      Telecommunications, vol. 2, no. 1, 2010.
[8]   Sherif Akoush and Ahmed Sameh, "Mobile User Movement
      Prediction Using Bayesian Learning for Neural Networks," in
      IWCMC '07 Proceedings of the 2007 international conference on
      Wireless communications and mobile computing, 2007.
[9]   J Amar Prathap Singh and M Karnan, "Intelligent location
      management for UMTS networks using Fuzzy Neural Networks,"
      Journal of Engineering and Technology Research, vol. 2, 2010.




                                                                            69                             http://sites.google.com/site/ijcsis/
                                                                                                           ISSN 1947-5500