Wireless Personal Communications: Emerging Technologies for Enhanced Communications [1 ed.] 0792383591, 9780792383598, 9780306470462

Citation preview

WIRELESS PERSONAL COMMUNICATIONS Emerging Technologies for Enhanced Communications

THE KLUWER INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER SCIENCE

WIRELESS PERSONAL COMMUNICATIONS Emerging Technologies for Enhanced Communications

edited by

William H. Tranter Theodore S. Rappaport Brian D. Woerner Jeffrey H. Reed Virginia Polytechnic Institute & State University

KLUWER ACADEMIC PUBLISHERS New York, Boston, Dordrecht, London, Moscow

eBook ISBN: Print ISBN:

0-306-47046-2 0-792-38359-1

©2002 Kluwer Academic Publishers New York, Boston, Dordrecht, London, Moscow

All rights reserved

No part of this eBook may be reproduced or transmitted in any form or by any means, electronic, mechanical, recording, or otherwise, without written consent from the Publisher

Created in the United States of America

Visit Kluwer Online at: and Kluwer's eBookstore at:

http://www.kluweronline.com http://www.ebooks.kluweronline.com

TABLE OF CONTENTS PREFACE

I

ix

SMART ANTENNAS AND DIVERSITY 1.

Effects of Directional Antennas with Realizable Beam Patterns on the Spaced-Time Correlation T. B. Welch, M. J. Walker and R. E. Ziemer

2. Frequency Reuse Reduction for IS-136 Using a Four Element Adaptive Array

1

11

J. Tsai, R. M. Buehrer

3.

Pseudo-Blind Algorithm for SDMA Application J. Laurila and E. Bonek

4.

Integrated Broadband Mobile System (IBMS) Featuring Smart Antennas M. Bronzel, J. Jelitto, M. Stege, N. Lohse, D. Hunold and G. Fettweis

5.

II

CDMA Smart Antenna Performance M. Feuerstein, J. T. Elson, M. A. Zhao and S. Gordon

23

35

49

PROPAGATION 6.

7.

8.

9.

Wireless RF Distribution in Buildings Using Heating and Ventilation Ducts C. P. Diehl, B. E. Henty, N. Kanodia and D. D. Stancil

61

Predicting Propagation Loss from Leaky Coaxial Cable Terminated With an Indoor Antenna K. Carter

71

Building Penetration and Shadowing Characteristics of 1865 MHz Radio Waves M. Panjwani and G. Hawkins

83

Maximizing Carrier-to-Interference Performance by Optimizing Site Location J. Shi and Y. Mintz

91

10. Azimuth, Elevation, and Delay of Signals at Mobile Station Site A. Kuchar, E. A. Aparicio, J. Rossi and E. Bonek III

INTERFERENCE CANCELLATION 11. A New Hybrid CDMA/TDMA Multiuser Receiver System U. Baroudi and A. Elhakeem 12. Multiuser Multistage Detector for Mode 1 of FRAMES Standard A. Boarin and R. E. Ziemer 13. Self-Organizing Feature Maps for Dynamic Control of Radio Resources in CDMA PCS Networks W. S. Hortos

IV.

111

123

129

EQUALIZATION 14. Complex Scaled Tangent Rotations (CSTAR) for Fast Space-Time Adaptive Equalization of Wireless TDMA M. Martone

V.

99

143

15. An Effective LMS Equalizer for the GSM Chipset J. Gu, J. Pan, R. Watson and S. Hall

155

16. Self-Adaptive Sequence Detection via the M-algorithm A. R. Shah and B. Paris

167

17. Soft-Decision MLSE Data Receiver for GSM System M. Lee and Z. Zvonar

179

MODULATION, CODING AND NETWORKING 18. Turbo Code Implementation Issues for Low Latency, Low Power Applications D. E. Cress and W. J. Ebel

191

19. Evaluation of the Ad-Hoc Connectivity with the Zone Routing Protocols 201 Z. J. Haas and M. R. Pearlman VI.

INVITED POSTERS PRESENTED AT THE 1998 SYMPOSIUM 20. CDMA Systems Modelling Using OPNET Software Tool P. Gajewski and J. Krygier

vi

213

21. Signal Monitoring System for Fault Management in Wireless Local Area Networks J. F. Vucetic and P. A. Kline 22. Computer-Aided Designing of Land Mobile Radio Communication Systems, Taking Into Consideration Interfering Stations M. Amanowicz, P. Gajewski, W. Kolosowski and M. Wnuk 23. Adaptive Interference Cancellation with Neural Networks A. Zooghby, C. Christodoulou and M. Georgiopoulos 24. Calibration of a Smart Antenna for Carrying Out Vector Channel Sounding at 1.9 GHz J. Larocque, J. Litva and J. Reilly 25. Implementing New Technologies for Wireless Networks: Photographic Simulations and Geographic Information Systems H. P. Boggess, II and A. F. Wagner, II 26. Envelope PDF in Multipath Fading Channels with Random Number of Paths and Nonuniform Phase Distributions A. Abdi and M. Kaveh 27. Radio Port Spacing in Low Tier Wireless Systems H. Yeh and A. Hills 28. A Peek Into Pandora’s Box: Direct Sequence vs. Frequency Hopped Spread Spectrum R. K. Morrow, Jr. 29. On the Capacity of CDMA/PRMA Systems R. P. Hoefel and C. de Almeida INDEX

223

235 247

259

269

275 283

305 315

327

vii

7KLVSDJHLQWHQWLRQDOO\OHIWEODQN

PREFACE The papers appearing in this book were originally presented at the 8th Virginia Tech/MPRG Symposium on Wireless Personal Communications. This symposium, which is an annual event for Virginia Tech and MPRG, was held June 10-12, 1998 on the Virginia Tech campus in Blacksburg, Virginia. The symposium brings together leaders from industry and academia to discuss the exciting future of wireless and current research trends. The symposium has been an important part of MPRG's activities since the inception of the group in 1990.

As can be seen from the Table of Contents, the papers included in this book are divided into six sections. The first five of these correspond to symposium sessions and are devoted to the following topics: Smart Antennas and Diversity, Propagation, Interference Cancellation, Equalization, and Modulation, Coding and Networking. These session titles reflect current research thrusts as the wireless community strives to enhance the capabilities of wireless communications. This year an added feature of the symposium was the inclusion of externally contributed poster papers. Ten of these poster papers are included in this book as the sixth section.

The first group of contributions, consisting of five papers, relate to smart antennas and diversity. The first paper, Effects of Directional Antennas with Realizable Beam Patterns on the Spaced-Time Correlation, by T. B. Welch, M. J. Walker and R. E. Ziemer, considers the performance achieved with directional antennas at a base station. The authors consider the relationship between the bit error probability and the space-time correlation coefficient and illustrate the degradation in system performance that results when this correlation drops below one. The next paper, Frequency Reuse Reduction for IS-136 Using a Four Element Adaptive Array, is co-authored by J. Tsai and R. M. Buehrer. They present simulation results for two-element and four-element adaptive arrays and various frequency reuse factors. The third paper in this group, Pseudo-Blind Algorithm for SDMA Application by J. Laurila and E. Bonek, presents a novel pseudo-blind spacetime equalization algorithm for application to spatial division multiple access systems. Simulation results are presented which show performance as a function of various parameters including the number of antenna elements. The next paper, Integrated Broadband Mobile System (IBMS) Featuring Smart Antennas by M. Bronzel, J. Jelitto, M. Stege, N. Lohse, D. Hunold and G. Fettweis, explores the use of smart antennas to adaptively enable a trade-off between mobility and data rate. The authors present experimental data illustrating the relationship between beamwidth and delay spread. They point out that reduced delay spread allows the use of higher-order spectrally efficient modulation techniques. The final paper in this group, CDMA Smart Antenna Performance, is co-authored by M. Feuerstein, J. T. Elson, M. A. Zhao and S. Gordon. The strategy adopted is not to generate an optimum antenna pattern for each channel but rather operates on a per sector basis.

Each sector can be assigned patterns with the objectives of balancing traffic, managing handoff and controlling interference. Propagation issues constitute the theme of the next group of five papers. The first paper in this group is titled Wireless RF Distribution in Buildings Using Heating and Ventilation Ducts, and was authored by C. P. Diehl, B. E. Henty, N. Kanodia and D. D. Stancil. They present a novel method for distributing RF signals in buildings using heating and ventilation ducts as waveguides. The use of existing infrastructure is attractive and losses are low compared to direct propagation or a leaky coax. The following paper, Predicting Propagation Loss from Leaky Coaxial Cable Terminated With an Indoor Antenna by K. Carter, focuses on the development of a model for a leaky coaxial cable for use in the design of indoor microcell systems. The model presented in this paper exhibits a mean error of 2 dB with a standard deviation of 3dB. The paper, Building Penetration and Shadowing Characteristics of 1865 MHz Radio Waves, by M. Panjwani and G. Hawkins, presents experimental data for penetration loss and shadowing loss for seven buildings in urban environments in the Netherlands. They observed a high correlation between penetration loss and shadowing loss. This implies that penetration loss can be estimated with reasonable accuracy from shadowing loss, which is easier to measure. The fourth paper, Maximizing Carrier-to-Interference Performance by Optimizing Site Location by J. Shi and Y. Mintz, examines performance enhancement of a cellular system through the maximization of the carrier-to-interference ratio. This is accomplished by optimizing base station location according to traffic. The final paper in the propagation section, titled Azimuth, Elevation, and Delay of Signals at Mobile Station Site and co-authored by A. Kuchar, E. A. Aparicio, E. Bonek and J. Rossi, presents an analysis of channel sounder measurements made at 890MHz in a dense urban environment in Paris, France. Their objective is a thorough study of propagation mechanisms in their target area. Extensive data is presented and the results of their study will be incorporated into future propagation models developed by the authors. The third group of papers relate to interference cancellation as a technique for enhancing system performance. There are three contributions in this group. The first paper, A New Hybrid CDMA/TDMA Multiuser Receiver System by U. Baroudi and A. Elhakeem, considers a novel traffic control scheme which allows use of a dual-mode receiver. A decorrelating multi-user receiver is used in bursty

slots and a single-user receiver is used on non-busrty slots. In the following paper, Multiuser Multistage Detector for Mode 1 of FRAMES Standard by A. Boarin and R. E. Ziemer, the authors consider a multistage detector that combats both multiple access interference and intersymbol interference in code/time division multiple access systems. They show that the complexity of the detector is proportional to the number of users. The third and final paper in this group, Self-Organizing Feature Maps for Dynamic Control of Radio Resources in CDMA PCS Networks by W. S. Hortos, considers the application of self-organizing feature maps to the channel assignment in CDMA systems in which radio resources are regulated to minimize interference.

X

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

[L

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avr(ji)), then the excess cells Nexcess

(variable bit rate components)

should be directed to the other time slot (i.e. Ts2)· Therefore, the assignment of slots to a part or all of user traffic is done priori through an agreement between the base station and the user. The motivation for this traffic control is to make it practical for implementing the Multi-User Detection strategy at the receiver of the bursty traffic population.

112

(ji): means class (j) belongs to category (i.e. bursty and interactive categories); the average number of cells emitted from all bursty users to : the average number of cells emitted from class (i) bursty users to the ratio of a user’s average bit-rate to the basic bit-rate ; the ratio of a user’s bit-rate to the basic bit-rate ; the instantaneous number of users; and Nclass (i) : the number of classes of users in category (i). The motivation behind the choice that the system is working on a CDMA platform is the attractive features of CDMA technique over the others in the wireless environment. However, CDMA is suffering from two main problems. First, it is well known that the performance of CDMA techniques (in particular DSCDMA) is very sensitive to the processing gain

Hence, the user’s transmission rate is

not expected to be high enough to support multimedia applications nor a wireless ATM system where the objective is on-demand availability of bandwidth at peak rate as high as 10Mb/s [4]. The second problem is that CDMA technique is an interference-limited technique.

Multi-Code CDMA: Considering this heterogenous traffic, and for analysis convenience, we propose that various traffic classes should represent their bit rates as n multiple of a basic rate

or a 1/n multiple of

; n is an integer. For

practical rate users of rate R b /n, the generated cells will be buffered till their gross rate matches with Motivated by the desire to provide remedy for the first problem, the high rate classes (say class i, where the bit rate is of

), where each user needs to transmit with a rate greater than

be converted into

low-bit basic streams, where

rate equal to

, where

, the high-bit stream shall

. Each new low-bit basic stream has a bit

is the basic bit duration.

, where

is the bit duration of

the high-bit rate traffic coming from class i. Consequently, the low-bit rate stream traffic is packetized in cells as per ATM standards. This strategy introduces the concept of Multicode CDMA (MC-CDMA) [5] where each low-bit rate stream is spread with a different code.

3. Statistical Traffic Model The structure of the voice and video traffic is fairly complex due to the high correlation among arrivals [6].

113

Furthermore, the activities of such sources play a key role in modelling generated. The correlation between voice cells generated during a call can be modelled as an Interrupted Poisson Process (IPP) as shown in Fig. 2. This model is the simplest model and is widely used to model voice traffic. The transition from talk-

spurt (on state) to silent (off state) during a time period equal to cell transmission time on the channel occurs with the probability

and the transition from OFF to ON occurs with the probability

Consider-

ing this discrete event system, the ON and OFF states periods are geometrically distributed with the mean respectively. The total number of cells generated from all active calls (during the ON

period) follows a Bernoulli distribution. For traffic class i, the probability Pi (n) that n out Ni voice sources are in the ON state, where Ni is the

number of independent active voice sources is derived as

where

which is the average activity probability. When I independent different classes of

users are multiplexed, the aggregated cell arrivals are governed by the number of voice sources in the ON state from all types of users. Having assumed that all classes of traffic users been are independent identically distributed random variables (iid), the compound traffic distribution of m cells generated by all active calls of all I classes is given by

4. The Decorrelator Receiver The decorrelator receiver uses a linear transformation to obtain an estimate of the transmitted symbol. It is

basically composed of two stages: matched filter bank and then followed by the linear transformation process. It is easy to represent the matched filter outputs [7] in a vector notation as shown in eq. (5).

114

From (7), we observe that the solution (i.e. the bit vector matrix

) can be obtained by inverting the correlation

which is invertible in most cases of interest. Hence, the linear transformation process is done

by

This results in

This linear detector has many attractive features. First of all, its computational complexity is much reduced comparing with the optimum receiver. The compositional complexity of the decorrelator detector increases linearly with the number of users (i.e. O (K)). Further, the linear decorrelator receiver exhibits the same degree of near-far resistance as the optimum multiuser detector [3]. In addition, when the users energies are unknown, the decorrelator receiver is the optimal approach. However, a significant limitation of this technique is the computational complexity due to the inversion of the correlation matrix [8] where entries depend on the number of active users, signature sequences, and the delays of the users. Further, any change in one of these parameters changes the correlation matrix and consequently a need for updating the multiuser detection process. Moreover, the uncertainty in the actual number of active users is another serious problem that might degrade the system performance very severely [9]. Fortunately, by adopting our traffic control policy, the last problem is resolved.

4. System Performance Analysis In the following, we analyze the performance of the proposed CDMA/TDMA system employing the above traffic control. We consider that there are four classes of bursty traffic users each of transmission rate Rj (j=1,2,3,4). These rates are multiple of Rb. On the other hand, we assume that the interactive traffic category has two classes of users. In this work, we study the effects of (MC-CDMA) intracell interference on the reverse link (i.e. mobile to base). Each low-bit rate will be spread with a different gold code. The information bit are modulated using the binary phase shift keying (BPSK). Here, we just consider the additive gaussian interference. We can present the received signals from both categories of traffic as follows. The received signal in Ts1

(i.e. r1(t)) is composed of just a part of the

bursty traffic as explained before. On the other hand, r2(t) received in Ts2 is composed of two parts: 1)

115

the interactive users signals and 2) the excess traffic from the bursty users. Thus,

n(t) is an additive white gaussian noise with two sided spectral density

kj is the number of

active users belong to class j in the interactive traffic category and it is a random variable. Then, the out-

puts of the correlator filters will be processed as explained before, where r 1 (t) is processed by the decorrelator multiuser receiver, while r2(t) is processed by the conventional detection strategy (single-user).

In the following, we shall evaluate the performance of the proposed approach using the ATM-cell errorrate criterion assuming synchronous transmission. According to the above traffic control, the bit error

probability of the bursty traffic is composed of two portions: 1) bit error performance for the cells using Tsl and detected by the multiuser receiver, and 2) bit error performance for the excess traffic cells using

Ts2 and detected along with the interactive cells by the single-user receiver. Then, these performances are averaged over the whole bursty traffic population. Following eq. (4), it is easy to figure out the probability distributions for both bursty traffic components. Considering first the bursty traffic category, we get the following probability distribution:

where k is the number of cells and const is the value when eq, (4) is evaluated for

On the

other hand, the probability distribution of the interactive traffic is straight forward from eq. (4). From the

above discussion, we can easily get the average bit error rate for the bursty traffic as derived in eq. (10).

is the average bit probability of error for the whole category of bursty traffic.

is the probability

of error given [3] by eq. (11) when there are k cells (i.e. average components of the bursty traffic) in Ts1.

P1(k) is the probability distribution of the bursty traffic as defined in eq. (9). Sig is the maximum

116

expected interactive traffic low bit rate's cells. Navr1 is the average number of low-bit rate’ cells of the

bursty traffic. Nexcess is the number of excess bursty traffic low bit rate’s cells.

is the bit probability

of error given [10] by eq. (12) when there are m (i.e. m=l+j) cells in Ts2 .

where the Q-function

where PG is the number of chips spanned one bit period (i.e. PG = Tb/Tc

).

Following the same argument as above, we obtain a similar equation for the performance of the interactive

traffic class;

5. Results We compare our proposal to the conventional system where both categories of traffic are using the whole available bandwidth and the received signals are processed by the conventional single- user receiver. Fur-

ther, both systems (proposed and conventional) are compared (when applicable) to the single users system

(where there is just one user) which is considered the lower bound for any Multiple Access System. The common criterion used to evaluate the performance of the multiple access technique is the packet error

rate. Considering the ATM structure for the transmitted packets, where each packet (called cell) is composed of 53 bytes, the packet error rate is

where Pc is the probability of correct bit-decision, and n is the number of bits in the ATM cell. The results

presented here is of the MC-BPSK-CDMA/TDMA hybrid system. Fig. 3 illustrates how the packet error rate is evaluated. First, for every set of bursty traffic parameters, and for every fixed number of active

117

users, we obtain all possible combinations of cells that each class of users might send and each user will split its stream of bits into certain number of low-rate streams. Furthermore, each low-rate stream will be spread using a set of pseudo-random codes (Gold codes) where the ratio of T b/Tc = L = 1023. Though, in the case of two slots CDMA/TDMA, the bursty rate should be higher than 1/T b . Here, we assume that the bursty rate at each time slot is 2/T b (i.e. 20 Kbps). Hence, to maintain the same bandwidth, the processing gain should be half the one used in the wide CDMA (i.e. conventional system in our case), that is 512. Therefore, the bit energy Eb is the same for both systems (i.e. conventional and proposed). Then, the bit error rate calculated using the classical multiuser detection and single user detection are averaged over ail the possible combinations assuming that each combination is equiprobable. This is very important, because each class of traffic has different traffic characteristics. Hence, to obtain a fair and clear picture of the system performance, the results should be averaged over all possible combinations.

Table I summarizes the traffic characteristics used in evaluating the proposed multiple access protocol. The basic rate is assumed the same for all classes of both traffic categories that is Rb = 10Kbps. Consider-

ing the bursty traffic classes, we let each class to have a range of transmission rates such that the proposed

system will be tested under more reliable parameters. Therefore, each bursty traffic class has a minimum bit rate, maximum bit rate and an average bit rate. For the purpose of investigating the capacity of the pro-

118

posed protocol, we study the new system on what we call combination of rates. In other words, the minimum-bit rate combination is where the system is tested assuming that all classes are operating on their lowest bit rates. The maximum-bit rate combination is where the system is tested assuming that all classes

are operating on their maximum bit rates and the same thing applies for the average rates. Further, each class of traffic has its own transmission activity (i.e. pi,j, (i,j) means class j belongs to traffic category i).

The system studied, here, has a maximum of 60 bursty users and 13 interactive users.

It is worth to emphasis that the number of active users admitted into the system is not necessarily equal to

the number of mutual cells processed by the system, because the number of cells depends on other parameters and not just the number of active users. The instantaneous number of cells is given by eq. (15).

Fig. 4 shows the packet error rate for the proposed approach as well as the conventional and the single user

systems. It is clear that the packet error rate performance is very similar to the conventional one which is irreducible. This is due to the fact that in the original proposed approach, the user should stick to a certain

average bit rate and as such at high rate transmission, the cells received by the multiuser detector is just a small portion of the bursty traffic. Hence, when we average the performance of both receivers over the

whole traffic population, the cells received by the conventional receiver will have more statistical weight (binomial distribution). To overcome this problem, we modify the approach such that more cells shall be received by the decorrelator multiuser receiver, and this can be accomplished as follows. If the system per-

formance is observed to be worsened by employing the original version of the proposed approach, then

119

both the user and the base station negotiate again a new average bit rate which should be higher than previ-

ous average bit rate

and so on. Fig. 4 shows an improvement in the system performance

when this modification is applied. Further, the system capacity (i.e. the number of users the system can handle in an acceptable packet error rate) also has been improved as shown in Fig. 5.

6. Conclusions The new hybrid CDMA/TDMA system has been presented. We examined the new system under a wide range of expected traffic characteristics (bit rate, transmission activities, etc.) for the future wireless networks. The results show the superiority of the proposed system compared with the conventional one. Further, this improvement in the performance is attributed to the novel traffic control scheme that could

balance between the traffic load on each detection algorithm and makes the implementation of the decorre-

lator receiver more practical.

7. References [1] S. Vedru, “Minimum Probability of Error of Asynchronous Gaussian Multiple-Access Channel”, IEEE Trans. on Inform. Theory, Vol. IT.-32, No. 1, Jan. 1986, pp. 85-96. [2] A. Duel-Hallen, J. Holtzman, and Z. Zvonar, “Multiuser Detection for CDMA Systems”, IEEE Personal Communications, Vol.2, No. 2, April 1995, pp. 46-57.

[3] R. Lupas and S. Verdu, “Near-Far Resistance of Multiuser Detectors in Asynchronous Channels, IEEE Tran. Commun., vol. Corn-38, no. 4, April. 1990, pp. 496-508. [4] “Special Issue on Wireless ATM,” IEEE Personal Commun., August 1996. [5] Chih-Lin I, and R. D. Gitlin, “Multi-Code CDMA Wireless Personal Communications Networks,” Proc. of ICC '95, pp. 1060-1064. (6] J. J. Bae, and T. Suda, “Survey of Traffic Control Schemes and Protocols in ATM Networks,” Proc. IEEE, vol. 79, no. 2, Feb. 1991, pp. 170-189. [7] J. G. Proakis, Digital Communications, 3rd ed., New York: McGraw Hill, 1995. [8] M.J. Juntti, and J.O. Lilleberg, “Implementation Aspects of Linear Multiuser Detectors in Asychronous CDMA Systems”, proc. of ISSSTA ’96, pp. 842-846. [9] E. S. Esteves and R. A. Scholtz,“ Bit Error Probability of Linear Detectors in the Presence of Unknown Multiple Access Interference”, Proc. of IEEE Globecom ‘97, pp. 599-603. [10] M. B. Pursley, “Performance Evaluation for Phase-Coded Spread-Spectrum Multiple Access Communication - Part I: System Analysis”, IEEE Tran. Commun., vol. Com-25, no. 8, Aug. 1977, pp. 795-799.

120

121

7KLVSDJHLQWHQWLRQDOO\OHIWEODQN

1 Multiuser Multistage Detector for Mode 1 of FRAMES Standard Adrian Boariu and Rodger E. Ziemer1 Electrical and Computers Engineering Department University of Colorado at Colorado Springs 1420 Austin Bluffs Pkwy Colorado Springs, CO 80907, USA [email protected], [email protected]

Abstract In code/time division multiple access (C/TDMA) systems, both multiple access interference (MAI) and intersymbol interference (ISI) arise. MAI is present due to the CDMA format while the ISI is due to the channel multipath. In this paper we present a type of multistage detector that overcomes these problems and in addition has complexity proportional to the number of CDMA users, is computationally efficient, and is suitable for pipeline implementation that allows fast data processing. The simulation results are compared to the single user bound for the average bit error probability (BEP).

1. Introduction The FRAMES (Future Radio Wideband Multiple Access System) standard defines third generation multiple access schemes for personal communication systems (PCS). Mode 1 [1] is wideband TDMA. The frame is 4.615 ms in duration and has eight slots. One option of Mode 1 is the CDMA feature. Each slot can bear a number up to eight CDMA users. The slot format includes two data bursts, separated by a midamble, and ending with a guard interval. The spreading factor is 16 chips/symbol and modulation employed is QPSK. The TDMA feature guarantees that all K users arrive synchronously at the receiver. Subsequently we will analyze only the uplink (mobile to base station) because this is the one that often limits the capacity of the system. 1 This work was supported by the BMDO and managed by the Office of Naval Research under Contract N00014920-J01761/P0004

2. The detector description In a mobile radio environment the signal is subject to fading and multipath propagation effects. As a result, the receiver must deal with intersymbol interference (ISI) due to the channel multipath in addition to the multiple access interference (MAI) due to the CDMA feature. Perfect estimation of the channel via midamble is assumed. COST-207 defines the power-delay profiles and Doppler power spectrum for the mobile channel [5].

Previous solutions to this problem can be found in [2]. The main approach is to estimate

the entire data burst for all users at once, using well known detection methods: decorrelation, minimum mean-square error, block feedback, etc. The main drawback is that such detection

methods must deal with a large matrix that must be inverted or Cholesky factorized.

Our solution is different [4]. For the coherence of the paper we briefly describe the

method. Figure 1 shows the ISI introduced by a bad urban (BU) radio channel, where is the equivalent impulse response, hc and c being the impulse response of the channel and the spreading codes, respectively. Since there is a guard interval at the end of each time slot, there is

no ISI over the [0, Ts) interval. Hence, we can detect the first symbols of all users by performing the integration of the received signal over one symbol interval. This is the first stage. In the next stage we integrate the signal over two symbol intervals and we use the estimates from the previous stage to cancel the ISI. The same procedure is used for the third stage. Thus, succeeding

stages add more energy in the detection process due to the extended integration interval.

124

Let

be the output of the matched filters at the kTs moment with the received signal r(t) integrated over a length of iTs, N being the data package length, and let

be the cross-correlation matrices for lag kTs with the received signal integrated over a length of iTs. For the other lags the cross-correlation matrices are zero. The equations that describe the stage detector for BU case are

For each of these stages, the feedback detector based on Cholesky factorization with xk detected successively (due to fi triangular form) is employed. A final stage based on the method described in [3] is added to improve the performance. The proposed detector will have

four stages for the BU situation and three stages for the typical urban (TU) case. The single user lower bound can be computed based on

where W is the number of channel paths,

are the eigenvalues of the 1/2 rc diag(P)

matrix, and the P vector contains the power delay profile samples while rc is a matrix which is formed by the autocorrelation function of the specific user spreading code [6].

3. Simulation results

Figure 2 presents some simulation results for the BU and TU channel cases. With each stage employed the performance is better. For reference, simulation curves using the block feedback detector (using Cholesky factorization) [2] are also depicted. We see that the detector

125

presented in this paper gives an improvement of around 0.75 dB over the block feedback detector.

126

4. Conclusions A multiuser multistage detector has been presented that estimate data on a symbol-by-

symbol basis and not the entire package of data at once as block feedback detectors do. The BEP performance improves with the number of stages employed. The lower bound for the BEP is less than 0.5 dB off from the simulation results for the eight-user case. The method uses matrices

having dimension equal to the number of users and is independent of the data burst length. It is

suitable for pipeline implementation because the stages can run independently.

References [1] T. Ojanpera, et al, “Comparison of multiple access schemes for UMTS”, Proc. IEEE Veh.

Technol Conf., pp. 490-494, May 1997

[2] P. Jung and J. Blanz, “Joint detection with coherent receiver antenna diversity in CDMA mobile radio systems”, IEEE Trans. on Veh. Technol, vol. 44, pp. 76-88, Feb. 1995

[3] M. K. Varanasi and B. Aazhang, “Multistage detection in asynchronous CDMA communications”, IEEE Trans. on Commun., vol. 38, pp.509-519, April 1990 [4] A. Boariu and R. E. Ziemer, “Stage detector for C/TDMA systems”, accepted for

publication at Development & Application Systems 1998 Conference, Suceava, Romania [5] “COST-207: Digital land mobile radio communications”, Final Rep., Luxembourg, Office for Official Publications of the European Communities, 1989 [6] J. Omura and T. Kailath, “Some useful probability distributions”, Technical Report, no. 7050-6, Stanford University, Sept. 1965

127

7KLVSDJHLQWHQWLRQDOO\OHIWEODQN

2 Self-organizing Feature Maps for Dynamic Control of Radio Resources in CDMA PCS Networks William S. Hortos Florida Institute of Technology, Orlando Graduate Center, 3165 McCrory Place, Suite 161, Orlando, FL 32803 ABSTRACT The application of artificial neural networks to the channel assignment problem for cellular codedivision multiple access (CDMA) cellular networks has previously been investigated. CDMA takes advantage of voice activity and spatial isolation because its capacity is only interference limited, unlike time-division multiple access (TDMA) and frequency-division multiple access (FDMA) where capacities are bandwidth-limited. Any reduction in interference in CDMA translates linearly into increased capacity. To satisfy the demands for new services and improved connectivity for mobile communications, small cell systems are being introduced. For these systems, there is a need for robust and efficient management procedures for the allocation of power and spectrum to maximize radio capacity. Topology-conserving mappings play an important role in biological processing of sensory inputs. The same principles underlie Kohonen’s self-organizing feature maps (SOFMs), which are applied to the adaptive control of radio resources to minimize interference, hence, maximize capacity in direct-sequence (DS) CDMA networks. The approach based on SOFMs is applied to published examples of DS/CDMA networks. Results of the approach for these examples are informally compared to the performance of Hopfield-Tank algorithms and genetic algorithms for the channel assignment problem.

1. INTRODUCTION During network system planning, radio test data are used to estimate radio propagation profiles and user demands for mobile cellular service. Based on these stationary estimates, a so-called compatibility or interference matrix and traffic-demand vector for the planned network are derived. The channel

assignment problem (CAP) is the allocation of the required number of channels to each cell to meet traffic demands subject to interference conditions. This network representation has been used to develop energy functions for Hopfield-Tank neural networks (HNNs) as well as genetic algorithms (GAs) for the CAP in cellular radio networks 1,2. Unfortunately, in systems of smaller or non-uniform cells, used to

provide personal communication services (PCS), the rapid dynamics of cell-to-cell handoff, changes in service grade, and transmit power control render algorithms based on stationary estimates inadequate. Channel assignment to the network cells can be either static (SCA) or dynamic (DCA) 3. In SCA a

set of channels is permanently assigned to each cell, or, coverage area of a base station. In DCA the channels are assigned to cells on a call-by-call basis to improve network capacity and spectral utilization. The DCA approach is suited to non-uniform and time-varying traffic demands. To adapt to changing

network conditions, the past observed dynamics of the network can be used to update radio resource allocation (RRA) procedures. The past history of direct-sequence (DS)/code-division multiple access

(CDMA) networks in the service areas of the base stations allows iterative updates of interference statistics, limiting new measurements only to recent network events. The RRA for a new service request, i.e., the assignment of power, base-station antenna, and call activity monitoring to provide a channel for the call, depends globally on the position of the mobiles as well as on previous assignments that may interfere with the contemplated RRA.

Extensions of

Kohonen’s self-organizing feature map (SOFM) are constructed to adjust the spatial and temporal configurations of mobiles and base stations, called the “state” of the network. The approach is based on a composite mapping, from radio resources into the set of interference levels in the service areas of the base stations, then from the interference set into the set of CDMA channels. In 1991 Kunz proposed the first Hopfield neural network (HNN) model for solving SCA in cellular FDMA networks 1, 4 . Smith and Palaniswami 5 are first to apply a method based on Kohonen’s SOFM for the SCA problem in FDMA and TDMA networks. HNN convergence has been improved recently using auxiliary conditions on the energy function’s coefficients to ensure hill-climbing (HC) on the hypercube of feasible solutions

2, 5

. Further convergence improvement has been realized for hill-

climbing HNNs (HC-HNNs) via Abe’s non-uniform integration step size 2, 6. The effect of sectored cells on the convergence of HC-HNN and genetic algorithms (GAs) for the CAP has been examined in DCA 2. 2. COVERAGE, CAPACITY AND INTERFERENCE IN DS/CDMA PCS NETWORKS The primary goal of DS/CDMA PCS network design is management of the three-way tradeoff among call quality, coverage and capacity in order to maximize overall system performance economically. The advantages of DS/CDMA derive from soft/softer handoff, power control, activity monitoring, and reuse of the same spectrum in all cells and sectors. The cell coverage for the reverse link in a DS/CDMA PCS network is determined by radio design and operating environmental factors, as indicated by the following simplified link budget:

where Pm is the maximum mobile station transmit power in dBm, is the receiver sensitivity in dBm, Gb (Gm) is the base station (mobile) antenna gain in dBi, Lp is the vehicle penetration loss in dB, Lc is the cable loss in dB, GHO is the handoff gain in dB,

is the lognormal fade margin in dB, and

is the

channel loading. The transmission parameters Pm, Gm, Lp, and Lc are established at network installation or not adaptive during network operation. Gb can be modified through antenna selection. Actual mobile transmit power Pm(t) is controlled to limit interference.

The latter two can be considered radio

resources of the network. The remaining terms in (1) depend on propagation and traffic conditions. The

130

following equation simplifies the relationship between CDMA capacity and radio parameters on the

reverse link:

where m is the number of simultaneous users per cell or area, B/Rb is the processing gain (PG), v is the voice activity factor, r is the frequency reuse efficiency, and g is the sectorization gain. Capacity improvement of a CDMA system over TDMA and FDMA comes from the universal frequency reuse capability, soft handoff, power control, and voice activity monitoring. For CDMA, the frequency reuse efficiency r is represented by the ratio of interference from the service area of the given base station to the total interference from that service area and all neighboring service areas. Thus, r is sensitive to propagation conditions. Due to power control, the signal as well as intra-cell interference remains the same on the reverse link, so that r is only inversely proportional to inter-cell interference. Ambient noise establishes the required received signal power at the cell site, and thus establishes the user's power or 7 cell radius for a given transmit power level. In a reuse system, capacity is alternately defined by Lee as

where M is the total number of channels, s is the number of sectors and reuse factor r can be expressed as

where d is the co-channel cell separation, that is, the minimum distance, and R is the cell radius. In a DS/CDMA system, the broadband frequency channel can be reused in every adjacent cell, so that r is close to 1. In a practical sense, d=2R, i.e., all available frequencies can be used in each cell, so from (5), r = 1.33. Radio capacity is also based on (3) and (4). However, in DS/CDMA, r is fixed but M, the total number of available channels is variable, and depends on the level of interference, i.e., M(I). Interference, in turn, is the result of the radio resources assigned to each call. As shown in Figure 1, interference at the mobile arises both from the home cell and the adjacent cells, all operating on the same frequency f1. The received carrier-to-interference (C/I) can be written as

where Eb is the energy per bit, I0 is the interference per hertz

where N0 is thermal noise per

hertz), Rb is the information rate, B is the bandwidth per channel, and B/Rb = PG. Processing gain is used to overcome I, and, hence, determine the number of channels that can be created. Given two sets of values for Eb/Io and Rb/B, corresponding C/I values are found from (6). From Figure 1, C/I at the mobile location A can be used to for a worst-case scenario. If

131

For an IS-95 system, let B = 1.228 MHz and Rb = 9.6 kbps, then PGdB = 21 dB. Voice quality at a -2

frame error rate (FER) of 10 typically corresponds to Eb/I0 = 7 dB. Given the values of PG and Eb/Io,

C/IdB = 7 - 21 = -14 dB or C/I = 0.03981. From this value of C/I, the number of traffic channels mi in each coverage area of Figure 1 can be obtained by

where mi and

are the number of channels and power level at cell i, respectively. Solving (7), one has

In the case of a single radio cell, that is, channels/cell. In the case

3. INTERFERENCE, TRAFFIC DEMAND, AND RADIO RESOURCE ALLOCATION The following conditions summarize the interference constraints in cellular network operation.

1. The co-channel constraint (CCC) is that the same channel cannot be assigned simultaneously to certain pairs of radio cells. The CCC is determined by the co-channel interference (CCI), which, in turn, depends on the interference control methods applied in coverage areas of the base stations.

132

2. The adjacent channel constraint (ACC) is that channels adjacent in their domain's distance metric (frequency, time slot or code) cannot be assigned to adjacent radio cells simultaneously. The ACC is related to the channel reuse factor r. 3. The co-site channel constraint (CSC) is that any pair of channels assigned to a radio cell must be at a minimum distance in their domain. In DS/CDMA, this distance depends on the interference level produced by the radio resource assignment used in the service area of each base station. For a network of N base stations, the constraints are commonly described by an NxN symmetric matrix, called the interference matrix C. Each off-diagonal element cij in C represents the minimum separation distance between a channel assigned to cell (or sector) i and a channel assigned to cell (or sector) j. The CCC is represented by cij =1, while the ACC is represented by cij = 2. Setting cij = 0 indicates that base stations i and j are allowed to assign the same channel to users in their service areas. In DS/CDMA,

with control. Each diagonal element cii represents the minimum separation

distance between any two channels assigned to cell (or sector) i. This is the CSC and

is always

satisfied, provided that, in sectored networks, sectors are equivalent to cells. Since DS/CDMA capacity can only be increased by reducing other-user interference, a departure can be taken from the model of a two-dimensional interference matrix for assigning N base stations a fixed M channels. In a DS/CDMA network, each base station i establishes a radio coverage area that can carry any of mi(I) channels, where

the total number of channels in the network, and mi(I)

represents the local capacity of the coverage area i. As discussed by Gilhousen, et. al, using the elements of directional antenna sectorization, voice activity monitoring, and reverse-link power control, the system can adaptively regulate intra-cell and inter-cell interference, I, and, thereby, manage channel capacity 8. Considering these elements as network radio resources suggests representing their effect on network capacity by a composite mapping,

on a 3N-dimensional lattice,

into the set M of

vectors of channel capacities in the coverage areas of the base stations, where

represents the set of cell sectorization values in any coverage area; activity monitoring in each area;

is the set of states of voice

is the set of discrete power control levels;

is the real bounded

interval of all possible interference values; and M is the subset of the set of N-dimensional vectors of non-negative integers, whose ith component is the call capacity at base station i. This mapping relates an RRA to network capacity, through the interference level that the assignment produces. Traffic demand for channels, in a network of N base stations, is represented by an N-vector called the traffic demand vector T. The component ti of T represents the number of active channels (new calls and

133

handoffs) to be assigned to the coverage area of base station i. Let qik denote the kth active user

assigned to station i. Then, the constraints described by the interference matrix are given by

Dynamic network conditions are modeled by time-varying entries in C and T, and lead to dynamic

radio resource allocation (DRRA) problems with transient traffic demands and interference thresholds.

4. THE SELF-ORGANIZING FEATURE MAP APPROACH Learning in a self-organizing system is due to the processes of competition and inhibition. Presented with an external input, neurons compete with each other to claim the input. Synaptic weights are adapted to reflect the result of the competition. The idea of the radio resources in a DS/CDMA network 9

“competing” to be assigned calls suggests an approach is based on Kohonen’s SOFM . The Kohonen

map is modified to solve discrete-space optimization problems among the lattice of radio resources. The

approach is first developed for SRRA problems, then later extended to DRRA problems. All feasible solutions to the RRA problem, as formulated by the HNN approach, lie at the vertices of a 3N-dimensional hypercube

N is the number of base stations, and

is the total number of available radio resource combinations. The image of these vertices under

also

intersects the constraint hyperplane defined by the interference matrix, traffic demand vector and interference constraints (9) resulting from the RRAs. Since ti is an integer for all i, the

-image of the

radio resource constraint set is an integral polytope. Variables on this hypercube, are defined as

which, for convenience, are denoted by Xj, r. Let X denote the NxSxVxPdimensional array of these variables. Since the range of values for each radio resource and local

interference level is assumed to be bounded, each range can be normalized to the interval [0,1]. Without

loss of generality, the set of radio resources and its image under

in the interference set are both

contained in unit hypercubes. Suppose one approaches a vertex continuously from within the unit hypercube, starting from a point on the constraint hyperplane and inside the hypercube, which represents a feasible, non-integer solution to the RRA problem. One continues to move along the constraint hyperplane, gradually approaching a feasible vertex. The continuous variable in the interior of the

hypercube is denoted by wr,

Thus, for a metric Q, and W defined as the array (wr, j), Q(W) = Q(X) at the vertices. The value wr, j denotes the probability that the variable in the r, j (actually position j.

of array X is activated. Self-organization is applied to the array of probabilities or synaptic weights, W.

134

5. THE SOFM ALGORITHM FOR PCS NETWORKS The structure of the discrete-space form of the SOFM is shown in Figure 2, where W is expressed as a

RxN matrix, with vector ô replaced by a scalar index

into the lattice of allowable RRA

vectors. It consists of an input layer of N nodes and an array of R output nodes. The output nodes correspond to the solution array of capacities resulting from each RRA, while the input layer denotes the

N coverage areas in the PCS network. The weight connecting input node j to node (corresponding to r) of the output array is given by

. A cell in which an RRA is required is presented to the

network through the input layer at node j. Physically, a new call or handoff is presented to the PCS network in the service area of base station j. The nodes of the output layer compete with each other to determine which column of the solution matrix (channel produced by the RRA) accommodates the input

vector with the least impact on quality or cost. The weights are then adapted to indicate the RRA decision using the neighborhood topology.

Consider the case where an RRA is required in coverage area j*. An input vector x is presented with a “1” in position j* and “0” elsewhere. For each node r ofthe outer layer, Vr,j* , the cost to theobjective function ofassigning r to coverage area j*, is computed. The cost potential Vr,j* of node r for a given

input vector

is defined as

where the degree of interference caused by the RRA is represented by the weight Pj,

i, d+1 ,

called the

cost or proximity indicator and

the distance in the channel capacity domain between

the images of RRAs r and s.

then the interference cost should be at a 135

maximum, with cost decreasing until the two channels are sufficiently separated so that interference is below a threshold. Elements of the proximity indicator array P are recursively defined from C as

The dominant node, m0, of the outer layer is the network node with minimum cost potential for a particular input vector, that is,

for all nodes r and fixed j*. The neighborhood of the

dominant node m0, is the set of nodes

ordered according to the ranking of the cost

potentials corresponding to each node, that is,

where

is

the size of the neighborhood in the SOFM network for the coverage area of base station j*.

Unlike Kohonen’s original formulation 9, where ordering corresponds to the physical structure of the

network, the neighborhood of the dominant node is not defined spatially here, but depends on the ranking of the computed cost criterion of each output node for a given input vector. Thus, dominant

nodes and their neighborhoods are determined by competition according to the cost criterion; the weights are modified according to Kohonen’s rules within the dominant neighborhood. The size of the dominant

neighborhood depends on which base station is receiving input.

At the completion of Kohonen weight updating, the weight array W may be moved off the constraint hyperplane, resulting in an infeasible solution. In the next stage of the SOFM, the weights of the nodes outside the dominant neighborhood are organized around the modified weights, so that W remains a feasible solution. This stage can be performed by a HC-HNN. Expressing the array W as a vector w, w is considered the vector of states of a continuous HNN. The HNN performs random and asynchronous

updates on w, excluding the weights in the dominant neighborhood, to minimize the energy function:

where

and

is the projection onto the constraint hyperplane given by

is the identity operator. The energy function (14) is expressed in terms of a

solution vector x„ constructed from the solution array X, by reordering the X i,r where r = (s, v, p), in an ordering of four-integer indices. In terms of x, the demand constraints can be expressed as Dx = T,

where T is the demand vector and array D consists of N rows of 1’s and 0’s. Some argue that the HNN need not be an HC-HNN discussed in10, since the weights need only intersect the constraint hyperplane and, as such, there is no optimal point5. However, reducing the time to minimize the energy function

136

(14) speeds the update of non-dominant weights and, consequently, the convergence rate of the global weight adaptation. It may even improve the initial configuration of weights when the next service

requirement is input to the SOFM network. The next randomly selected requirement is input to the SOFM network to begin a new update stage of the algorithm, where the dominant node and its neighbors

are determined and their weights modified. This procedure is repeated until the SOFM weights stabilize

to a feasible 0-1 solution which is a local minimum of the RRA problem. During convergence, the magnitude of the weight modifications and the size of the neighborhoods are decreased. Initially, the size of the neighborhood for each subarray of W, given by

large, but is decreased linearly until

is

the caller demand for base stationj. Since the SOFM weight

modifications depend on the order in which the resource requirements are input, the approach is inherently stochastic.

In this form, the network must be run repeatedly to arrive at different local

minima and may be subject to stability problems that plagued earlier HNN algorithms for CAP 1,6,10. The following SOFM algorithm can be applied to the SRRA problem in DS/CDMA networks. 1. Initialize the network weights as

under any

is the maximum network capacity possible

This yields an initial feasible, although non-integer solution.

2. Randomly select a new radio resource requirement for a base station. Represent this requirement as

the input vector x. Find the position j* (base station coverage area) which is active, i.e., x j* = 1. 3. Compute the potential V r,j* for each index r in the output layer array according to (10).

4. Determine the dominant node, m0, by competition such that its neighboring nodes

and identify

is the size of the neighborhood for input at j*.

5. Update the weights in neighborhood of dominant node according to the rule

a modification of Kohonen's slow updating rule, α and σ are monotonically decreasing, positive

functions of sampled time, γ is a normalized weighting vector used in tie-breaking at a node. For all weights outside the updated neighborhood,

Weights are updated as

6. The weights may no longer lie on the constraint hyperplane, so an HC-HNN is applied to return to a feasible solution. The vector w is modified around SOFM weight adaptations so that Dw = T.

7. Repeat Step 2 until radio resource requests at all base stations have been selected as input vectors to

the SOFM network. This forms one period of the algorithm. Repeat this procedure for K periods. Parameters in Step 5, α and σ are decreased according to some monotonically decreasing function.

137

8. Repeat Step 2 until

this condition is considered stable convergence of the weights

for a given neighborhood size. Decrease the neighborhood sizes 9. Repeat Step 8 until

linearly for all j.

for each base station j, j = 1, · · ·, N.

The weighting vector is a heuristic used to damp oscillations in SRRA algorithm updates11:

Each component in the vector γ is normalized. The SRRA SOFM parameters are selected heuristically5:

The SOFM approach is suited to the dynamic rearrangement of radio resources. The DRRA problem begins at the point in the SRRA solution where a new call cannot be assigned without a rearrangement of existing RRAs. A time-varying demand vector T(n) and interference constraints are satisfied when D(n)x(n) = T(n) at sample time n. Each sample period represents the arrival of a single call or multiple calls to the network and, during that period, input vectors, corresponding to the areas in which calls are

placed, are presented to the SOFM network at a rate determined by the demand distribution at that time. Since feasibility is always restored during the second stage (Step 6) of the SOFM, any rearrangement of

the existing calls to enable a new call is automatic. If no rearrangement is possible, either the SOFM may not converge to a feasible set of RRAs, or a feasible rearrangement may be found by allowing the

interference to increase above pre-determined levels. Higher interference levels could, in turn, diminish channel capacity. In this case, the call can be blocked and the previous state of the system reinstated. To improve SOFM convergence, Steps 5 and 6 for updates of non-dominant weights can be made more robust by adopting Abe's approach 10 to ensure that an HC-HNN only leads to feasible RRAs that

are stable points of the system of update weights, i.e., satisfy D(n)x(n) = T(n) at sample time n. A piecewise linear saturation function can replace the exponential in Step 5. In Step 7, faster updating can

be obtained based on Abe’s convergence acceleration for HC-HNNs to optimize integration step sizes, now applied in only K = 1 period 6. With little time to wait for stable convergence of the weights, once the SRRA is completed, the DRRA algorithm omits Step 8. The vector

of neighborhood sizes

also be initially set to the demand vector T(n) at sample time n, at the start of the current update. 138

can

6. SIMULATION OF THE SOFM FOR RRA PROBLEMS The performance of the discrete-space SOFM for SRRA and DRRA is examined through simulations of network examples considered by Kunz1 , Lai 12, and Funabiki, et. al.13. Interference matrices and traffic demand vectors for these examples, shown in Figure 3, represent cellular networks with non-

homogeneous traffic. Homogeneous traffic is modeled by demand vectors with equal components. Time-varying traffic is approximated by the cyclic rotation of T or periodic replacement with a new

vector during the simulations. In the simulations, the radio resources at each base station form a triple from the selection of cell

sectorization: 1 (omni), 3 (120°), 6 (60°) sectors in a coverage area; voice activity monitoring: 0 for “off” or 1 for “on”; and power control level: 0 for no power control, 3 levels, 6 levels, 10 levels. Each triple of resource values r = (s, v, p) is mapped to an interference estimate Ir based on statistics derived from measurement studies of microcellular networks such as

14, 15

. The interference establishes the

DS/CDMA reuse factors, and, hence, determines the number of available channels. The uniform cost vector for each r is c = (1, 1, 1). Individual RRA cost terms r.c are summed over the number of active

base stations in the network. This sum is added to the quality or cost criterion. Simulations were performed on a personal computer using adaptive learning models in MATLAB’s

Neural Network Toolbox, which were extensively modified to implement the SOFM outlined in Steps 1 through 9. Algorithm performance is measured alternately on the basis of the average probability of call blocking (ABP) or the total active calls in the network (transient capacity), along with the average

number of iterations (ANIs) required for asymptotic convergence, based on a prescribed error value An omnicell network of N= 21 cells is considered with the interference matrix in Figure 3(c) with M = 221, the minimum number of channels. Entries in T are set initially to t i = 4 for each cell i and the SRRA SOFM run with K= 100 different initial states. Parameter t i is incremented to 6, 8 and 16. Each

case is repeated for K = 100 different initial states and the results averaged. The ABPs for the cases are 0.010, 0.024, 0.089, and 0.117, respectively. The corresponding ANIs are 38.4, 49.2, 102.5 and 213.9. Resource allocations for each initial state at each base station area are set to minimum values, (1,0, 0),

and then allowed to adapt to the demand vector. This may account for large ANIs in the test cases. Cell sectorization shows the greatest sensitivity to a uniform traffic increase, and is at six 60° zones at simulation termination with ti = 16. Power control and voice activity monitoring are activated in a

greater number of coverage areas as homogeneous demand increases from 4 to 16. The optimal RRAs

effectively reduce the cii from 5 to 2 in the coverage areas of the corresponding base stations i.

139

Direct comparisons are tenuous when using similar, but not identical, examples. As an indirect

comparison, a modified HC-HNN is run to solve the CAP for the same 21-omnicell network, with a TDMA scheme and homogeneous traffic successively set to ti = 4, 6, 8 16 for each cell i. ABPs are 0.021, 0.044, 0.138, and 0.26, respectively; corresponding ANIs are 17.4, 24.7, 38.2 and 47.5. Cells sectored into three 120° zones reduce cii from 5 to 4 and improve ABP to about 0.153 with ANI of 33.8 when ti = 16. A GA with forced channel reassignment for the CAP of the 21-cell network, where each

cell has three channels available and tk= ti=1, attains an ABP of about 0.017, averaged over 100 runs 2.

A non-homogeneous network example, introduced by Kunz, is based on an area measuring 24 km by 21 km around Helsinki, Finland 1. There are 25 base stations, distributed non-uniformly over the area, and 73 channels to satisfy the interference conditions. The interference matrix and traffic demand vector in Figure 3(a) and 3(b) were generated from this data. The SOFM of the SRRA problem for this network is simulated, with K = 100 initial states and all RRAs initialized to (1, 0, 0) in each coverage area. The ABP is 0.078, with ANI equal to 421.6. While this is far better than convergence in 2450 iterations reported by Kunz for his version of HNN l, the SOFM is much slower than either the HC-HNN or GA

approaches to the CAP for the same example2 . This may be due to the SOFM “learning” the correct

RRAs iteratively over the search space. The SOFM does attain a lower ABP due to its ability to increase

140

capacity to meet demand. The algorithm slowly increments the sectorization values to 6, sets the voice activity monitoring “on”, while power control assignments vary over the 25 base stations. To evaluate performance of the SOFM modified for DRRA problems, the components of T2 in the Kunz model are cyclically shifted 5 positions down every

periods, with

and 100, to

represent dynamic local traffic. Algorithm sensitivity to initial RRAs, in response to demand shifts, is examined by initially using the following three patterns in the coverage areas: (1) all RRAs are (1, 0, 0);

(2) all RRAs are (3,1,1); and (3) the RRAs are set to the final values after K = 100 periods of the SRRA. In response to these cyclic demand shifts, the ABPs for the DRRA with initial RRA pattern (1) increase from 0.068 to 0.381, as the number of periods

for the algorithm to learn the new network demands,

decreases from 100 to 10, respectively. For initial RRA pattern (2), the ABPs increase from 0.032 to

0.294, as

decreases from 100 to 10, respectively. Lastly, using the final resource patterns from the

SRRA problem for the Kunz network to initiate the RRAs in the dynamic resource control of the same network causes the ABPs to range from 0.002 to 0.169, as

decreases from 100 to 10, respectively.

These results can be compared informally to a GA simulation for the CAP of the Kunz model, with cross-over

mutation

and bias weights A = 1.0 and B = 1.1. The components of T2

are cyclically shifted 5 positions down every 100 generations to test convergence sensitivity to demand

shifts. The ABP for this GA varies from 0.01 to 0.343 depending on the size of the demand shift 2. All simulation runs of this GA converge within 200 generations.

7. CONCLUSIONS The Kohonen self-organizing map has been extended to the RRA problem to incoming calls in a

DS/CDMA PCS network. The problem statement generalizes the determination of the minimum number of channels required to obtain an interference-free assignment. The SOFM application better reflects practical systems, in which radio resources are regulated to minimize network interference. The Kohonen SOFM is extended to perform discrete-space optimization of SRRA problems, then further modified for more robust performance in DRRA problems. Both RRA algorithms have been simulated

for known network examples, with results informally compared to published simulations of HC-HNN

and GA methods applied to the CAP for similar networks. Future investigations must refine the approach based on more accurate models of PCS network behavior. Duque-Antón, et. al. observe that the interference produced by the simultaneous use of a channel in two cells is not fully known in the PCS

environment, since traffic dynamics are more difficult to model with increasingly smaller cell size. Further research is required to learn the interference constraints in such an adaptive environment by

statistically correlating carrier and interference power levels with call activities14, 15.

141

REFERENCES 1. D. Kunz, “Channel assignment for cellular radio using neural networks,” IEEE Trans. Veh. Technol., vol. VT-40, pp. 188-193, 1991.

2. W. Hortos, “Comparison of neural network applications for channel assignment in cellular TDMA networks and dynamically sectored PCS networks,” Appl. and Sci. of Artificial Neur. Net. III, Proc. of SPIE, vol. 3077, pp. 508-524, Orlando, FL, 1997. 3. P. T. H. Chan, M. Palaniswami, and D. Everitt, “Neural network-based dynamic channel assignment for cellular mobile communication systems,” IEEE Trans. Veh. Technol., vol.VT-43, no. 2, pp. 279-

288, 1994.

4. J. J. Hopfield and D. W. Tank, “ Neural computation of decisions in optimization problems,” Biological Cybern., vol. 52, pp. 141-152, 1985. 5. K. Smith and M. Palaniswami, “Static and dynamic channel assignment using neural networks,”

IEEE J. Selected Areas Comm., vol. 15, no. 2, pp. 238-249, 1997. 6. S. Abe, “Convergence acceleration of the Hopfield neural network by optimizing integration step

sizes,” IEEE Trans. Syst. Man and Cybern.-Pt.B, vol. 26, no. 1, pp. 194-201, 1996. 7. W. C. Y. Lee, “Applying the intelligent cell concept to PCS,” IEEE Trans. Veh. Technol., vol.VT-43, pp. 672-679, 1994.

8. K. Gilhousen, I. Jacobs, R. Padovani, A. Viterbi, L. Weaver, and C. Wheatley, “On the capacity of a cellular CDMA system,” IEEE Trans. Veh. Technol., vol.50, no. 2., pp. 303-312, 1991.

9. T. Kohonen, “Self-organized formation of topologically correct feature maps,” Bio. Cybern., vol. 43, no. 1, pp. 59-69, 1982. 10. S. Abe, “Global convergence and suppression of spurious states of the Hopfield neural nets,” IEEE Trans. Circuits Syst., vol. CAS-40, no. 4, pp. 246-257, 1993. 11. K. N. Sivarajan, R. J. McEliece, and J. W. Ketchum, “Channel assignment in mobile radio,” in Proc.

39th IEEE Veh. Technol. Soc. Conf., pp. 846-850, San Francisco, CA, 1989. 12. W. K. Lai and G. G. Coghill, “Channel assignment through evolutionary optimization,” IEEE Trans. Veh. Technol., vol. VT-43, pp. 91-96, 1996.

13. N. Fubaniki and Y. Takefuji, “A neural network parallel algorithm for channel assignment problems in cellular radio networks,” IEEE Trans. Veh. Technol., vol. VT-41, no. 4, pp. 430-437, 1992. 14. M. Duque-Antón, D. Kunz, B. Rüber, and M. Ullrich, “An adaptive method to learn the compatibility

matrix for microcellular systems,” in Proc. IEEE 44nd Veh. Technol. Conf. VTC-94, pp.848-852, Stockholm, Sweden, 1994. 15. M. Duque-Antón, D. Kunz, and B. Rüber, “Learning the compatibility matrix for adaptive resource management in cellular radio networks,” Eur. Trans. Telecommun., vol. 6, no. 6, pp. 657-664, 1995.

142

3 Complex Scaled Tangent Rotations (CSTAR) for Fast Space-Time Adaptive Equalization of Wireless TDMA Massimiliano (Max) Martone, Member, IEEE

WATKINS-JOHNSON COMPANY Telecommunications Group 700, Quince Orchard Rd. Gaithersburg, MD 20878-1794, USA E-mail: [email protected]

Abstract A new update algorithm for space-time equalization of wireless TDMA signals is presented. The method is based on a modified QR factorization that reduces the computational complexity of the traditional QR-Decomposition based Recursive Least Squares method and maintains numerical stability and tracking capability. Square roots operations are avoided due to the use of an approximately orthogonal transformation, defined Complex Scaled Tangent Rotation.

I . I NTRODUCTION The space-time equalization concept was first proposed in [1] and subsequently applied to

wireless TDMA in [2] where it was conjectured the implicit optimality of the scheme. The use of

different feedforward filters at the antennas and one single feedback filter was demonstrated to be effective because it was able to simultaneously combat signal fading, intersymbol and cochannel interference. The implementation of the joint update requires special attention because low signal to noise ratio and fast frequency selective fading channels result generally in ill-conditioned

adaption. Recursive Least Squares based on QR Decomposition [3] is a well known and numerically well behaved method to perform the filters update. However the high computational complexity of the method has always been considered a remarkable problem. We propose in

this work a new algorithm based on an approximated QR factorization which improves in terms of computational efficiency over existing schemes mainly because square roots operations are avoided. The approach uses a generalization of the scaled tangent rotation of [5] to update the Cholesky factor of the information matrix without needing to form it. The performance of the method is compared to more traditional algorithms by means of computer simulations in fixed point arithmetic for a realistic scenario as specified in the standard IS-136 [6], [7] for cellular communications in the US.

II. S PACE -T IME QR-BASED MMSE E QUALIZATION Consider a K-antenna receiver. At the k-th antenna

delayed and attenuated replicas of

the signal are received (k = 1,2, ...,K). The impulse response of the two multipath diversity channels can be expressed as

are delay, amplitude, phase of the m-th path as received at the k-th antenna.

The complex baseband modulated signal is

are the complex symbols defining the signal constellation used for the particular digital modulation scheme1. The filter

is a square root raised cosine shaping filter with roll-off

factor equal to 0.35, T is the signaling interval. The baseband signal received at the k-th antenna is

where

frequency, and

is additive Gaussian noise.

usually in the range fractionally spaced samples

is the carrier

is sampled at

rate (R is an integer

and square root raised cosine filtered to obtain the complex I/Q The optimum combining/equalization scheme

has K feedforward filtering sections (one per antenna) and one feedback filtering section (see Fig.

1 for K = 2). Since the algorithm jointly optimizes the taps of these filters as to minimize the Mean Squared Error using samples of the received signal taken at different points in time and in space, this architecture is called Space-Time Minimum Mean Square Error (MMSE) equalizer. We can express the output of the MMSE Space-Time equalizer in vector notation as

with

for k = 1,2,...,K,

in the US standard for cellular communications [6]. are fractionally spaced taps, while c(n) is updated once per symbol.

144

are symbol spaced taps. Moreover

for k = l,2,...,K, and

The adaptive algorithm minimizes the Mean Squared Error defined as and the goal of the adaption process is to adjust

1)) to converge toward the solution

. The sequence

is

generated using the known training symbols (during training) and using past decisions (during data demodulation, in decision directed mode). The QR approach: The equalization problem can be reconducted to solving at each n + 1 step

the problem

where

is the forgetting factor

is the data matrix and

The normal equations define the desired minimizer as

The use of orthogonal transfor-

mation to solve least squares problems is well established as is the inadvisability of using the normal equations [3]. Suppose that a matrix

is known from the previous step such that

with Q orthogonal and

upper triangular matrix, then the problem

stated in (2) is equivalent to

, because Euclidean distance is preserved by orthogonal transformations. The traditional QR-RLS approach [3] obtains c(n + 1) from the solution of the triangular

linear system

where

is obtained by sweeping the row vector

in (3) through orthogonal transformations represented by

applying the same orthogonal transformations to III. CSTAR TRANSFORMATION AND THE NEW METHOD

The novelty of the method we present is in the following two points:

145

is obtained

1. the algorithm tracks the variation in c(n) from step n to step n + 1 rather than c ( n ) itself. This saves some computations and simplifies part of the algorithm.

2. the orthogonal transformations are approximated by Scaled Tangent Rotations rather than the traditional Givens rotations.

Define

From the previous Section

with

upper triangular. Since

is the solution of3

where

where

Hence

can be found by solving the triangular system

satisfies

It is then evident that all we need to solve the system (6) can be obtained by forming

and sweeping the bottom part of this matrix using plane rotations4. trix 3 4

can be found as a product of

Just substitute It should be clear now that by tracking

elements of the last column of

The orthogonal ma-

Givens rotation matrices [3]:

we have simplified the rotation step in

because the first L

are equal to zero. We save 2L complex multiplications with respect to the

QR-RLS of [3] and, more important, the dynamic range required to represent the elements of the last column of the augmented matrix

is reduced.

146

A single Givens rotation element

annihilates the L + l, l-

using where

and

The considerable drawback of the method is in the computation

of the angles for the Givens rotations. Square-root computations are not easily implemented in

DSP processors and are even more problematic in VLSI circuits. Usually they involve iterative procedures whose convergence is not always guaranteed. Scaled Tangent Rotations (STAR) were proposed in the context of RLS adaptive filtering for real time-series in [5]. We generalize the rotation to the complex domain but there is a difference that is important to note. In STAR [5]

scaling is necessary to prevent instability caused by the fact that the tangent function may become infinite. Our modification for complex signals still contains a scaling operation but the factors are not normalized to unity as in [5] because this would involve again a square-root operation. The

CSTAR (Complex Scaled Tangent Rotation) transformation in analogy to the Givens rotation is defined in terms of each

as in Table 1, where the

complex sign function is

and the sweep is

applied to the augmented matrix

(see Table 2). Observe that the Givens elementary com-

plex rotation has two remarkable properties. First of all it zeroes selectively one predetermined

element of any complex matrix, which is indeed needed to triangularize the data matrix. Sec-

ond, it is an orthogonal transformation, that is

is an orthogonal matrix that preserves

the original least squares problem. The CSTAR elementary transformation

maintains

the first property but it is not an orthogonal tranformation. Note in the flow diagram of the

transformation (Table 1) that, whenever scaling is required, we obtain a non-orthogonal elementary matrix (and of course a non-orthogonal T(n)). So the CSTAR solution deviates from the optimum least squares solution5. However it is possible to show along the same guidelines of [5] that the deviation from orthogonality is limited to the first few adaption steps because as 5

In fact at any given step

where

and

are the triangular matrices obtained from the sweep performed applying Givens

Rotations and CSTAR Rotations, respectively.

147

new samples are processed the scaling operation becomes more and more unnecessary. In other words the algorithm has the property that

where I is the identity matrix and we have used the notation for the Frobenius norm of a

complex matrix M whose generic i,j element is

Our experimental results show that the effect of this initial bias is negligible. The algorithm can be summarized as in Table 2. In Table 4 the computational complexity of the CSTAR method is compared to the QR approach using Givens Rotations (defined QRG) and the EWRLS (exponentially windowed RLS of [4]) in terms of real multiplications, reciprocals, square roots and additions.

IV. CSTAR-RLS WITHOUT THE B ACKSUBSTITUTION STEP The backsubstitution step is implicitly a serial process: the unknowns are obtained one by one

with a complexity

. In addition L reciprocals are needed. A simple derivation [8] shows

that a very elegant solution exists for

that avoids divisions and make the complexity

It is possible to prove the two following facts using the matrix factorization lemma. Fact 1: The inverse hermitian of

can be recursively computed

using

where

is the (asymptotically) orthogonal matrix that sweeps the bottom part of (8) using Plane Rotations. Fact 2: The solution of the triangular system (6) is obtained as:

where

is obtained from (10) and

is obtained from (7).

The updating algorithm can be summarized as in Table 3. If the matrix with the matrix

where

is any real number, then

148

must be initialized to

is initialized

V. S IMULATIONS A dual-diversity (K = 2) TDMA system for cellular communications has been simulated according to [6], [7]. We assume a two-ray Rayleigh fading diversity channel

[7]. The delay interval is the difference in time of arrival between the two rays at each antenna. The speed of the transmitter mobile defines the time-varying characteristics of the channel. The

frame is constituted by 162 symbols and 14 of them are dedicated to the training sequence. Delay interval for both diversity channels is equal to

(one symbol period) to describe

an environment severly affected by intersymbol interference.

The length of the feedforward

sections is 3, the length of the feedback section is 2. Sampling rate is 2/T (R = 2). The described algorithm, the EWRLS algorithm (traditional RLS [4]) and the QR-RLS algorithm (traditional QR-based RLS, [3]) have been implemented using 24 bits of resolution in the fixed

point arithmetic representation. Fig. 2 shows performance of the CSTAR algorithm compared to

EWRLS and to QR-RLS in terms of Mean Squared Error estimated and averaged over 100 runs. The value of is 0.855 for

Fig. 3 shows Bit Error Rate (BER)

results. The EWRLS algorithm reveals numerical problems directly impacting BER performance.

The CSTAR algorithm achieves performance similar to the traditional Givens-based QR-RLS.

VI. C ONCLUSIONS We have presented a new method to update the digital filters of a MMSE K-antenna spacetime decision-feedback receiver. The algorithm is particularly suited for fixed-point arithmetic implementations because it preserves the numerical stability and performance of a QR-based

approach but it is less computationally demanding due to the absence of square root operations. Experimental results were presented for the IS-136 North-American [6] standard for cellular

TDMA communications to validate the method and to confirm the effectiveness of the approach.

R EFERENCES [1] P. Monsen, “MMSE equalization of interference on fading diversity channels”. IEEE Trans. Comm., vol. 34, pp. 5-12, Jan. 1984. [2] C. Despins, D. Falconer, S. Mahmoud, “Compound strategies of coding, equalization, and space diversity for

wide-band TDMA indoor wireless channels”. IEEE Trans. Vehicular Tech., vol. 41, pp. 369-379, Nov. 1992. [3] S. Haykin. “Adaptive Filter Theory”, Englewood Cliff, N.J Prentice All, 1986. 6

In general low speeds require larger values of

there is marginal BER degradation if

to get the best performance out of the tracking scheme. However

is kept fixed to 0.855.

149

[4] J. G. Proakis. "Digital communications". McGraw-Hill, 1989. [5] K. J. Raghunath and K. K. Parhi “Pipelined RLS Adaptive Filtering Using Scaled Tangent Rotations”. IEEE Trans. Signal Proc., vol. 44, No. 10, pp. 2591-2604, Oct. 1996. [6] TIA/EIA/IS-136.1-A, “TDMA Cellular/PCS - Radio Interface - Mobile Station - Base Station Compatibility -

Digital Control Channel” and TIA/EIA/IS-136.2-A, “TDMA Cellular/PCS - Radio Interface - Mobile Station - Base Station Compatibility - Traffic Channels and FSK Control Channel”, October 1996. [7] TIA/EIA/IS-138, “800 MHz TDMA Cellular - Radio Interface - Minimum Performance Standards for Base Stations”, December 1994. [8] B. Yang and J. F. Bohme, “Rotation-based RLS algorithms: unified derivations, numerical properties, and parallel implementation”, IEEE Trans. Sig. Proc., vol. 40, pp. 1151-1167, May, 1992.

150

Table 2: The CSTAR algorithm. Initialization • 0) Inputs

• 1) Compute the prediction error

• 2) Form matrix

• 3) Sweep

using L CSTAR Rotations

• 4) Solve by backsubstitution the triangular system

• 5) Obtain

Table 3: The CSTAR algorithm without Backsubstitution. Initialization • 0) Inputs

• 1) Compute the prediction error

2) Sweep

using L CSTAR Rotations

• 3) Obtain

151

152

153

154

4 An Effective LMS Equalizer for the GSM Chipset Jian Gu, Jianping Pan, Renee Y. Watson, and Steven D. Hall CommQuest Technologies, Inc. 527 Encinitas Boulevard, Encinitas, CA 92024 (760)-634-6181, [email protected]

Abstract An effective equalization approach to implementing receivers for Global System for Mobile

communications (GSM) is presented. The approach utilizes a linear transversal filter using an Least Mean Square (LMS) algorithm for tap coefficient training. Effective soft decision variables are generated from the output of the LMS equalizer. It is well known that an LMS equalizer has low complexity and can be easily implemented into an Application Specific Integrated Circuit (ASIC). Performance comparisons with 16-state Viterbi equalization are included. This

approach is incorporated into CommQuest’s GSM chipset design for GSM receivers using its Communication Application Specific Processor

technology.

I. Introduction

Many communication systems have to deal with multipath propagation channels, especially for those used for cellular/PCS mobile communications. The Global System for Mobile communications (GSM) is such a system for mobile communication and has been adopted by

many countries worldwide. Gaussian filtered Minimum Shifting Keying (GMSK) modulation is

used for GSM. The GSM standard [9] requires any receiver for demodulation of the GMSK modulated signal should be able to handle multipath propagation channels with delay spread up to

In contrast, the bit duration or the symbol duration of the GMSK signal is about 3.69 Therefore, a delayed and probably attenuated, and phase-rotated signal through the longest

path of the propagation channel may arrive at the receiver up to about a five-symbol duration after the signal traveling through the shortest path. Without utilizing any advanced equalization techniques, it is impossible to meet the Bit Error Rate (BER) requirements specified in the

standard [9]. There are many equalization methods that can be used in receivers for GSM systems. A wellknown and widely used approach is called Viterbi equalization -- an approach using the

155

Maximum Likelihood Sequence Estimation (MLSE) [2]-[7]. The Viterbi equalizer needs a

propagation channel estimator to generate all possible signal sequences which are resulted from being transmitted through the estimated propagation channel. Given the channel memory length, the number of such signal sequences is limited. These generated signal sequences are compared

with the received signal sequence and the generated signal sequence which is most like the received signal is selected. The data sequence associated with the selected signal sequence is the recovered data sequence. Propagation channel estimation is performed by detecting a known data

sequence called the midamble, which is embedded in the middle of the burst. When the propagation channel becomes very noisy and/or strong interfering signals appear, both of which are common in the mobile communication environment, performance of the Viterbi equalizer

could be significantly compromised due to the fact that a reliable estimation of the propagation

channel is not available. Another typically used method is equalization with a linear transversal equalizer [1] [2]. It is said that the Least Mean Square (LMS) algorithm does not converge fast enough and therefore

some fast converging algorithms like Recursive Least Square (RLS) with decision feedback could be used for GSM receivers [1]. However, crashes of the decision-directed equalizer due to error-propagation were observed [10], and some techniques that restart the equalizer may have to be used. In this paper, we present an approach of using a transversal equalizer with LMS algorithm for GSM receivers. Considering error propagation and quantization effects (Using 8-bit arithmetic), no decision feedback is introduced and thus preventing possible complications. With proper initialization and training, the LMS equalizer converges fast enough to provide the same or better

performance in comparison with Viterbi equalization in very noisy and/or strong interference environments, especially near the reference sensitivity level and reference interference level specified in the standard ETSI/GSM 05.05 [9]. The transversal equalizer is also less complicated than the Viterbi equalizer; and therefore, it is very easy to implement.

Section II describes the demodulator structure with the transversal equalizer. Section III presents simulation results for both the transversal equalizer with the LMS algorithm and the Viterbi equalizer. Some conclusions are drawn in Section IV.

II. Structure of the Demodulator

156

Fig. 1 shows the block diagram of the demodulator. The converter in Fig. 1 includes all functions needed to convert the received radio signal into a pair of in-phase and quadrature digital samples, denoted by multiple of bit interval

is the sampling duration which is a

that is about

Access (TDMA) and GMSK with

and

for GSM. GSM utilizes Time Division Multiple where B is the 3-dB bandwidth of the Gaussian filter

is the bit duration. Since the GMSK modulation is a binary modulation scheme, its

symbol interval is equal to the bit interval

The dash-lined block in Fig. 1 includes functions

for the baseband I-Q signal processing. The GMSK modulated signal is transmitted in burst format. Fig. 2 depicts the structure of bits in a burst which is to be transmitted in a time slot. There are 116 data bits which are separated by a synchronizing sequence of 26 bits which is called the midamble, preceded by 3 tailing bits and followed by 3 tailing bits. After being differentially encoded, a burst of modulating bits is formed and fed into the GMSK modulator. The demodulator recovers the 116 data bits of a burst from the received signal. The baseband demodulator consists of a sample buffer, a midamble correlator, a processing unit, a decimator, a transversal equalizer, a power measurement unit, a Signal-to-Noise-Ratio (SNR) estimator, a soft decision variable generator, and a timing recovery unit. The sample buffer stores the baseband I-Q samples collected at a rate

of the bit rate

which is a multiple

The timing recovery unit controls when to start collecting samples of the

received burst signal.

Fig. 3 shows a diagram of the midamble correlator which is a Finite Impulse Response (FIR) filter modeled by a

tapped delay line but with L non-zero tap coefficients, where

The L tap coefficients are the conjugates of the I-Q samples corresponding to the midamble portion of the received baseband signal at sampling intervals of it correlates L received I-Q samples spaced at

At each cycle of

in the sample buffer. Under multipath

propagation conditions, the relative delay between the first ray and the last ray of the radio signal can be over

If

is the maximum relative delay between the first ray and the last ray, we

use a midamble search window of

since we want to align the nominal timing with the

center of the search window. Therefore, the time span of the samples in correlation is and the total number of output values is 2MN. Nominal timing is provided by the timing recovery unit.

157

The processing unit executes a number of functions including calculation of the Multipath Intensity Profile (MIP), peak detection, estimation for the propagation conditions of the dominant ray, and tap reduction for the transversal equalizer based on the MIP. The MIP is obtained by calculating the square of the magnitude of the I-Q samples at the

output of the correlator over the midamble search window. Assuming the auto-correlation

function of the midamble signal is ideal: if there is only one ray of the received signal, then the

MIP has only one peak and timing of the peak corresponds to the arrival timing of the received

single ray signal. However, if there are several rays with different arrival timings, then there would be several peaks each of whose values represents the strength of the rays, and the timings

of the peaks relate to arrival timing of the rays. The maximum magnitude value corresponds to the strength of the dominant ray of the propagation path. The output of the midamble correlator is

the propagation channel response, if there is only one ray, or is an approximation of the propagation channel response, if there are several rays. For example, if there is only one ray and the propagation channel response is

and if we scale the received signal by

then the

resulting signal can be perfectly demodulated without any equalization, where

is the

conjugate of C . The decimator has a rate of M: 1. The epoch of the decimator is aligned with the timing of the dominant ray from the peak detection. Decimation on

results in a subsequence

where K is related to the epoch of the decimator

and is an integer between 0 and M-1. The transversal equalizer shown in Fig. 4 is a

tapped delay line equalizer in training

mode with 2N+1 taps. The input signal of the equalizer is trained it is simply a FIR filter. The center tap is initialized with

After being where C is the

propagation condition estimate of the dominant ray of the received signal. The remaining taps

are set to null before training. A tap-reduction algorithm is used to determine the number of taps

needed to ensure good performance based on the estimated dispersion of the propagation channel. After the tap reduction algorithm, some of the taps are excluded from the training process, i.e. these taps will not be updated from their initial null values during the training process. The well-known LMS algorithm [8] is used in the training process. Referring to Fig. 4 and letting

and

be the k-th symbol of the midamble and the estimated received k-th 158

symbol

of

the

midamble

respectively,

then

the

updated

tap

coefficients

are

j may assume values of a smaller range as a result of tap reduction.

is the step size and

the (k – 1 )-th training and its final value is

is the tap coefficient after

After training, all tap coefficients remain fixed

during demodulation of the current received burst. Based on the MIP, the 2MN samples of the squared magnitude are divided into 2N+1 subsets.

Fill an array mip[.] by summing up M samples of magnitude and assigning the sum to an element

of the array. Note that mip[N] is the sum of M samples in the middle of the 2MN samples and

that mip[0] and mip[2N] are a sum of M/2 samples. Each element of mip[.] corresponds to a tap of the transversal equalizer. Define start_tap as the index of the first non-zero tap and end_tap as the last non-zero tap. The following is the C program of the tap reduction algorithm:

where p is a scale factor to establish a threshold for the tap reduction algorithm to operate when

the propagation channel is less dispersive.

and

are called aggressiveness factors and both are

less than 1. The larger the aggressive factors the more aggressive the tap reduction. Also, we can

treat non-causal tap and causal taps differently by making

different.

The power measurement unit calculates the average signal power of the midamble portion, which is an average of

of the related I-Q samples. Note that this average power includes

noise power and/or interference power; hence, it can be written as S+N, where S is the wanted signal power and N is the power of undesired signals including noise and/or interference. The

159

SNR estimator calculates S/(S+N-S), where S is the wanted signal power which is obtained at the output of the midamble correlator since the correlator is in fact a matched filter to the midamble

and rejects most of the noise and interference. The soft decision variable generator scales the outputs of the equalizer by the estimated SNR and transforms them to samples with a given bit-width. The I-Q combiner, shown in Fig. 5, converts the complex representation of the signal to a one-dimensional sample sequence. The sign bit of each sample in the sequence is the hard-decision data bit. For a simpler

implementation, the combiner can be placed before the soft decision variable generator. The timing recovery unit collects the timings of the dominant rays of several bursts from the

processing unit and calculates the average timing. Only the timings associated with “good” bursts are used in the averaging process, where “good” means that the estimated SNR of a burst is large. It then compares the new average timing with the previous average timing and adjusts the

clock/counter as necessary. The adjusted timing is called new nominal timing and is applied to the midamble correlator and other units for processing of the next burst.

III. Simulation Results The LMS equalizer has been incorporated into our fixed-point system simulation of the

receiver including Surface Acoustical Wave (SAW) filters, Intermediate Frequency (IF) amplifiers with Automatic Gain Control (AGC), and an IF bandpass sampling Analog-to-Digital Converter (ADC). Simulation results show that the LMS equalizer based receiver fully satisfies, with substantial margins, the sensitivity and interference performances specified in the GSM

standard. Assuming the overall RF/IF system noise figure is 8 dB, the receiver has a BER performance margin of about 5 dB at the reference sensitivity level for the TU50 channel conditions. The receiver’s co-channel interference performance has about 1.8 dB margin at the

reference interference level for TU50. These results and other simulation results for all other propagation conditions are consistent with measured results from actual hardware.

In the following, we make some comparisons with the performance of a receiver using 16-state

Viterbi equalization. The same IF stages and a similar tap reduction algorithm are used for the 16-state Viterbi equalization cases. All simulation runs last 20,000 bursts.

Note that the noise power in the SNR calculation is the white noise power over a frequency band of 200 kHz and that if the overall system noise figure is 8 dB then SNR = 11 dB corresponds to a signal level of -102 dBm. 160

Fig. 6 illustrates the raw BER curves for both the transversal equalizer (labeled with LMS) and Viterbi equalizer (labeled with VE). Fig. 7 illustrates the Residual Bit Error Rate (RBER) of Class 1b and Frame Erasure Rate (PER) of the two cases, where RBER is defined as the ratio of

the errors detected over the “good” frames to the number of bits in the “good” frames where

“good” frame means the frame is not erased. The propagation channel condition is TU50 which

is Typical case for Urban area with vehicle speed at 50 kilometers per hour. Fig. 8 and Fig. 9 show co-channel interference performances over TU50.

Note in Fig. 6-Fig. 9 that the two equalization approaches have close performances over those SNR or Carrier-to-interference-Ratio (CIR) ranges. Moreover, for the coded BER (RBER of Class 1b and FER) the LMS equalization is generally better than Viterbi equalization over most

parts of the range. IV. Conclusions

We have presented an effective demodulator with an LMS equalizer for GSM handheld units. The demodulator uses a number of techniques to ensure performance and efficiency. Proper tap

coefficient initialization and burst timing setting allow the LMS equalizer to converge at a faster rate. The tap reduction algorithm retains only those taps that are necessary for handling the propagation condition of a given burst. The tap reduction algorithm also treats non-causal and causal taps differently in order to retain significant taps. The retained taps are to be trained while the other taps are set to zeros. Fewer taps results in lower residual error when the channel dispersion is low. Compared with the commonly used Viterbi equalizer, the transversal equalizer has the same or better performance for uncoded BER when SNR is low (signal levels near the reference sensitivity level of -102 dBm) or when CIR is near the reference interference level which is 9 dB

for co-channel interference. This is mainly due to the fact that the channel response estimator for the Viterbi equalizer does not provide a good channel response estimation at low SNRs or in severe interference conditions. In addition, the LMS equalizer has a better coded BER performance than that of the Viterbi equalizer when SNR and/or CIR is low. The LMS equalizer based receiver has been implemented into a digital Application Specific Integrated Circuit (ASIC) called Communication Application Specific Processor

and

incorporated into the CommQuest’s chip set for GSM handheld units. Reception tests of the units have verified the simulation results. Testing shows the receiver has about 6 dB margin in

161

sensitivity performance for the TU50 channel conditions and 7 dB margin for static channel

conditions. The receiver has about 2 dB margin for co-channel interference performance for TU50. The units also satisfy all other BER performances specified in GSM standard [9] with

substantial margins.

References: [1] Giovanna D’Aria, Roberto Piermarini, and Valerio Zingarelli, “Fast Adaptive Equalizers for Narrow-Band TDMA Mobile Radio,” IEEE Trans. on Vehicular Technology, Vol.. VT-40, pp.

392-404, May 1991.

[2] John G. Proakis, “Adaptive Equalization for TDMA Digital Mobile Radio,” IEEE Trans. on Vehicular Technology, Vol.. VT-40, pp. 333-341, May 1991.

[3] Renato D’Avella, Luigi Moreno, and Marcello Sant’Agostino, “An Adaptive MLSE Receiver

for TDMA Digital Mobile Radio,” IEEE J. Select. Areas Commun., vol. SAC-7, pp. 238-247, Jan. 1989. [4] G. Benelli, A. Garzelli, and F. Salvi, “Simplified Viterbi Processors for the GSM Pan-

European Cellular Communication System,” IEEE Trans. on Vehicular Technology, Vol. VT-43,

no.4. pp.870-878, Nov. 1994. [5] G. Benelli, A. Fioravanti, A. Garzelli, P. Matterini, “Some Digital Receivers for the GSM

Pan-European Cellular Communication System,” IEE Proc.-Commn. Vol. 141, No.3, June,

pp.168-176, 1994. [6] J. C. S. Cheung and R. Steele, “Modified Viterbi Equalizer for Mobile Radio Channels

Having Large Multi-path Delays,” Electronics Letters, Vol. 25, No. 19, pp.1309-1311, Sept. 1989.

[7] E. Del Re, G. Benelli, G.Castellini, R. Fantacci, L. Pierucci, and L. Pogliani, “Design of a Digital MLSE Receiver for Mobile Radio Communications,” 1991 GLOBECOM, pp.1469-1473. [8] John G. Proakis, Digital Communications. New York: McGraw-Hill, 1983.

[9] European Telecommunications Standard Institute (ETSI), ETSI/GSM 05.05, 1996.

[10] E. Eleftheriou and D. D. Falconer,”Adaptive equalization techniques for HF channels,” IEEE J. Select. Areas Commun., vol.SAC-5, pp. 238-247, Feb. 1987.

162

163

164

165

166

5

Self-Adaptive Sequence Detection via the M-algorithm Ali R. Shah Ericsson Inc., 740 East Campbell Rd., Richardson TX 75081 Bernd-Peter Paris Department of Electrical and Computer Engineering Center of Excellence in C3I George Mason University Fairfax, VA 22030

April 29, 1998 Abstract

The problem of implementing self-adaptive equalization algorithms in real-time is addressed. Self-adaptive equalization determines the transmitted sequence without using a training sequence. The advantage over current adaptive equalization techniques is discussed. Tree search procedures have been shown to be more effective than dynamic programming. Simulation results for tree search procedures based on the M-algorithm are presented. The focus is on the effects of channel order, sequence length and modulation format on the BER. The performance of the M-algorithm is compared with

traditional approaches.

1 Introduction Many problems in digital communications can be modeled by means of a discrete-time finitestate Markov process representing the signal which is observed in independent identically distributed noise. We are considering the case when the process parameters are unknown. We are investigating methods to exploit the structure and finiteness of the state space of the signal to determine the most likely state sequence without resorting to a known training sequence. We will refer to this approach as self-adaptive sequence detection (SASD). The problem is also referred to as self-adaptive or blind equalization in the communication literature. Self-adaptive sequence detection (SASD) has several advantages over techniques where the channel coefficients are estimated via a training sequence. In digital mobile radio, data

167

is transmitted over channels whose impulse response changes over time. In traditional approaches, a training sequence is sent with each data packet to re-estimate the channel. An

overhead of 10-20% is introduced, which could be eliminated through the use of self-adaptive techniques. The intersymbol interference (ISI) can be represented by a finite impulse response filter whose coefficients are sequence of equally likely symbols drawn from a discrete and finite alphabet A, is input to the channel. For slowly varying channels, the coefficients can be assumed to be unknown but constant for each data packet. The channel model is shown in Figure 1.

The goal is to determine the most likely input sequence out of candidate sequences where M is the alphabet size and N is the sequence length. The paper is organised as follows. We begin with the mathematical preliminaries and

describe the metric that needs to be computed. Then current techniques to find the transmitted sequence are also presented. Then we explain the algorithm and two methods to recursively update the metric. This is followed by simulation results and the performance of the M-algorithm.

2

Mathematical Preliminaries

We propose approaches similar to Feder and Catipovic [2] and Ghosh and Weber [3]. We briefly describe the approach to perform channel equalization when the channel coefficients are known. This is referred to Maximum Likelihood Sequence Detection (MLSD). The most likely sequence is the one that maximizes the joint probability density function (jpdf) of the observation v given the transmitted sequence and channel coefficients:

Assuming that the noise is Gaussian, we can explicitly write the jpdf in (1) as:

The sequence that maximizes (2) also maximizes the log of the density function. Taking the log yields the likelihood expression which allows simplification of the exponential term.

168

Maximizing this expression over all sequences s yields the maximum likelihood sequence estimate. Let denote the estimated sequence and s and v are vectors comprising the input symbols and observations, respectively. The most likely sequence can be estimated as:

where S is an matrix whose columns are shifted versions of s. The Viterbi algorithm [4] can be used to find the most likely sequence. When are unknown, they can be estimated

We can reformulate our criterion by replacing the estimate of

from (4) in 3.

After some manipulation, we can formulate our criterion as:

where

is a projection matrix. Among all possible

input sequences, we are looking for the one which maximizes the metric in (6). The optimal sequence is the one that spans the signal sub-space containing the largest portion of the received signal. For developing sequential algorithms, we do not use variants of the Viterbi algorithm like the Generalized Viterbi Algorithm (GVA) [5] and Per-Survivor Processing (PSP) [6]. The basis for our adaptation are tree search algorithms, originally developed for decoding of convolutional codes. In this paper we consider the M-algorithm [7].

3 The M algorithm The M-algorithm retains sequences that are the best in terms of the criterion in Equation (6). It extends each sequence to M branches to obtain paths, where M is the alphabet size The metric is computed for each sequence. The paths are then sorted in descending order and the best paths are retained while the rest are deleted. The M algorithm can also be referred to as the extend all nodes of stack-algorithm. A flow diagram of the M algorithm is in Figure 2. As compared to dynamic programming, there is an additional burden of sorting the sequences which is of the order Dynamic programming is infeasible for SASD as the metric includes the quantity that cannot be written in an additive form. Therefore, the metric for self-adaptive equalization depends on the whole sequence and not

169

170

just the last L symbols, which in turn implies that the concept of a “state variable” is no longer useful. The implication is that Bellman’s principle of optimality [8] does not apply to SASD. Good surviving sequences ending in the same “state” lead to eliminations in this framework, that are not justified by the principle of optimality. In this paper, the results show that the M algorithm does much better than PSP. In the simulation results we consider the effect of filter order (L), sequence length (N), SNR, stack size and modulation format (BPSK, QPSK) on the bit error rate (BER). A comparison is made with the clairvoyant detector (that knows and is nothing but the Viterbi algorithm).

3.1

Recursive Computation of the Metric

The objective function in (6) can be maximized in an iterative manner. The covariance matrix is decomposed using either a factorization or the RLS (recursive least squares) algorithm.

3.2

The RLS algorithm

The RLS factorization can be used to compute the metric in the following manner:

where the inverse autocorrelation matrix is denoted by The quantity updated via the RLS algorithm is:

where is the vector of last L + 1 detected symbols. The number of computations per recursion step for the RLS algorithm are shown in Table 1. We note that the order of computations is using the RLS algorithm.

3.3

The UDU factorization:

The UDU factorization [9] can be used to compute the metric in the following manner:

The recursion of this algorithm is based on updating the illustrates the number of computations per recursion step.

171

and

matrices. Table 2

172

4 Simulation Results In the simulation results we study the effects of different factors on the BER of the algorithm. Influence of Filter Order The simulation results indicate a degradation in performance with an increase in the filter order. This is intuitive as a large L implies a greater extent of intersymbol interference and therefore, greater loss in performance. The effect of L on the BER is studied in Figures 3, 4 and 5 for L = 2, 3 and 4 respectively. The performance degrades if the number of paths remain a constant for an increasing value of L. Influence of Number of Stored PathsThe stored sequences for the clairvoyant detector, that uses the Viterbi algorithm, are referred to as the survivors. They are the best sequences ending at each of the states. As there is no concept of a state for tree-search algorithms, the stored sequences for this case, can be greater or less than The bit error rate (BER) is plotted for different values of the number of stored paths, in Figures 3, 4 and 5. The number of sequences needed by the M-algorithm to achieve the performance of the clairvoyant detector is greater than or equal to A large value of also implies an increase in the computational complexity of the M-algorithm, as each path is processed in parallel. By choosing a larger stack, we come close to the global search and therefore, the likelihood of finding the best path increases. That is shown in the simulation results. Per-Survivor Processing A dynamic programming approach referred to as per-survivor processing (PSP) is also utilized in Figures (3), (4) and (5). In this approach, the state, is used to obtain the surviving sequences. In dynamic programming, the metrics do not have to be sorted at each stage as the survivors are decided at each state. We observe that PSP yields poorer performance as compared to the M-algorithm. The reason being that the estimate of the channel coefficients is unreliable. Decisions to discard sequences based on these estimates in PSP, results in the loss of good candidate sequences. Influence of Frame Length The effect of frame length is studied in Figure 6. The SNR

173

174

is 8 dB, L = 2 and the frame length is varied on log scale from 10 to 200. The various plots

are for different values of Shorter frame lengths imply that the algorithm has a lesser number of observations to obtain the distance metric. The performance of the self-adaptive algorithms is not good for small frame lengths. But, as the frame-length increases, the BER approaches the performance of the clairvoyant detector. The plot suggests that using the M-algorithm on frames of length of 250 or more, should yield the same performance as the clairvoyant detector. Influence of modulation scheme The previous results were with respect to BPSK. Using QPSK improves bandwidth utilization. The results for QPSK are shown in Figure 7. The symbol error rate (SER) is plotted versus the SNR. The M-algorithm works equally well for a larger stack size for QPSK. The performance of the M-algorithm approaches the

performance of the clairvoyant detector at high SNR. Effect of an initial global search An initial global search till k = 10 should improve the performance of the M-algorithm at the cost of increase in complexity. For BPSK, a global

search implies searching over possible paths. In Figure 8 a global search is performed initially followed by the tree search. The results are compared with the M-algorithm without the global search. The simulation results indicate an improvement in performance for an initial global search compared to the M-algorithm without a global search especially for cases when the stack size is small.

5 Conclusions This paper presents an approach for blind equalization which is referred to self-adaptive sequence detection. Substantial savings in bandwidth are possible over current techniques

using a training sequence. The main contributions in this paper are implementing a treesearch algorithm for SASD (self-adaptive sequence detection). The tree search approach used

175

176

is the M algorithm where the best candidate sequences in a stack are retained and the rest are deleted at each stage. Simulations are used to quantify the effects of storage size, modulation format, sequence lengt and channel order are performed. Two different update algorithms are also presented. The processing requirements of each algorithm is The performance of the self-adaptive sequence detection algorithm is compared with the clairvoyant detector (that knows the channel coefficients). If sufficient storage is available, the SASD algorithm works as well as the clairvoyant detector for different lengths of the channel filter, It works for various modulation formats including BPSK and QPSK. Blind equalization (or SASD) algorithms usually require a large number of observations for convergence. This approach shows that sequences as small as 200 symbols can be used and the same performance is achieved. Utilizing self-adaptive sequence detection implies elimination of considerable overhead which implies extra bandwidth. The current results suggest that if sufficient storage is available, implementing such algorithms in real-time is viable.

References [1] M. R. L. Hodges, “The GSM radio interface,” Br. Telecom Technol. J., vol. 8, pp. 31– 43, January 1990. [2] M. Feder and J. Catipovic, “Algorithms for joint channel estimation and data recovery— application to equalization in underwater communications,” IEEE Journ. Oceanic Engineering, vol. 16, pp. 42–55, jan 1991. [3] M. Ghosh and C. Weber, “Maximum-likelihood blind equalization,” Optical Engineering, vol. 31, pp. 1224–1228, jun 1992.

[4] G. D. Forney, “Maximum-likelihood sequence estimation of digital sequences in the presence of intersymbol interference,” IEEE Trans. Information Theory, vol. IT-18, pp. 363– 378, May 1972. [5] N. Seshadri, “Joint data and channel estimation using trellis search techniques,” IEEE Transactions on Communications, vol. 42, February/March/April 1994. [6] R. Raheli, A. Polydoros, and C.-K. Tzou, “Per-survivor processing: A general approach to MLSE in uncertain environments,” IEEE Trans. Communications, vol. COM-43, February/March/April 1995. [7] J. B. Anderson, “Limited search trellis decoding of convolutional codes,” IEEE Transactions on Information Theory, vol. 35, September 1989. [8] R. E. Bellman and S. E. Dreyfus, Applied Dynamic Programming. Princeton, NJ: Princeton University Press, 1962.

[9] G. J. Bierman, Factorization Methods for Discrete Sequential Estimation. New York, NY 10003: Academic Press, Inc., 1977.

177

7KLVSDJHLQWHQWLRQDOO\OHIWEODQN

6 Soft-Decision MLSE Data Receiver for GSM System Martin Lee ALOGOREX Inc. 33 Wood Avenue South Iselin, NJ 08830 [email protected]

and Zoran Zvonar Analog Devices Inc. Communications Division 804 Woburn Street, Wilmington, MA 02118 [email protected]

Abstract: The great success of GSM as a second generation digital cellular standard is largely due to the advances in integrated solutions for the GSM terminals, in particularly handsets. With the user population constantly growing and the price of handsets plummeting, focus of the design effort has moved to efficient implementations of GSM data receiver. In this paper we will give a brief overview of the operating modes of GSM handset, present the framework for the development of the data receiver and propose a new soft-decision based MLSE receiver which allows efficient hardware/software partitioning for the implementation.

1.

Introduction

Since its introduction, the GSM cellular standard has become a world wide success and has been adopted in many countries either as a cellular (GSM800) or PCS (GSM 1800, GSM 1900) standard. One of the major contributions to GSM’s acceptance has been its good performance in terms of quality of service, and this has in turn allowed manufactures to take advantage of economy of scale and reduce cost of equipment. Nevertheless, there is still considerable drive to lower cost even further as well as increasing user desired attributes such as talk and standby times. In order to achieve these goals, optimsation of the building blocks for a GSM mobile station (MS) is critical [1]. One of the major signal processing blocks for a GSM MS is the data receiver (DR). The importance of this is highlighted by the fact that the type approval procedure is based heavily on this part of the MS. It is the aim of this paper to present and discuss some techniques for the realisation of a GSM mobile station data receiver which is capable of meeting all the functions and requirements as specified in the GSM recommendations. It will be shown that the GSM receiver for demodulating the basic GSM traffic channel can be achieved with a MLSE structure which can be implemented with a combination of hardware and software modules which combined offers an attractive alternative to the standard methods.

2.

System Overview

The GSM data receiver function can be broken down into three main functional requirements, namely acquisition, synchronization and demodulation, as shown in Fig.1. As can be deduced some functions of the DR are only activated at certain instances while others are practically on all the time the mobile station is powered on. The fact this is the case puts different requirements on the different parts of the DR’s sub-functions. In acquisition mode, the DR must be able to continuously process the expected received signal to allow the MS to lock on to the infrastructure. This by its nature can be an process intensive task and thus must be optimized such that the MS is allowed to perform other task while it is in acquisition mode. In synchronization mode, the DR must be able to lock on to the infrastructure and set all the MS’s internal settings such that it can communicate in synchronism. The accuracy of the estimated settings dictates largely how fast the mobile can communicate effectively with the infrastructure. In normal demodulation mode, the DR’s main task is to provide reliable received symbols (bits) to the rest of the system. In addition it must be able to provide reliable estimates about the conditions of the MS’s setting such that it can be used to correct for any drifts after the acquisition and synchronization modes. Other task could involve providing reliable measures on signal strength and quality which are necessary for optimal control in a cellular system. In this paper we will focus on the normal demodulation mode of the receiver, assuming that first two functions of acquisition and synchronization have been achieved.

180

3.

GSM Data Receiver Requirements

GSM mobile station data receiver must be capable of meeting all the functions and requirements as specified in the GSM recommendations. The relevant GSM specification for the performance of the data receiver is given in the Rec.05.05 series [2]. In Fig.2. the complete data path in GSM system is depicted, indicating the functions of the handset that have been completely specified, and the functions that have been only specified by required performance and realization is left to the manufacturer. The desired performance of a GSM receiver is specified both for coded and uncoded data bits. In order to satisfy the performance and provide cost-effective solution known realizations of the GSM receiver have been implemented either as DSP software solution or alternatively as custom logic solution. While DSP implementation is flexible, it may not be preferable in terms of power consumption and cost. Consequently, the goal is to achieve a solution which will provide satisfactory performance while reducing the complexity of the implementation and preserving certain degree of flexibility. Complexity problem may be addressed on two different levels. On the algorithmic level complexity can be reduced by using suboptimal approaches, e.g. reduced-state sequence detection or decision-feedback equalization. On the architectural level complexity and performance constraints are usually addressed by the design of co-processors or accelerators for specific function, such as Viterbi algorithm for equalization and decoding.

181

4.

System Model and Receiver Structure

GSM system employs a Gaussian MSK (GMSK) modulation with BT=0.3 providing a net 270.8 kbit/sec rate at the air interface. Intersymbol interference (ISI) is introduced deliberately to improve spectral efficiency of the system. In addition, time dispersion of the propagation channel introduces additional ISI. The diagram of the system model under consideration is shown in Fig.3. GMSK signal can be interpreted as linearly modulated signal where the input bits are precoded to form a new symbol [3]. The new symbols are alternating between real and imaginary branch, therefore the transmitted data symbols are independently received in the quadrature branches, and spaced by twice the original symbol period. The received signal, subject to ISI lasting L symbol periods, can be approximated as

where

182

The received symbols from the two quadrature branches are received offset in time by one bit period. Thus in the absence of any errors in the sampling and any degradation one can

transform the data onto one branch only by performing a constant phase rotation in the signal i.e. multiplying the received by thus providing the real only stream of samples for consequent processing. This is often referred as serial receiver realization and consequent equalization can be performed on real signal only.

Given the knowledge of the combined channel response of the system one can derive the original transmitted sequence. The original design of GSM took this aspect into account, and in order to aid the estimation of the channel a midamble is inserted into every slot. A normal traffic slot structure is shown in Fig.4. The training sequence is designed to exhibit an auto-correlation property with a distinctive peak and minimal side-lobe content. The overall channel impulse response (CIR) can be estimated by correlating the received and expected training sequence. These can then be used to form the MF coefficients which by definition will maximize the SNR at the output of the filter. The impulse response of the MF is the time-reversed complex conjugate of h(m). In order to compute the values of h(m), one simply cross-correlate the received signal y(i) and the known signal c(i) which by design is the training sequence. The above have described a system which employs symbol sampling rate. Fractional rates can also be used which theoretically is able to give higher performance with the expense of added complexity. In order to estimate CIR one must require some knowledge of where the midamble situates in the TDMA burst. This “global” timing is derived from a higher layer synchronisation procedure. It is sufficient to assume that during normal burst demodulation a good estimate of the start of the TDMA slot is known in the receiver. Thus in order to estimate the CIR one does

183

not have to carry out the cross-correlation over an exhaustive range, but only for a small window over which the midamble burst is expected. The procedure can proceed as soon as enough samples have been collected, i.e. up to end of the midamble. Then a portion of the expected

sequence, usually 16 bits long, will be cross-correlated with the incoming sequence. The number of taps required for data receiver operation is usually restricted to L=5. Thus those L successive coefficients of the estimate h(m) which have maximum energy, will be chosen as the best estimate of the CIR, and thus provide MF coefficients. The above procedure for computing the MF coefficients and hence the timing is based on the fact that the signal quality is reasonable good such that the cross-correlation procedure can result in good estimates. When the signal is poor or when it is fluctuating rapidly, then the described method will in general give a poor CIR. During such conditions the search range and the known signal length (maximally 26) can play a significant role in reducing the estimation error of the CIR and hence timing.

4.1.

Equalization Strategies for GSM

Following the MF some form of equalization can be applied in order to minimize the effects of the channel. The optimal approach is to use Maximum Likelihood Sequence Estimation (MLSE) which is computed using Viterbi Algorithm (VA) [4]. GSM data receiver based on MLSE has been reported in [5], and presents common solution in nowdays realizations. Consequent efforts to reduce the complexity on the algorithmic level include the suboptimal sequence estimation approaches [6], decision-feedback equalization [7] and application of block detection techniques [8]. The generic MLSE-VA was extended by Hagenauer to include soft decision (SD) outputs to give the Soft-Output Viterbi Algorithm (SOVA) [9]. However, this algorithm and consequent simplification presented in [10] require memory for storing reliability information. One of the efficient way of using co-processor for MLSE is to provide soft information to the decoder without using the memory, as suggested in [11]. We propose new method for determining soft outputs targeting the reduction of complexity and efficient software-hardware partitioning of the algorithm [12]. The classical VA consists of three basic operations: • Calculation of the branch metric contribution (BMC) • Combine the BMC and the accumulated path metric (APM) and decide on which branch(s) to keep and which branch(s) to discard. • Update the APM so that it can be used in the next epoch. Soft information can be achieved by making use of the parameter D which is the difference between the “survivor APM” and the “discarded APM” as presented in [9]. The larger the metric difference, the more reliable is the “hard decision”. Block diagram of the receiver is given in Fig.5., where z(n) denotes the MF output. As there is an interleaver after the demodulator, it is also necessary to relay information concerning the slot’s average reliability. This can take the form of a SNR estimate of the individual symbols. However, this is difficult to estimate in general, but a slot duration based

SNR estimate would be easier. In order to compute the noise one requires a reference which is conveniently provided by the midamble bits. Thus in order to compute the SNR we can simply determine the difference between the expected midamble (which is in the form of + 1/-1), and the received midamble (which is in the form of real values in general) to give an estimate of the noise during the instant of the midamble.

184

The combination of the parameter D and the slot SNR can be combined into a single parameter, SD which is refered to as “soft decision” information: The unsigned soft decision information can be combined with the hard decision bits by

to give the soft output symbols where the sign information indicates the bit decision, and the

magnitude gives an indication of the confidence in the bit decision. If the values are to be routed along to different parts of the overall receiver system, then it would be desirable to quantise its range. This can be done by gathering the expect dynamic range of SD and setting up the appropriate ADC range to cover the region.

185

4.2.

Simplified Soft Information Generation Method

The method of computing the soft information SD is costly to implement in software or hardware as it requires either many instructions or hardware storage to store the array SD. One method to overcome this is to split up the task of generating the hard and soft information. This decoupled SOVA is shown in Fig.6. The hard information block (HIB) is same as the standard VA, however the soft information block (SIB) is activated after the completion of the HIB. The processing of the SIB is described below.

Referring to Fig.7, the objective is to determine the SD value at epoch node N. Assuming that the HIB has managed to perform reasonably in detecting the correct symbols, then the following observations can be made : • The paths that has led up to epoch node N originated from the same past state node and will end up at a known future node. • The distance between the originator node to the epoch node N is comparable to that of the CIR length L.

If the above is true then remembering that the ISI can be completely determined given knowledge of the past and future symbols and the CIR, then we can estimate what the difference is between the desired survivor and the discarded node. In the limit that there is no errors in the detected symbols, the estimated should be exactly the same as the true SD. Therefore, to determine the SD value at N, we simply sum up the APM beginning from the known start node to the known end node.

186

Let

The implementation of the above is however not trivial as it requires computation of many terms. Fortunately, many terms between the correct path APM and the discarded APM are common and can be ignored. By expanding above and eliminating the common terms the following is obtained :

The term is in fact not necessary as we are only interested in the magnitude of SD(N). As evident the computations required to calculate the soft information is very simple as it requires only a summation of the CIR coefficients. From the equation above it is clear that the SD is simply a measure of the ISI from the previous and future symbols which is what one expects intuitively.

5.

Simulation Results

In order to assess the applicability of the proposed algorithm extensive simulations were conducted using a commercially available package COSSAP. The reference performance curves are based on the COSSAP library MLSE implementation of a GSM receiver [13], which are compared to the GSM receiver detailed in this report. Propagation conditions described in [2] were tested, including static, typical urban at 3 and 50Km/h (TU3, TU50), rural area at 250Km/h (RA250), hilly terrain at 100Km/h (HT100) and equaliser test at 50Km/h (EQ50) channels. Also, in addition to these fading channels at so-

187

called reference receiver sensitivity input levels, adjacent and co-channel tests at reference interference input levels were also conducted.

The COSSAP GSM reference receiver model employs CIR estimation block as described in Section 3. However, it does not employ matched filtering and the Viterbi algorithm is based on a full Euclidean distance measure, utilising both the real and imaginary parts of the received signal. A full soft decision measure is employed together with an optional internal PLL operating on a sample by sample basis. Further details can be found in [13]. Few examples from extensive simulation set will be presented to illustrate major trends. Comparison of different receiver structures is presented in Fig.8. for TU50 propagation condition. The comparison is between parallel MLSE receiver, serial receiver using standard SOVA (SOVA1) and serial receiver with simplified soft decision calculation presented in this paper (SOVA2). The difference in the performance is within simulation error margin. For the optimization of hardware block the performance of the receiver is evaluated for different wordlengths used to quantise the soft decision information. Example is presented in Fig.9. It has been concluded that 4 bits of quantisation are sufficient to provide resolution in the co-processor that can be used both for equalization and decoding. The main points from the simulation results can be summarised below : • Overall performance of serial (real) MLSE is comparable to that of parallel (complex) MLSE. • The BER performance of SOVA1, SOVA2 and COSSAP is essentially the same. • For soft decision dependent measures (FER,RBER) the COSSAP (complex) MLSE performs slightly better (0.5 to 1 dB) for FER, but the difference is small for RBER. • The simpler SOVA2 is equivalent in performance to the full SOVA1 implementation. • Soft decision word length can be limited to 4-bits while maintaining comparable performance. • Eb/No of is required to satisfy the reference sensitivity conditions according to Rec.05.05. • reference interference performance is met using SOVA2.

188

6.

Conclusions

In this paper a concept of a simpler MLSE receiver was described. It was shown that the conventional MLSE can be broken down into a two-stage process to reduce computation while maintaining performance. The advantage of the new partitioning is that it enables a mix of hardware and software to implement the MLSE. The performance of the simplified SOVA2 receiver was assessed according to the GSM recommendations. The required Eb/No figure to meet the reference sensitivity level is comparable to the optimal solution, and should give the RF front end enough margin to allow a cost effective design.

References [1] Z.Zvonar and R.Baines, “Integrated Solutions for GSM Terminals”, International Journal of Wireless Information Networks, Vol. 3, No. 3, 1996, pp147-162. [2] GSM Technical Specifications, Rec.05.05, ETSI. [3] P.Laurent, “Exact and approximate construction of digital phase modulations by superposition of amplitude modulated pulses”, IEEE Trans. Commun., Vol. 34, 1986, pp 150160. [4] G.Ungerboeck, “Adaptive Maximum-Likelihood Receiver for Carrier-Modulated DataTransmission Schemes”, IEEE Trans. Commun., Vol. 22, No. 5, 1974, pp 624-636. [5] R.D’Avella, L.Moreno and M.Sant’Agostino, “An Adaptive MLSE Receiver for TDMA Digital Mobile Radio”, IEEE Journal Selec. Area Commun., Vol. 7, No. 1, 1989, pp 122-129. [6] G.Benelli, A.Fioravanti, A.Garzelli and P.Matteneini, “Some Digital receivers for the GSM

Pan-European Cellular Communication System,” IEE Proc. Commun., June 1994, pp 168176. [7] P.Bune, “A Low-Effort DSP Equalization Algorithm for Wideband Digital TDMA Mobile Radio Receivers,: In Proc. of ICC’91, pp 25.1.1-25.1.5. [8] B.Bjerke, J.Proakis, M.Lee and Z.Zvonar, “A Comparison of Decision Feedback Equalization and Data Directed Estimation Techniques for the GSM System,” In Proc. ICUPC’97, San Diego, CA, 1997. [9] J.Hagenauer and P.Hoeher, “A Viterbi Algorithm with Soft-Decision Outputs and its Applications”, In Proc. IEEE Globecom’89, Dallas TX, Feb. 1989, pp 47.1.1-47.1.7. [10] B.Rislow, T.Masen and O.Trandem, “Soft Information in Concatenated Codes,” IEEE Trans. Commun., Vol. 44, No. 3, 1996, pp 284-286. [11] S.Ono, H.Hayashi, T.Tanaka and N.Kondoh, “A MLSE Receiver for the GSM Digital Cellular System,” In Proc. VTC’94, Stockholm, 1994, pp 230-233. [12] M.Lee, Trellis Decoder with Soft Decision Output, patent pending. [13] COSSAP Model Libraries, Vol. 3, September 1996, Synopsys Inc.

189

190

7 Turbo Code Implementation Issues for Low Latency, Low Power Applications D. Eric Cress [email protected]

and William J. Ebel, Member IEEE [email protected] Mississippi State University Electrical and Computer Engineering Box 9571 Mississippi State, MS, 39762 USA Abstract: In this paper, four important and interrelated issues are discussed which relate to the performance of Turbo codes for low latency and low power applications: (1) interleaving, (2)

trellis termination, (3) estimation of the channel noise variance, and (4) fixed point arithmetic effects on decoder performance. We give a method for terminating both constituent convolutional encoders in a known (all zero) state by assigning specific binary values to information-sequence

bits that are dependent upon the full set of user input information bits. This method causes a slight restriction on the set of allowable interleavers that can be chosen for the scheme but does not

compromise performance. Also, we give a robust method for estimating the conditional channel variance given pre-thresholded random variable samples measured directly from the channel.

Finally, performance results are shown for fixed-point number representations.

A. Introduction

An exciting development in recent years in the field of error correcting codes was the introduction of Turbo codes [1]. Empirical results indicate that these codes approach the Shannon limit for reliability improvement on an AWGN channel. In this paper, four important and interrelated issues are discussed which relate to the performance of Turbo codes for low latency and low power applications: (1) interleaving, (2) trellis termination, (3) estimation of the channel

noise variance, and (4) fixed point arithmetic effects. A typical Turbo encoder is shown in Figure 1. The binary data sequence x is input to a rate 1/2 recursive convolutional encoder (RCC) and at the same time it is input to an interleaver which generates the output sequence y and subsequently input to a second RCC. The output consists of the original data along with the parity resulting from the two constituent convolutional

191

encoders.

For the purposes of this discussion, the modulator is taken to be binary and is implemented by the mapping z = 2x-l where z represents a transmitted symbol and x represents a binary number. In this way, a binary 1 maps to the real number 1 and a binary 0 maps to the real number -1. Therefore the transmitter power is

= 1 per transmitted bit.

The channel is taken to be AWGN and is implemented by adding a Gaussian random variable (RV) to each transmitted symbol. If noise ratio (SNR) of the system is and, sequences with variance

is the variance of the added noise RV, then the signal-toThe received sequences are given by where

and

are independent Gaussian noise

We will use primes to denote variables that have an added noise

component. The parallel configuration of the encoder along with the systematic implementation of the convolutional encoders allows the decoder to operate in an iterative fashion. Figure 2 illustrates, in concept, a simplified Turbo decoder. First the received input sequence parity sequence

and the received

are paired and input to the first Maximum A-Posteriori (MAP) decoder [2][3].

In concept, the output is an estimate of the original input data in the form of a probability measure that each bit in the sequence is a binary “1”. The rest of the decoder operates as shown. Turbo codes have traditionally been shown to yield remarkable performance for long blocklengths (large interleaver), usually on the order of many tens of thousands of bits [4][5], i.e. 192

to

bits. The large interleaver is used to reduce the multiplicities of low weight

codewords, known as spectral thinning [6], however it also plays an important secondary role in eliminating the effect of poor decoded-bit estimates caused by unknown or unreliable constituent

convolutional encoder state terminations. To be more specific, the constituent convolutional

encoders of a Turbo code are generally configured to be terminated using tail bits that are transmitted in addition to the data sequence and parity sequences. These tail bits do not benefit from the diversity effect of the interleaver in the iterative decoder and result in poor decoded bit estimates. To illustrate this, Figure 3 shows the number of errors that occurred per information bit

position in the data sequence for a rate 1/3 Turbo code using 8-state recursive convolutional codes, an information sequence length of 30 bits, and with an SNR of 1.5dB. The algorithm using tail bits for trellis termination resulted in 8,262 bit errors out of 330,000 total bits. The algorithm

using full interleaving with trellis termination, described in Section B below, resulted in 1,449 bit

errors out of 300,000 total bits. The bulk of the additional errors resulted from the poor tail bit estimates and their residual effect on other bits. When large interleavers are used, these poor estimates have a negligible effect on the overall performance of the code. In wireless applications, however, the blocklengths are necessarily much smaller, on the order of a few hundred bits or less. The main issue in any practical solution is code performance per unit

complexity of the hardware realization, especially in a commercial application involving handheld, battery-powered electronic devices. In Section B, we describe a method for terminating the constituent convolutional encoders to improve the performance in short blocklength

193

applications.

In this paper, we also give a robust method for estimating the channel noise variance directly from the received data. This method is empirically shown to yield decoded error probability performance that is nearly identical to that which results when using the true value.

Finally, we show fixed-point arithmetic results for Turbo decoders. If the hardware

implementation of a Turbo decoder is not configured carefully, unstable behavior can be observed. We show that for some simple codes, as few as 4 bits of precision provide a reasonable coding

gains suggesting an interesting trade-off between complexity and performance.

B. Interleaving and Trellis Termination In this section, an algorithm is described as first suggested by Barbelescue [7], which terminates the trellis of the two constituent encoders by properly selecting the first 2m bits of the

information sequence, where m is the number of delays in each constituent encoder. We call these first 2m bits precursor bits. A Turbo encoder using precursor bits for trellis termination is shown

in Figure 4. The precursor bits are denoted by the length-2m sequence s. The encoder requires two passes. In the first pass, the precursor bits are set to zero and data is sequenced into the two constituent RCC’s without regard for the final state. After the data is input, the final state of the

each encoder is sent to a ROM where the precursor bits are read out and positioned at the 194

beginning of the input sequence. These precursor bits will guarantee that the two constituent

RCC’s end up in the all zero state at the end of the second encoding pass. The ROM is stored off line by iterating through all possible precursor sequences s and storing the final encoder state for

each RCC. Since the encoder is linear, the second pass will result in final RCC states which are the sum of the final state due to the data and the precursor bits. Since these final states are identical and the sum of two identical GF(2) numbers is always zero, this results in final RCC states that are zero.

There is only one issue to contend with here. In order for this to work, the off-line procedure for building the ROM must result in a one-to-one correspondence between each precursor binary number and each of the possible states in the second RCC. It is possible that the random

interleaver will position the precursor bits in y in such a way so that iterating through all possible precursor binary numbers does not result in an exhaustive set of final states for the second RCC.

We say that such an interleaver is not proper. In our simulation, the random interleaver is reselected until a proper one is found. Each interleaver is checked to see if the set of possible precursor bits result in a duplicate encoder state for the second RCC. In our simulation, the

195

interleaver is generally reselected from 0 to 8 times before a proper one is found.

In this configuration, the entire input sequence, including the precursor bits, is interleaved.

This eliminates the bias in the error probability introduced by the non-interleaved trellis tail bits

for trellis termination. Also note that the code rate for this configuration is

which is

approximately 1/3 for large N. The Turbo decoder corresponding to this Turbo encoder is shown in Figure 5. The

permutation resulting from the interleaver and deinterleaver are identical to that used in the

encoder except that the entries are real numbers rather than binary numbers.

C. Channel Noise Variance Estimation

In this section, a method for estimating both the conditional mean and conditional variance for the received data which enters the Turbo decoder is described. The problem can be stated as follows. Let Z be a random variable (RV) with pdf given by

196

where

is the Gaussian pdf with mean m, variance , and with independent variable z.

The goal is to estimate the conditional statistics given only measured statistics of RV Z. An illustration of

computing m and

is shown in Figure 6. If

is small, then an efficient method for

is to find the mean and variance for the

The error caused by the tail

overlap will be negligible. However, a more practical situation occurs when the conditional variance is large, corresponding to a low channel SNR. In this case, the parameters can be estimated by computing the

and

moment of Z, denoted

and

respectively.

Since the mean of Z is zero, the second moment is

Each integral represents the second central moment of a Gaussian RV with mean m and variance and evaluates to the same expression, given by

Therefore, the second moment of Z is (1) Since

is directly measurable from a set of samples of Z, this gives one equation in terms of the

unknowns m and

Similarly, the

moment of Z can be shown to be (2) 197

Solving (1) and (2) simultaneously, gives

and

Of course, since the measured moments are themselves RV’s, there is always a chance that these

calculations will fail. Surely we must have fail, which is most probable if

If either of these conditions

is small, then the first method should be used to estimate m

and In Table I below, the true channel noise variance is compared with the estimate using the method outlined in this section. The code chosen has a blocklength of 100 bits and 10 blocks were Table I. Comparison of true and estimated channel noise variance

combined to form the variance estimate. As the SNR increases, the estimate smoothly converges to the true estimate. In any case, the estimate is close to the true SNR value above 0dB and there was no appreciable difference in the performance of the decoded error probability when the estimated variance was used in place of the true variance.

D. Fixed-Point Arithmetic Results In Table II below, some experimental results for a hardware realizable turbo coding system are shown. The system uses a log-likelihood ratio decoder to minimize the hardware complexity by

198

reducing costly fixed-point divides and multiplies. The code used is an optimal convolutional code of memory order 2. The interleaver is a block length 100 pseudo-random interleaver Data is presented for varying degrees of fixed-point precision. The coding gain that is presented is for a channel modeled with AWGN and a signal-to-noise ratio (Eb/No) of 1.76dB. Note that the performance results for a dynamic range of less than -7 to 8 was very poor.

These results show that the performance loss due to the fixed-point precision is tolerable down to

4 bit numbers for the specific Turbo code implemented. These results also show that the block

error probability increased by a factor of 2 as the fixed-point number size was reduced from 8 bits

to 4 bits. This suggests that with 4-bit numbers, there were twice as many block decoding failures but there were not many bit errors in the additional block errors.

E. Conclusions

In this paper, four important and interrelated issues were discussed which relate to the 199

performance of Turbo codes for low latency and low power applications: (1) interleaving, (2) trellis termination, (3) estimation of the channel noise variance, and (4) fixed point arithmetic

effects on decoder performance. We described a method for terminating both constituent convolutional encoders in a known (all zero) and also gave a robust method for estimating the

conditional channel variance given pre-thresholded random variable samples measured directly from the channel. Finally, performance results are shown for fixed-point number representations and arithmetic. Bibliography [1] Berrou, C., Glavieux, A., and Thitimajshima, P., “Near Shannon Limit Error-Correcting Coding and Decoding: Turbo-Codes (1)”, International Communications Conference, Geneva Switzerland, 1993, pp. 1064-1070.

[2] Bahl, L.R., Cocke, J., Jelinek, F., and Raviv, J., “Optimal Decoding of Linear Codes for Minimizing Symbol Error Rate”, IEEE Transactions on Information Theory, March 1974, pp. 284-287. [3] Forney, G.D. Jr., “The Forward-Backward Algorithm”, in Proc. 34th Allerton Conf. Commun., Contr., Computing, Allerton, IL, Oct. 1996. [4] Divsalar, D., and Pollara, F., “On the Design of Turbo Codes”, JPL TDA Progress Report 42123, JPL, November 15, 1995, pp. 99-121. [5] Divsalar, D., and Pollara, F., “Turbo Codes for Deep-Space Communications” TDA Progress Report 42-120, JPL, February 15, 1995, pp. 29-39. [6] Perez, L.C., Seghers, J., and Costello, D.J., “A Distance Spectrum Interpretation of Turbo Codes”, IEEE Transactions on Information Theory, Vol. 42, No. 6, November 1996, pp. 16981709.

[7] Barbulescu, A.S., and Pietrobon, S.S., “Interleaver design for turbo codes”, Electronics Letters, Vol. 30, No. 25, December 8, 1994, pp. 2107-2108.

200

8 Evaluation of the Ad-Hoc Connectivity with the Zone Routing Protocols Zygmunt J. Haas and Marc R. Pearlman School of Electrical Engineering, Cornell University, 323 Rhodes Hall, Ithaca, NY, 14853 Tel: (607) 255-3454, fax: (607) 255-9072, e-mail: [email protected] URL: http://www.ee.cornell.edu/~haas/wnl.html

Abstract In this paper, we evaluate the novel routing protocol for a special class of ad-hoc networks, termed by us the Reconfigurable Wireless Networks (RWNs). The main features of the RWNs are: the increased mobility of the network nodes, the large number of nodes, and the large network span. We argue that the current routing protocols do not provide a satisfactory solution for routing in this type of an environment. We propose a scheme, coined the Zone Routing Protocol (ZRP), which dynamically adjusts itself to the operational conditions by sizing a single network parameter - the Zone Radius. More specifically, the ZRP reduces the cost of frequent updates of the constantly changing network topology by limiting the scope of the updates to the immediate neighborhood of the change – the Zone Radius. We study the performance of the scheme, evaluating the average number of control messages required to discover a route within the network. Furthermore, we compare the scheme’s performance, on one hand, with reactive flooding-based schemes, and, on the other hand, with proactive distance-vector schemes. 1. Introduction

A Reconfigurable Wireless Network (RWN) is an ad-hoc network architecture that can be rapidly deployed without relying on preexisting fixed network infrastructure. The nodes in a RWN can dynamically join and leave the network, frequently, often without warning, and without disruption to other nodes’ communication. Finally, the nodes in the network can be highly mobile, thus rapidly changing the node constellation and the presence or absence of links. Examples of the use of the RWNs are: • tactical operation - for fast establishment of military communication during the deployment of forces in unknown and hostile terrain; • rescue missions - for communication in areas without adequate wireless coverage; • national security - for communication in times of national crisis, where the existing communication infrastructure is non-operational due to a natural disaster or a global war; • law enforcement - for fast establishment of communication infrastructure during law enforcement operations; • commercial use - for setting up communication in exhibitions, conferences, or sale presentations; • education - for operation of wall-free (virtual) classrooms; and • sensor networks - for communication between intelligent sensors (e.g., mounted on mobile platforms. Nodes in the RWN exhibit nomadic behavior by freely migrating within some area, dynamically creating and tearing down associations with other nodes. Groups of nodes that have a common goal can create formations (clusters) and migrate together, similarly to military units on missions or This work is supported by the US Air Force/Rome Labs, under the contract number C-7-2544 and a grant from Motorola Corporation, the Applied Research Laboratory. 2 Micro-Electro-Mechanical-Systems

201

similarly to guided tours on excursions. Nodes can communicate with each other at anytime and without restrictions, except for connectivity limitations and subject to security provisions. Examples of network nodes are pedestrians, soldiers, or unmanned robots. Examples of mobile platforms on which

the network nodes might reside are cars, trucks, buses, tanks, trains, planes, helicopters, ships, UAVor UFO-s. In this paper, we concentrate on the issue of designing a routing protocol for the RWN. In particular, we address routing in a flat ad-hoc networks, as opposed to hierarchical ad-hoc networks that have been investigated in the past (e.g., [Lauer86,Westcott84]). The proposed protocol, the Zone Routing Protocol, allows efficient and fast route discovery in the RWN communication environment (i.e., large geographical network size, large number of nodes, fast nodal movement, and frequent topological changes). In what follows, we explain the elements of the proposed scheme.

2. Previous and Related Work In this work, we address routing in a flat ad-hoc networks, as opposed to hierarchical ad-hoc networks that have been investigated in the past (e.g., [Lauer86,Westcott84]). Although routing in hierarchical ad-hoc networks involves simpler procedure, some salient features of the flat architectures, as mentioned above, make them that much more attractive for communication in the RWN environment. Comparison of the two architectures is outside the scope of the paper and the reader is referred to [Haas98] for further discussion on this topic. The wired Internet uses routing protocols based on topological broadcast, such as the OSPF [Moy97]. These protocols are not suitable for the RWN due to the relatively large bandwidth required

for update messages. Routing in multi-hop packet radio networks was based in the past on shortest-path routing algorithms [Leiner87], such as Distributed Bellman-Ford (DBF) algorithm [Bertsekas92]. These algorithms suffer from very slow convergence (the “counting to infinity” problem). Besides, DBF-like algorithms incur large update message penalties. Protocols that attempted to cure some of the shortcomings of DBF, such as Destination-Sequenced Distance-Vector Routing (DSDV) [Perkins94], were proposed. However, synchronization and extra processing overhead are common in these protocols. Other protocols that rely on the information from the predecessor of the shortest path solve the slow convergence problem of DBF (e.g., [Cheng89] and [Garcia-Luna-Aceves93]). However, the processing requirements of these protocols may be quite high, because of the way they process the

update messages. Routing protocols that are based on a source initiated query-reply process have also been introduced. Such techniques typically rely on the flooding of queries to discover a destinatio. In [Corson97] the route replies generated are also flooded, in a controlled manner, to distribute routing information in the form of directed acyclic graphs (DAGs) rooted at each destination. In contrast, other schemes unicast the route reply back to the querying source, typically by means of reversed routing information gathered during the query phase. In the case of [Perkins97], this routing information is in the form of next-hop routes to the querying node, while in [Johnson96], a route accumulation procedure is employed during the route query, allowing the route reply to be returned via source routing. The on-demand discovery of routes can result in much less traffic than standard distance vector or link state schemes, especially when innovative route maintenance schemes are employed. However, the reliance on flooding may still lead to considerable control traffic in the highly versatile RWN environment. [Murthy95] and [Murthy] present a new distance-vector routing protocol for packet radio networks (WRP). Upon a change in the network topology, WRP relies on communicating the change to its neighbors, which effectively propagates throughout the whole network. The salient advantage of WRP is the considerable reduction in the probability of loops in the calculated routes, as compared with 3

Unmanned Aerial Vehicles

202

other known routing algorithms, such as, for example, DBF. Compared with our routing protocol, the main disadvantage of WRP is in the fact that routing nodes constantly maintain full routing information in each network node, which was obtained at relatively high cost in wireless resources. Our protocol, in contrast, rapidly finds routes, only when transmission is necessary. Moreover, multiple routes are maintained, so that when some of these routes become obsolete, other routes can be immediately utilized. This is especially important when the network contains large number of very fast moving nodes, as is the case in the RWN architecture.

3. The Notion of a Routing Zone and Intrazone Routing A routing zone is defined for each node and includes the nodes whose minimum distance in hops from the node in question is at most some predefined number, which is referred to here as the zone radius. An example of a routing zone (for node S) of radius 2 is shown in Figure 1. Note that in this example nodes A through K are within the routing zone of S. Node L is outside S’s routing zone. Peripheral nodes are nodes whose minimum distance to the node in question is equal exactly to the zone radius. Thus, in Figure 1, nodes G-K are peripheral nodes. Zones of different nodes overlap heavily. Related to the definition of a zone is the coverage of a node’s transmitter, which is the set of nodes that are in direct communication with the node in question. These nodes are referred to as neighbors. The transmitter’s coverage depends on the propagation conditions, on the transmitter power, and on

the receiver sensitivity. In our simulation, we define conceptually a radius,

which is the

maximal distance that a node’s transmission will be received without errors. Of course, it is important that each node be connected to at least one other node. However, more is not, necessarily, better. As the transmitter’s coverage includes all the nodes with distance 1 hop from the node in question, the larger the is, the larger is the content of its routing zone. A large routing zone requires large amount of update traffic. For the purpose of simplification, we will depict zones as circles around the node in question. However, one should keep in mind that the zone is not a description of distance, but rather nodal

connectivity (measured in hops). Each node is assumed to maintain the routing information to all nodes that are within its routing zone and those nodes only. Consequently, in spite of the fact that a network can be quite large, the updates are only locally propagated. We assume that the protocol through which a node learns its zone is some sort of a proactive scheme, which we refer to here as the IntrAzone Routing Protocol (IARP). In this paper, we use a modification of the Distance Vector algorithm. However, any other proactive scheme would do. Of course, in principle, the performance of the ZRP depends on the choice of IARP. However, our experience suggests that the tradeoffs are not strongly affected by the particular choice of the proactive scheme used. 3.1 Interzone Routing and the Zone Routing Protocol IARP finds routes within a zone. The IntErzone Routing Protocol (IERP), on the other hand, is responsible for finding routes between nodes located at distances larger than the zone radius. IERP relies on what we call bordercasting. Bordercasting is a process by which a node sends a packet to all its peripheral nodes. A node knows the identity of its peripheral nodes by the virtue of the IARP. Bordercasting can (and should) be implemented by multicasting, if multicasting is supported within the subnet.4 Alternatively, unicasting the packet to all the peripheral nodes achieves the same goal, albeit at much higher cost in resources. 4 It is not clear whether multicasting is, indeed, feasible in a highly dynamic network topology. Examination of applicability of multicasting in ad-hoc networks is outside the scope of this paper. Here, we assume that bordercasting is performed using unicasting.

203

The IERP operates as follows: The source node first checks whether the destination is within its zone.5 If so, the path to the destination is known

and no further route discovery processing is required. If the destination is not within the source Routing Zone, the source bordercasts a route request (which we call simply a request) to all its

peripheral nodes.6 Now, in turn, all the peripheral nodes execute the same algorithm: check whether the destination is within their zone. If so, a route reply (which we call simply a reply) is sent back to the source indicating the route to the destination (more about this in a moment). If not, the peripheral node forwards the query to its peripheral nodes, which, in turn, execute the same procedure. An example of this Route Discovery procedure is demonstrated in Figure 2. The source node S sends a packet to the destination D. To find a route within the network, S first checks whether D is within its routing zone. If so, S knows the route to node D. Otherwise, S sends a query to all the nodes on the periphery of its zone; that is, to nodes C, G, and H. Now, in turn, each one of these nodes, after verifying that D is not in its routing zone forwards the query to its “peripheral” nodes. In particular, H sends the query to B, which recognizes D as being in its routing zone and responds to the query, indicating the forwarding path: S-H-B-D. A nice feature of this distributed route discovery process is that a single route query can return

multiple route replies. The quality of these returned routes can be determined based on hop count (or any other path metric7 accumulated during the propagation of the query. The best route can be selected based on the relative quality of the route (e.g., choose the route with the smallest hop count, or shortest accumulated delay). Two main issues need to be addressed: When sending a reply to the source node, how does the “last peripheral node” know the whole path, to be included in the reply to the source? (A related question is, how does the responding node know how to send the reply to the source?) The second question is, how does the IERP process terminate? Let us start with the first question. The process by which the node receiving a query knows the path back to the source of the query is the Route Accumulation procedure. In the Route Accumulation procedure, each node that forwards a query writes into the query packet its identification. The sequence of these identifications represents a path from the source node to the current node, and, by reversing the order, a path from the current node to the source node. Thus, the routes within the network are specified as a sequence of nodes, separated by approximately the zone radius. A node, which identifies that the destination is in its zone, simply adds its own identification to the query and returns the accumulated route to the source. The second issue, that of termination of the IERP process, is a more difficult one. Of course, similarly to the standard flooding algorithm, a node that previously received the query will discard it. This, however, does not solve the whole problem, since as the zones heavily overlap, the query will be forwarded to many network nodes. In fact, it is very possible that the query will be forwarded to all the network nodes, effectively flooding the network. But a more disappointing result is that, due to fact that bordercasting involves sending the query over a path of length equal to the zone radius, the IERP

5

6 7

Remember that a node knows the identity, distance to, and a route to all the nodes in its zone. Again, the identity of its zone peripheral nodes are known to the node in question. Typical path metrics include hop count, delay, capacity, etc.

204

will result in much more traffic than the flooding itself! What is needed is a more efficient termination criterion.

Let us look at this problem more closely. The idea behind IERP is that the search for a node advances in the “quantum” of zone radius, instead of flooding the network by forwarding the query among neighbors. The gain that we expect is due to the fact that only some network nodes will be involved in such a “flood.” The challenge is to “steer” the search in the direction outwards of the original Routing Zone (see Figure 3), rather than going back into the areas that were already covered by other threads of the search. There are a number of ways that such a redirection of the search could be accomplished. We discuss here two possibilities. The first improvement, termed the Backwards Search Prevention (BSP), makes sure that peripheral nodes of the current node that lie within the routing zone of the previous node are not included in the next bordercast. To prevent the backward propagation of queries, a bordercasting node must send its queried peripheral nodes a list of its routing zone nodes (perhaps appended to the IERP query packet). Thus, in the example in Figure 4, after S bordercasts to the nodes F and C, the nodes A, C, and S will not be included in the consecutive bordercast by node F, as the nodes A, C, and S are all within the routing zone of the previous bordercasting node S. While the BSP may be impractical for large routing zones (due to the long list of routing zone nodes), it could be quite effective when used by nodes which maintain smaller routing zones.

The second improvement, which we call the Loopback Search Prevention (LSP), involves pruning any search that goes into areas previously searched. This is accomplished by terminating bordercast at nodes that either have received the query before or that have overheard the query transmitted by their

205

neighbors.8 Note that this includes all threads of the query. Additional modification include termination of a thread at nodes whose routing zone include any of the nodes in the currently accumulated path. An example is shown in Figure 5. In this example, S bordercasts to A, which bordercasts to B, which in turn bordercasts to C. C will terminate the search of this thread (i.e., will stop bordercasting), as S, who is in the thread's accumulated route, is within its routing zone. Both of these improvements reduce the amount of control traffic of the IERP protocol. From our simulation runs, we have learned that the contribution of the two schemes in reducing the control traffic is approximately equal. Note that the two techniques are not overlapping. The Backwards Search Prevention avoids sending a route request to nodes that should not forward it. On the other hand, the Loopback Search Prevention would send the route request to such a node, but will subsequently terminate the search thread at this node. The main advantage of the Zone Routing Protocol is in the fact that the number of "flood" messages to discover a route is significantly smaller, as compared with other reactive-type protocols. This decrease is due to the directed propagation of queries to specified peripheral nodes. Since for radius greater than one the routing zones heavily overlap, the routing tends to be extremely robust. Zone Routing, as described earlier, discovers multiple routes to a destination. However, the Route Discovery process can be made much more efficient in resources, at the expense of longer latency. This could be done by sequentially, rather than simultaneously, querying the peripheral zone nodes, either one-by-one or in groups. Thus, there is a tradeoff between the cost and the latency of the Route Discovery procedure. We omit here correctness proof of the ZRP. Am interested reader is referred to [Haas98-2]. 3.2 The Route Maintenance Procedure In the Route Discovery procedure, each node, proactively and continuously learns the topology within its zone radius and, reactively, on-demand, discovers routes by hopping in steps of the routing radius. Because the number of nodes within a zone is much smaller than the number of network nodes, the penalty for dissemination of routing information within a zone is limited. So is the cost of the route discovery, when the zone radius is sufficiently large. For a small radius (zone radius =1), the ZRP behaves as a reactive scheme (flooding). On the other extreme, for a large radius the scheme exhibits proactive behavior. In general, the size of the zone radius determines the ratio between the proactive and reactive behavior of the protocol. The Route Maintenance Procedure adaptively adjusts the zones' radii, as to reduce the "cost" of the Route Discovery Procedure. The adjustment may be performed, based on the value of Call-to-Mobility-Ratio (CMR) measured independently at each node. CMR is a ratio of the rate at which queries are initiated to the rate at which connections with the neighbors are broken. Large CMR indicates that the network mobiles are very active in connection initiation and, thus, larger zone radius would decrease its frequent route discovery costs. Small CMR suggests that mobiles rarely place outgoing connections and, to reduce the overall cost of learning the routing within the nodes' routing zones, a smaller zone radius is preferable. Similarly, for fast moving mobiles (small CMR) the local zone routing information becomes obsolete quickly. Thus, a smaller zone radius carries smaller penalty. The routing zone radii may be configured prior to network deployment, based on a priori knowledge of network call activity and mobility patterns. Alternatively, and more typically, the routing zones may be resized dynamically, allowing the ZRP to adapt to local changes in call activity or node mobility.9 The Route Maintenance Procedure also significantly reduces the routing costs by employing the Route Discovery procedure only when there is a substantial change in the network topology. More specifically, active routes are cached by nodes: the communicating end nodes and intermediate nodes. 8 This is done by having each node eavesdropping on all its neighbor communications and requires that the IERP communicate with the MAC layer. 9 Dynamic adjustment of the routing zone radius requires minor modifications to the basic Zone Routing Protocol. Discussion of these enhancements is outside the scope of this paper.

206

Inactive paths are purged from the caches after some timeout period.10 Upon a change in the network topology, such that a link within an active path is broken, a local path repair procedure is initiated. The

path repair procedure substitutes a broken link by a mini-path between the ends of the broken link. A path update is then generated and sent to the end points of the path. Path repair procedures tend to

reduce the path optimality (e.g., increase the length for shortest path routing). Thus, after some number of repairs, the path end points will initiate a new Route Discovery procedure to replace the path with a new optimal one. 4.0 Evaluation of the ZRP

We use the OPNET™ Network Simulator from MIL3, an event driven simulation package, to evaluate the performance of the ZRP over a range of routing zone radii, from reactive routing to proactive routing Performance is gauged by measuring the control traffic generated by the ZRP and its effects on the average session delay. Our results can be used to determine the optimum ZRP routing zone radius for a given nodal velocity and for a given route query rate. The ZRP control traffic consists of the intrazone (IARP) route update packets and the interzone (IERP) route request/reply/failure packets. While the neighbor discovery beacons could be considered control overhead, this additional traffic is independent of both mobile velocity and routing zone radius. Furthermore, the neighbor discovery process is not an exclusive component of the ZRP; various MAC protocols are also based on neighbor discovery. As such, the beacons do not contribute to the relative

performance of the ZRP and are not accounted for in our analysis. Because the IERP packets are variable length (due to the route accumulation procedure), we measure control traffic in terms of node ID fields, rather than packets. A meaningful measure of ZRP delay is the average route query response time which is

defined as the average duration from the time a route is initially requested by the Network layer until the route is discovered.11 If the destination appears in the routing tables (which will occur with probability (l–Prob[route discovery])), the query is immediately answered and the route query

response time is assumed to be zero.12 Otherwise, a route discovery is required (which will occur with probability (Prob[route discovery])) and the route query response time is measured as the time elapsed between the generation of the route request and the reception of the first route reply,

For a fixed network size and fixed nodal density, the probability of a route discovery for an initial query is only dependent on the routing zone radius. The behavior of the route reply time is far more

complicated. Not only is it dependent on the arrival rate of control packets, it is also affected by such factors as the network traffic load and the average length of IERP control packets. Our study provides some insight into the effect of these factors on the ZRP delay. Our simulated RWN consists of 52 mobile nodes, whose initial positions are chosen from a uniform random distribution over an area of 600 [m] by 600 [m]. Each node j moves at a constant 10

The determination of what constitutes an "active" or “inactive” path depends on the CMR of the network

nodes. A cache management algorithm that determines the path activity is outside the scope of this paper. 11 This delay metric does not reflect the delays associated with subsequent route repairs. We assume here that routes can be adequately repaired through the local route repair procedure described earlier. These limited depth

queries produce much less control traffic and much lower delays compared with the initial full depth query. We assume that the local processing time (e.g., table lookup) is negligible, compared with transmission delays.

207

speed, v, and is independently assigned an initial direction13, which is uniformly distributed between 0

and

When a node reaches the edge of the simulation region, it is reflected back into the coverage

area. Each simulation runs for duration of 125 seconds. No data is collected for the first 5 seconds of the simulation to avoid measurements during the transient period and to ensure that the initial intrazone route discovery process stabilizes. In order to measure the delay resulting only from the ZRP overhead, the network load is assumed to be low. Route failures are detected and acted upon. The route queries are generated according to a Poisson arrival process, with the arrival intensity being a simulation parameter. The route queries represent both the initial query performed at the beginning of a session and subsequent queries due to reported route failures. Each route query is for a destination selected from a uniform random distribution of all other nodes in the network. Since the average time between a node’s query for the same destination is longer than the expected interzone route lifetime, discovered interzone routes are effectively used only once and then discarded. For the purposes of our simulation, we have made a number of simplifying assumptions regarding the behavior of the lower network layers and channel. This simplified model helps to improve understanding of our routing protocol behavior by providing our performance measures with some immunity from lower layer effects. From the media access control (MAC) perspective, we assume that there is no channel contention. This assumption is necessary to separate the delays associated with a particular MAC scheme (e.g., collision avoidance algorithms) from the delays related to the routing protocol. These MAC independent results could be used as a benchmark for future analysis of the interaction between the Routing and the MAC layers Our assumption of a collision-free media access protocol means that the average SIR of a received packet is limited by the ambient background noise and receiver noise. For fixed transmitter and noise powers, we assume that the BER is reasonably low within a distance, which we call dxmit. Beyond dxmit, the BER increases rapidly. This behavior results from a rapid decrease in received power as the separation distance is increased. We approximate this rapid increase in BER by the following simplified path loss model:

We interpret this behavior as follows: any packet can be received, error-free, within a radius of dxmit from the transmitter, but is lost beyond dxmit. Since packet delivery is guaranteed to any destination in range of the source, we are able to further reduce the complexity of our model by eliminating packet retransmission at the data link level.

13

Direction is measured as an angle relative to the positive x-axis.

208

5.0 Performance Results

Results of our simulation are presented in the following figures. Figure 11 shows the dependence of intrazone control traffic on the routing zone radius, for various rates of network reconfiguration. All else being equal, the rate at which the networks reconfigure increases linearly with the speed of the nodes. For unbounded networks with a uniform distribution of nodes, we expect the increase in intrazone control traffic to be proportional to However, because our network is of finite size and the nodes are distributed randomly, we find that increase is actually somewhere between It should be noted that there is no intrazone control overhead for All nodes within a routing zone of are, by definition, neighbors. Consequently, the Neighbor Discovery Protocol provides all of the information needed to maintain connectivity within the routing zone. The performance of the reactive portion of the ZRP is exhibited in Figure 12. As we increase the routing zone radius, we find that the rate of interzone control traffic decreases. This decrease can be attributed to three factors. First, as the size of the routing zone increases, more destinations can be found within a routing zone, requiring fewer IERP route requests. Second, as routing zones become larger, the redundant route query traffic is reduced through the increasingly directed propagation of queries to peripheral nodes. Lastly, the average number of peripheral nodes between a querying source and destination is inversely proportional to the routing zone radius. Thus, as the routing zone radius increases, the IERP accumulated routes are specified, on average, by fewer node IDs. For all nodes are peripheral nodes and bordercasting is equivalent to flooding. For we observe a significant reduction in the interzone control traffic, indicating a potential benefit from a hybrid routing scheme compared to purely reactive routing. The total control traffic (i.e., the sum of the control packets from the intrazone and interzone protocols), depicted in Figures 13, gives an indication of the performance of our hybrid routing scheme. For low route query rates, we find that relatively reactive routing (i.e., small routing zone radii) produces the least amount of control traffic. As the route query rate increases, the control over-

209

overhead can be minimized by increasing the routing zone radius. For the network configurations and operational conditions that we assume in our simulation, configuring the ZRP for a routing zone of

is shown to reduce the rate of ZRP control traffic by approximately 45% of the purely reactive schemes. For larger the overhead required to maintain larger routing zones outweighs the benefits gained from bordercasting

Figures 14 show the performance of the ZRP as measured by the average route query response time. The delay characteristics appear to be heavily influenced by the behavior of the interzone route discovery protocol. Under the conditions that the average amount of control traffic is small to moderate (small routing zones, small query rates and moderate nodal velocities), most of the instantaneous network load is due to a single route discovery. When the routing zones become relatively large and the network topography more volatile, the overall ZRP control traffic becomes large and begins to have a noticeable impact on the instantaneous network load. This behavior is exhibited at (Figure 14c). We note that for the load of 0.5 and of 1.0 [queries/second] (representing short route lifetimes), a minimum in average route query response time appears at Although we’ve simulated the ZRP for a medium sized networks with routing zones of hops, it is reasonable to

210

assume that for wider networks and larger routing zones, minimum will also be present for relatively low velocities and low route query rates as well. Neglecting the effects of additional data traffic, we find that the ZRP can provide as much as a 50% reduction in the average route query response time compared with purely reactive routing. This improvement is somewhat smaller for networks with highly dynamic topologies, but even our most volatile networks exhibit significant improvements of 38% compared to purely reactive routing.

6.0 Summary and Concluding Remarks The Zone Routing Protocol (ZRP) provides a flexible solution to the challenge of discovering and maintaining routes in the Reconfigurable Wireless Network communication environment. The ZRP combines two radically different methods of routing into one protocol. Interzone route discovery is based on a reactive route request/route reply scheme. By contrast, intrazone routing uses a proactive protocol to maintain up-to-date routing information to all nodes within its routing zone.

211

The amount of intrazone control traffic required to maintain a routing zone increases with the size of the routing zone. However, through a mechanism, which we refer to as bordercasting, we are able to exploit the knowledge of the routing zone topography to significantly reduce the amount of interzone control traffic. For networks characterized by highly mobile nodes and very unstable routes, the hybrid proactive-reactive routing scheme produces less average total ZRP control traffic than purely reactive routing Purely reactive schemes appear to be more suitable for networks with greater route stability. Furthermore, for highly active networks (frequent route requests), more proactive networks produce less overhead (i.e., larger routing zones are preferred).

We note that for networks with low activity, the instantaneous network load is generally dominated by the control traffic from a single route discovery. Consequently, the ZRP exhibits minimum delay for relatively large routing zone radii (for the networks we simulated, even for

cases where relatively reactive routing minimizes the average ZRP control traffic. For highly volatile networks, the ZRP has been shown to provide 38% less delay than reactive routing. For slower, more stable networks, the optimal-delay ZRP configuration produced a nearly 50% reduction in delay compared to reactive routing. Based on the performance of the ZRP under heavy control traffic, we expect that additional data traffic will further reduce the optimal size of the routing zone.

7.0 References [Bertsekas92] D. Bertsekas and R. Gallager, Data Networks, Second Edition, Prentice Hall, Inc., 1992. [Cheng89] C. Cheng, R. Reley, S.P.R. Kumar, and J.J. Garcia-Luna-Aceves, “A Loop-Free Extended BellmanFord Routing Protocol without Bouncing Effect,” ACM Computer Communications Review, 19(4), 1989, pp.224-236.

[Corson97] M.S. Corson and V. Park, “Temporally - Ordered Routing Algorithm (TORA) Version 1 Functional Specification,” IETF MANET Internet Draft, Dec. 1997 [Ephremides87], A. Ephremides, J.E. Eieselthier, and D.J. Baker, “A design concept for reliable mobile radio networks with frequency hopping signaling," Proceedings of the IEEE, vol.75, pp.56-73, January 1987. [Garcia-Luna-Aceves93] J.J. Garcia-Luna-Aceves, “Loop-Free Routing Using Diffusing Computations,” IEEE/ACM Transactions on Networking, vol.1, no.l, February 1993, pp.130-141. [Gerla95] M. Gerla and J.T-C. Tsai, “Multicluster, Mobile, Multimedia Radio Network,” ACM/Baltzer Wireless Networks Journal, vol.1, no.3, pp.255-265 (1995). [Haas97] Z.J. Haas, “A Routing Protocol for the Reconfigurable Wireless Networks,” IEEE ICUPC’97, San Diego, CA, October 12-16, 1997. [Haas98] Z.J. Haas and S. Tabrizi, “On Some Challenges and Design Choices in Ad-Hoc Communications,” submitted for publication. [Haas98-2] Z.J. Haas and M.R. Pearlman, “Providing Ad-Hoc Connectivity with the Reconfigurable Wireless Networks,” submitted for journal publication. [Johnson96] D.B. Johnson and D.A. Maltz, “Dynamic Source Routing in Ad-Hoc Wireless Networking,” in Mobile Computing, T. Imielinski and H. Korth, editors, Kluwer Academic Publishing, 1996. [Lauer88] G. Lauer, "Address Servers in Hierarchical Networks," IEEE International Conference on Communications '88, Philadelphia, PA, 12-15 June 1988. [Leiner87] B.M. Leiner, D.L. Nielson, and FA. Tobagi, “Issues in Packet Radio Network Design,” Proceedings of the IEEE, vol.75, pp.6-20, January 1987. [Moy97] J. Moy, “OSPF Version 2”, RFC 2178, March 1997. [Murthy] S. Murthy and J.J. Garcia-Luna-Aceves, “An Efficient Routing Protocol for Wireless Networks,” MONET, vol.1, no.2, pp.183-197, October 1996. [Murthy95] S. Murthy and J.J. Garcia-Luna-Aceves, “A Routing Protocol for Packet Radio Networks,” Proc. of ACM Mobile Computing and Networking Conference, MOBICOM’95, Nov. 14-15, 1995. [Perkins94] C. E. Perkins and P. Bhagwat, “Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers,” ACM SIGCOMM, vol.24, no.4, Oct. 1994, pp.234-244. [Perkins97] C.E. Perkins “Ad Hoc On-Demand Distance Vector (AODV) Routing,”, IETF MANET Internet Draft, Dec. 1997. [Westcott84] J. Westcott and G. Lauer, “Hierarchical routing for large networks,” IEEE MILCOM’84, Los Angeles, CA, October 21-24, 1984.

212

9 CDMA SYSTEMS MODELLING USING OPNET SOFTWARE TOOL Piotr Gajewski, Jaroslaw Krygier Military University of Technology, Electronics Faculty

.Kaliskiego 2 str, 01-489 Warsaw, Poland phone: (+48 22) 685-9517

fax: (+48 22) 685-9038

E-mail: [email protected]

ABSTRACT The paper presents the result of CDMA network modelling using OPNET software. Some results of our model investigation are also described here. Simulation result can be achieved using elaborated computer programs or commercial software tools. OPNET (Optimized Network Engineering Tool) is an example of such commercial software elaborated and delivered by MIL-3. Elaborated model of CDMA system consists of mobile terminals, base stations, base controllers and

managing centre. This four levels project contains several models including user mobility models (Gaussian, triangular), traffic models (Poisson), channel assignment models (hybrid model with channel relocation), handover models as well as microcellular network model (regular). The user-defined parameters of these models can be introduced and changed. It gives a possibility of investigation of the call blocking probability versus number of channels, channels assignment method, mobile station mobility, priorities, traffic intensity as well as probability of mobile station inaccessibility.

Some results will be presented in this paper, including blocking probability versus total number of channels, number of fixed and dynamic channels and the area of the switching in proposed method.

1. INTRODUCTION

Recent years have shown stunning development of cellular networks and great interest in them among the users all over the world. Bigger and bigger demand is seen in the area of data transfer. Second generation systems enable provision of such services rather below the standard and at low transmission speed. Due to the demand for the so-called personal communication services PCS of worldwide reach, working out a uniform standard becomes essential. At present there is no explicit opinion concerning PCS standard solutions. The application of CDMA code access in radio-interface with spectrum type DS spread technique having the best possibilities of intensive radio-communication maintenance is being considered.

Communication Systems Institute of MUT undertook the realisation of a program application to investigate the characteristics of CDMA system with the use of OPNET software. Computer simulation method is generally used as a confirmation of theoretical consideration or planning an actual system activity. The first stage of system investigation is building a model. The model can be a formal introduction of the theory or a formal description of empirical

213

observation. Most often it is a combination of both. This paper shows projected model of cellular network using DS-CDMA. In the second paragraph mobile station model inside the system is described. The next paragraph describes soft switching model, characteristic for the system using radio-channels coded by pseudo-random sequences. Paragraph four contains the description of proposed allocation channel algorithm and paragraph 5 and 6 show model implementation method on OPNET base and present obtained results of simulation experiment.

2. MOVABILITY MODEL IN CDMA SYSTEM OPNET program makes the way to model the movability of mobile stations easy by drafting a movement motion trajectory of particular nodes with Network Editor. With modelling a good number of mobile station terminals-in-motion, it would cause a great difficulty and the lack of randomness of this type of occurrences. Therefore the model using random generating of motion trajectory was built. The following assumptions will be introduced for this aim: -

calls are generated uniformly inside every cell and call fluxes from separate callers are independent;

-

mobile station can move equiprobably in 4 directions perpendicular to each other; V is random variable meaning velocity of mobile station of normal distribution with mean velocity Vmean and standard deviation Vi is constant velocity of mobile station moving between (i-l) and i-th velocity variation; T is continuous random variable meaning the time between two successive velocity variations or between the time of velocity variation and the end of a call; this variable has exponential distribution with parameter Tmean;. Ti means the duration time between (i-1) and i-th velocity variation of mobile station, with To= 0;

-

the two random variables Vand T are statistically independent;

-

the representation of mobile station trajectory proceeds only during the call;

- T is continuous random variable with exponential distribution meaning full time of a call. Fig. 1 shows the way of movement of mobile station modelled with taking into account random variables mentioned above. The shaded range determines the handover area, in which a mobile station is connected with 3 base stations at the same time.

214

Fig. 2 shows details of trajectory generation. Random velocity of successive segments vi and time between velocity variations, coordinates of velocity variations (xi, yi) can be determined. Assuming constant measure time (tmeasure), path segments di between these measure points will shorten with velocity increase. Mobile station total path depends on randomised total time of call as well.

In places coordinated (xi, yi) a mobile station position is updated.

215

3. SOFT HANDOVER MODEL

A very important problem in cellular networks handing over calls being under way between two base stations at the moment of crossing the intercellular range limit by mobile station. Due to specificity of CDMA system connected with the possibility of using the same frequency band in the whole network region the soft handover method is often in use. It means two or more cells supervising the quality of radio connection in the case of the mobile station moving inside the socalled soft handover range and choosing the most convenient base station through which useful information exchange is carried on.

Fig. 3 shows the soft handover range by shadowed field, rims of which have the possibility to change the site. Physically it is possible due to signal power modification or transmission criteria variations, whereas in our case it is investigation result of switching range effect on the

quality of system service offer. Also fig. 1 shows switching zone limited by intermittent circle line. It can notice that mobile station can be within the range of three base stations. Decision of mobile terminal connection with neighbouring base stations is made at first measure time after soft handover range crossing (Fig. 3). Coming out of switching range is detachable similarly.

4. CHANNEL ALLOCATION ALGORITHM Cellular network systems differ from wire communication systems in limited number of channels. It is caused by limited allocated frequency bandwidth. Thus, they have to conduct a very complex channel allocation policy for calling mobile stations. Because of our modelling

216

code division system we can deal with one allocated bandwidth for the whole system, code divided into channels.

All such channels can be used then in all system cells or be divided into disjoint sets and allocated similarly to narrow frequency-divided bandwidth systems.

In modelled system hybrid policy of radio channels allocation is proposed. It means using two ways of allocation: constant typical for CDMA fixed channels and dynamic, using minimal interference spacing theory for co-channel cells. Channel allocation in cells method is presented in Fig. 4. Dynamic channel allocation principle is based on selecting channel in a given cell by the

way to the minimalisation of cells belonging to this cell interfering environment in which the used channel would be blocked. The so-called cost function is in use here. When the call is finished, the channel allocation is not optimal any more and reallocation is needed on similar

basis.

5. SYSTEM CDMA MODEL APPLICATION ON OPNET PLATFORM

The worked-out simulation model of CDMA system takes advantage of graphic modelling

possibilities of OPNET. Network Editor made it possible to model a system net taking into account accurate position coordinates of every base station. There are net elements like base

stations, base stations controllers, management node. Net nodes in the shape of a generator represent mobile station and wire communication system subscribers. Model implementation in Network Editor is shown in Fig. 5.

One of the targets is testing the influence of movement intensity on each system element. To have the possibility of that variable regulation we assume that each mobile station will not be

217

modelled separately as a single object in a net editor, but all mobile stations will be represented

by the movement generator. Its tasks are as follows: - random generating of fixed number of mobile stations sites at the moment of exchanging the information;

- random generating of velocity and direction of mobile station movement; - the user's behaviour simulation;

- undertaking decisions concerning the quality of received signal; - generation of signalling information.

Modelling of the user's behaviour is based on generating start and finish times of a call or information transmission as well as intensity of calls. It is assumed that call time or data

transmission is represented by exponential distribution. Its mean value is introduced at the beginning of simulation. The change of calls intensity is explicitly defined through the change of

time between successive calls. It is assumed that the time between calls is also represented by

exponential distribution, resulting in Poisson distribution flux of calls. User can call another subscriber of mobile station or wire communication system with the same probability. In case of rejecting the user does not call again, either. 218

Other users of cellular network or wire communication system can call mobile station. In

connection with it, at the given moment the station can be engaged or disconnected. It can be out of reach of the system. Velocity of mobile station and trajectory of movement are modelled in consistence with above mentioned arrangement.

General simulation model structure is displayed in Fig. 6. The model consists of five basic

levels: - opening data setting; - net (configuration, cell sizes, dimension of switching regions); - user's behaviour (generation of calls, mobility, soft handover effect); - call service (allocation and reallocation of channels, transfers, supervising of connections); - collecting simulation effects.

The following parameters can be determined in the input data set: -

mean number of mobile station users for one cell,

-

general amount of duplex channels for speaking,

-

amount of fixed and dynamic channels for one cell,

-

number of channels reserved for switched calls,

-

call duration mean time,

-

mean interval time for successive calls coming from any free subscriber,

probability of outer subscriber’s inaccessibility,

219

-

mean value and standard deviation of mobile velocity,

- cellular network modulus and switching region radius. The definite call generating algorithms, subscribers’ mobility simulation and mobile stations’ behaviour in switching range have been written down in internal language of the file

and concealed in generator node. Call flux from subscribers of wire communication system generating algorithm has been written down and concealed in PSTN node, while complicated algorithms of channel allocation are to be found in management node. Nodes bs00, bs01, ...and bsc0, bsc1, ... are of assistance and serve to transfer signal information among generator and management nodes and to gather and transfer output simulation data. OPNET editor enables representation of particular processes with the aid of transformation graph, where individual processes are described in internal language. Fig. 7 represents

implementation of mobile station generator model.

6.

SIMULATION RESULTS

On the beginning the above mentioned set of input set of input parameters was introduced.

Series of output data was received as a result of carried out simulation experiment. Fig. 8, 9, 10 show probability dependence of call blocking directed at the group of channels in cell from mean traffic intensity in case of different channel distribution in cell, for different region sizes of switching and for different mean values of mobile station velocities.

220

On the basis of the above presented results it can be noticed that:

- using hybrid method of channel allocation significantly decreased call blocking probability in

comparison with the constant method; - decreasing of mean traffic velocity values caused call blocking probability increase; - increasing of soft handover region caused call blocking probability decrease.

221

7. CONCLUSIONS

The authors have suggested and built a model of CDMA system and have implemented it on OPNET simulation tool. They have used graphic possibilities of the tool here and physical

network structure has been modelled. The dynamics of the system has been modelled by programming it in internal OPNET language. Hybrid method of channel allocation has been proposed. A simulation experiment has been conducted and a series of results has been obtained, part of which has been presented in paragraph 6. The obtained results have confirmed an opinion

of dynamic method usefulness of channel policy and proper behaviour of the system under the effect of soft handover region modification and mobile station velocity. The built model can be used to plan cellular network CDMA and to probe the system solutions. It is an open model and

can be developed with other parts of the system. REFERENCES

[1] G. S. Fishman: Computer simulation, notions and methods, (in Polish) PWE, Warsaw, 1981 [2] W. C. Lee: Overview of cellular CDMA, IEEE Trans. on Veh., Vol 40, No. 2, May 1991 [3] MILS, OPNET Manual Set Section.

[4] R. E. Fisher, A. Fakusawa, T. Sato: Wideband CDMA System for Personal Communication

Services, Oki America Inc. 1996IEEE. [5] E. Del, R. Fantacci, G. Gimbene: Handover and Dynamic Channel Allocation Techniques in Mobile Cellular Network, IEEE Trans. On Vehicular Technology, Vol. 44, No. 2, May 1995 [6] D. E. Everitt: Traffic Engineering of the Radio Interface for Cellular Mobile Networks, Proc.

of the IEEE, Vol. 82, No.9, September 1994 [7] M. Amanowicz: Professional Land Radiocommunication Systems Modelling (in Polish),

WAT, Warsaw, 1989.

[8] Szu-Lin Su, Jen-Yeu Cen, Jane-Hwa Huang: Performance Analysis of Soft Handoff in CDMA Cellular Networks, IEEE Journal on Selection Area in Comm., Vol.14, No. 9, December 1996.

222

10 SIGNAL MONITORING SYSTEM FOR FAULT MANAGEMENT IN WIRELESS LOCAL AREA NETWORKS

Jelena F. Vucetic, Paul A. Kline Dynamic Telecommunications, Inc. 12850 Middlebrook Road, Suite 302 Germantown, MD 20874, USA Phone: (301) 515-0403 (xl02) Fax: (301) 515-0037 Email: [email protected]

Abstract: In the last several years, various types of wireless networks have become available and cost-effective for commercial applications. In order to become a real alternative to traditional wireline telecommunications, wireless networks should provide competitive quality, reliability and availability of service. These requirements imply a need of wireless fault management system with features similar to those incorporated in its wireline counterpart. With expansion of advanced high-speed services (e.g. multimedia, ATM, etc.) into wireless networks, fault management becomes an imperative to ensure reliability and quality of service. Despite a significant need for fault management in wireless networks, there has been almost no such proposals nor deployments yet. Existing wireless fault management systems provide management of Mobile Switching Centers (MSC), rarely of base stations. Traditionally, base stations measure signal quality, and send alarms to a Network Management Center (NMC) if some of the signal parameters are out of allowed range. This solution has been proven as insufficient since it is capable to detect the signal quality at a base station's site, not at a user's site. The proposed Signal Monitoring System collects various signal parameters at a user's site (the Access Point of a wireless local area network (WLAN)), determines if these parameters are within the allowed range of values, and generates an alarm if it is not the case. The NMC receives alarms and handles them as in wireline networks (using an automated trouble-ticketing mechanism, and network reconfiguration if it is necessary). This system does not interfere with the regular network operations, nor it requires utilization of any extra voice channels (i.e. no reduction of the network throughput). This paper describes elements of the Signal Monitoring System (Signal Monitoring Units (SMU) and a Performance Manager), and how it detects and locates base station failures, signal degradation, co-channel and adjacent channel interference in a network. A mechanism of alarms generation and reporting is also described, as well as the interface between the Signal Monitoring System and existing Network Management Center.

1. INTRODUCTION This paper describes a fault management solution for wireless local area networks (WLAN) based on an overlay system that continuously evaluates signal quality in the network coverage area, generates and reports alarms to an existing network management system. The proposed fault management system provides features equivalent to the OSI standard [3,5] traditionally applied in wireline networks. It also includes management of operations and maintenance personnel engaged in the WLAN operations [7]. The proposed fault management system enables: • • • • • • • • •

Continuous measurement of various signal parameters in the network coverage area Detection and location of signal degradation Detection and location of co-channel and adjacent-channel interference Dynamic control of base stations' transmit power to adjust the overall network coverage Detection and location of base station failures Efficient trouble-shooting Management of operations and maintenance personnel Various types of statistics and reports Friendly graphical user interface (GUI)

2. SYSTEM OVERVIEW 2.1 SYSTEM ARCHITECTURE Figure 1 represents a general proposed network architecture, which consists of a WLAN, its Network Management Center (NMC), an overlay network of scanning receivers and the Performance Manager. Traditionally, in a wireless network the base stations measure signal quality, and send alarms to the NMC if some of the signal parameters are out of allowed range [6]. This solution has been shown as insufficient since it is capable to detect signal quality only at the base station's site, not at the subscriber's site.

To improve alarms generation in a WLAN, a Signal Monitoring System can be used, as shown in Figure 1. This system does not interfere with the regular network operations. It only "listens" to

224

the signals on the utilized channels, measures their parameters, and sends alarms to the NMC if some of the parameters are out of allowed range. The Signal Monitoring System includes Signal Monitoring Units (SMU) and a Performance Manager. An overlay network of SMUs is used for various types of measurements and statistical analysis of signal parameters. The SMUs can be placed at multiple locations within the WLAN to execute desired measurements. The SMUs are connected with the Performance Manager. They report their measurements to the Performance Manager according to a predefined reporting schedule. The Performance Manager collects measured data from the SMUs, analyzes them and determines if there is a poor or insufficient signal quality in certain area of the WLAN. Such an event may be an indication that the corresponding base station has a problem or a failure. If this is the case, the Performance Manager generates an alarm containing the description of the problem, the location (base station identifier) where the problem is detected, and the timestamp. Then, the

Performance Manager sends the alarm to the NMC, which decides which further actions should be taken (e.g. send trouble tickets, ignore or clear the alarm). There are many ways to implement this system. This paper considers the following high-level versions of implementation: 1. SMUs are connected with the centralized Performance Manager (e.g. via en Ethernet network, as shown in Figure 2). This version is especially suitable for indoors systems (e.g. hospitals, convention centers, factories, etc.). 2. Each SMU is collocated with a distributed Performance Manager (Figure 3), which processes measurements locally and sends alarms related to the corresponding base stations that it is covering to the NMC. The connection between Performance Managers and the NMC can be TCP/IP, dial-up or any other on-demand connection.

225

2.2 SYSTEM DESIGN CONSIDERATIONS Location and number of SMUs in a WLAN is critical to provide reliable fault management. However, in the design of the Signal Monitoring System, a tradeoff needs to be made between its reliability and cost. To minimize the cost of the system, the number of SMUs should be minimal. They should be placed at locations where signal degradation and interference from another base stations are beyond a tolerance threshold.

2.3 NETWORK ELEMENTS Managed network elements of the proposed Signal Monitoring System are Base Station

Transceivers (BST) and Base Station Controllers (BSC). For the purpose of fault management, BSTs are characterized by the following attributes:

• • • • • • • • • • •

BST identifier Associated BSC identifier Type of antenna (omni-directional or sectorized, non-diversity or diversity) Number of voice channels List of voice channels Number of control channels List of control channels Current status of channels (available, busy, faulty) BST status (operational, faulty, not-configured) List of backup BSTs Collected alarms

• •

BST location BST area

226

• • • •

BST region Responsible Supervisor Responsible Operator Responsible Technician(s)

For the purpose of fault management, BSCs are characterized by the following attributes: • • • • • • • • • • •

BSC identifier Associated BST identifiers Number of voice channels List of voice channels Number of control channels List of control channels Current status of channels (available, busy, faulty) BSC status (operational, faulty, not-configured) List of backup BSCs Collected alarms BSC location

• • • • •

BSC area BSC region Responsible Supervisor Responsible Operator Responsible Technician(s)

3. FAULT MANAGEMENT SYSTEM The proposed fault management system for WLAN includes signal measurements performed by SMUs, alarms generation and reporting performed by the Performance Manager, and alarms handling performed by the NMC.

3.1 THE PERFORMANCE MANAGER Depending on the organization of the overall fault management system, each base station in the monitored network can have its own Performance Manager (as shown in Figure 3), or a single Performance Manager can monitor several base stations (as shown in Figure 2). The former solution is recommended if the associated SMU is within the coverage area of a single base station. The latter solution is more suitable if the SMU is located within the area with overlapped coverage of more than one base station (distributed Performance Manager, as in Figure 3), or in the case of a centralized Performance Manager (Figure 2). In any case, the Performance Manager consists of the following functional blocks, as is shown in Figure 4: • • • •

SMU Interface Measurement Database Alarms generator NMC Interface

227

3.1.1 The SMU Interface In transmit mode, the SMU Interface downloads lists of scanning channel identifiers and types of desired measurements to the associated SMU(s). The list of channel identifiers is created by a network operator using the corresponding base station's frequency plan. The operator can enter the list either remotely from the NMC or locally from a database residing in the considered Performance Manager. For each base station monitored by the Performance Manager, the list of downloaded channels into the corresponding SMU(s) includes the channels allocated to the base station, as well as the list of their adjacent (higher and lower) channels. In the receive mode, the SMU Interface collects measurement data from SMUs according to the previously downloaded list of channel identifiers and types of measurements. The received data are then stored into the Measurement Database, whose organization (for a single base station) is shown in Figure 4.

3.1.2 The Measurement Database

228

The Measurement Database is a relational database residing in a Performance Manager. It consists of tables that define: which BSTs are monitored by the Performance Manager (BST_Table), which SMUs report measurements to the Performance Manager (SMU_Table), which SMUs measure signal characteristics for each base station (SMUList), and measured data for each channel allocated with each BST (Meas_Table).

The Measurement Database organization is shown in Figure 4. In this example, the Performance Manager monitors N BSTs using M SMUs. Each BST is associated with a set of SMUs that measure signal characteristics of this BST. This association is defined in SMUList tables.

For each controlled SMU and each BST monitored by the Performance Manager, there is a Meas_Data data block. It consists of a Channels Table and Measurement Table (RSSI, CoChInt, AdjChHi and AdjChLo columns). A detailed organization of a single Meas_Table (for a single BST and a single SMU) is presented in Figure 5. The number of entries of each of these columns is equal to the number of channels allocated to the corresponding BST. The RSSI column contains the average RSSI measured for each channel from the Channels Table. In the same fashion, the CoChInt column contains co-channel interference measurements, while AdjChHi column and AdjChLo column contain adjacent channel interference measurements for higher and lower adjacent channel, respectively.

3.1.3 The Alarms Generator The Alarms Generator periodically reads measurements from the Measurement Database for each controlled SMU and each monitored base station. Then, it evaluates if any of interference measurements exceeds predefined thresholds, or if any RSSI measurement is below a specified threshold. Then, the Alarms Generator correlates measurements obtained from the SMUs measuring signals of the same BST, as well as of adjacent BSTs. If correlated results indicate poor coverage in a certain area, the Alarms Generator generates an alarm and sends it to the NMC via the NMC Interface. These alarms contain the following information: • • • • •

BSC Identifier BST Identifier Faulty Channel Identifier Alarm Code Alarm Description

•

Out-of-Range Measured Data

• • •

SMU Identifier SMU Location Timestamp

229

•

Alarm Severity

3.1.4 The NMC Interface The NMC Interface enables communications between the Performance Manager and the NMC using any standard protocol (e.g. as is shown in Figures 2 and 3). In transmit mode, the NMC Interface sends alarms from the Alarm Generator to the NMC. In receive mode, the NMC Interface receives lists of channel identifiers for each base station monitored by the Performance Manager, and forwards the list (along with the list of adjacent channels) via the SMU Interface to the corresponding SMU. At the same time, the NMC Interface stores these lists in the Measurement Database.

3.2 THE NETWORK MANAGEMENT CENTER To be able to utilize the fault management capabilities of the described Signal Monitoring System, the existing Network Management Center (NMC) includes the following functional blocks: •

Performance Manager Interface

•

Management Information Base (MIB)

•

Fault Management Applications

•

Graphical User Interface (GUI)

If any of these blocks already exists in the NMC, they should be expanded to support fault management.

3.2.1 The Performance Manager Interface The Performance Manager Interface exchanges information between the NMC and the Performance Manager. In transmit mode, the Performance Manager Interface downloads into the Performance Manager the list of channels that are allocated to each of the BSTs monitored by this Performance Manager. In receive mode, the Performance Manager Interface receives alarms from the Performance Manager and stores them into the MIB.

3.2.2 The Management Information Base (MIB) The Management Information Base (MIB) is a relational database which stores all information on the managed network elements (NE), operations and maintenance personnel, that are relevant to the fault management applications.

3.2.2.1 The Fault Management Information Structure The fault management information [4] can be divided into three categories: • Information on regions of operations • Information on network elements •

Information on operations and maintenance personnel

230

Each information category is hierarchically organized and connected by relations. Different

categories of information are also interconnected through relations, as shown in Figure 6. In terms of regions of operations, a network may be divided into regions. Each region consists of several areas. In terms of managed network elements, a network includes BSCs and BSTs. Each region contains one or more BSCs and BSTs connected to them. Each area contains either several BSCs or several BSTs (in Figure 4, denoted by shaded blocks).

A network is operated and maintained by the personnel that includes a Superuser, Supervisors, Operators and Technicians. Each Operator is assigned to an area to control the corresponding set

of BSCs or BSTs. A team of Technicians is assigned to each area to install, maintain, repair and upgrade the corresponding BSCs or BSTs. Each Technician is responsible for a specified type and subset of NEs within a certain geographic area. Each region is managed by a Supervisor. Each Supervisor has a group of Operators and Technicians reporting to him/her. All Supervisors report to the Superuser, who is responsible for the whole network.

3.2.2.2 The MIB Organization The fault management information is implemented in a MIB as a relational database [2,4,5,7]. The general MIB organization is shown in Figure 7. The MIB consists of tables interconnected by relations. Each table consists of attributes. For simplicity, Figure 7 highlights only the attributes that illustrate how the tables are related in terms of fault management. The NE table contains all relevant network element (BSC or BST) information (the NE identifier, the affiliated area identifier, the responsible technician’s identifier, the responsible operator's identifier, the NE configuration, location, list and status of generated alarms, number of channels, etc.). The Area table contains all relevant information on an operation area (the area identifier, the affiliated region identifier, the responsible operator identifier, the area’s office address, telephone number and facsimile numbers).

231

The Region table contains all relevant information on an operation region (the region identifier, the responsible supervisor identifier, the region monitoring center address, telephone and facsimile numbers). The Supervisor table contains all relevant information on a supervisor who is responsible for a certain network region (the supervisor identifier, the affiliated region identifier, the name, address, telephone number, facsimile number, pager, working hours). The Operator table contains all relevant information on an operator who is responsible for a certain network area (the operator identifier, the affiliated area identifier, the supervisor's identifier, the name, address, telephone number, facsimile number, pager, working hours). The Technician table contains all relevant information on a technician who maintains a certain network element (the supervisor identifier, the affiliated area identifier, the name, address, telephone number, facsimile number, pager, working hours, network element identifier).

Figure 8 illustrates a top-down MIB structure. The Superuser table contains pointers to the lists of three types of MIB entities: managed network elements, personnel and managed areas. Assuming that the MIB and its tables have been created, the following operations can be executed on MIB entities: • Add new entity • Modify entity information • Read entity information • Delete entity These operations also include creation, modification or deleting of relations of the relevant entities.

3.2.3 The Fault Management Applications Fault management applications include:

232

The NMC filters alarms received from all Performance Managers based on various criteria, such as: geographic area, available operations personnel, alarms severity.

Based on the nature of a received alarm, the NMC can issue a trouble ticket with the directions to a technician to resolve the problem, clear the alarm, or escalate its priority to expedite its resolution. In some cases, the NMC can reconfigure the network in order to provide necessary coverage in the presence of a base station failure. This can be done by increasing the transmit power of a certain BST to provide service to an area affected by the failure of another BST. If interferencebased alarms are persistent, the NMC can modify the entire frequency plan of the managed network and/or the transmit power of its base stations to improve the overall coverage.

3.2.4 The Graphical User Interface (GUI) We assume that the NMC already includes a GUI containing a geographic map of the network coverage area, with an overlay graphical presentation of the managed network elements and their interconnections.

233

The existing GUI can be expanded with the following features to support fault management:

•

Changing the color of a network element symbol based on the severity of a relevant alarm

•

Blinking the network element symbol in emergency situations

•

Clicking on the network element symbol to obtain real-time information on the element's configuration and pending alarms.

•

Clicking on a selected alarm to automatically generate and send a trouble ticket to a field technician. The ticket includes: technician identifier, alarm description, faulty network element identifier and location. The NMC can send the ticket to a technician by: email, facsimile, pager, short messaging, etc.

•

Automatically clearing an alarm and modifying the status/color of the corresponding network element when the problem/failure is resolved.

4. CONCLUSION This paper describes an overlay Signal Monitoring System, which provides fault management of base stations in a WLAN. The system continuously measures various signal parameters using a set of SMUs deployed within the network coverage area. Based on these measurements, the system determines signal degradation or interference, and generates alarms that indicate either a failure in certain network element or inadequate coverage.

The proposed Signal Monitoring System improves reliability and quality of service of a WLAN by efficient failure detection and location. If it is applied in dynamic adjustment of the network frequency plan and transmit power control of the base stations, it can also improve the overall network coverage and availability of network resources.

REFERENCES [1]

[2]

[3] [4]

[5] [6] [7]

S. M. Dauber: “Finding Fault”, BYTE Magazine, McGraw-Hill, Inc. New York, NY, March 1991 O. Wolfson, S. Sengupta, Y. Yemini: “Managing Communication Networks by Monitoring Databases”, IEEE Transactions on Software Engineeringm Vol. 17, No. 9, September 1991 L. Feldkhun: “Integrated Network Management Systems”, Proceedings First International Symposium on Integrated Network Management, 1989 H. Yamaguchi, S. Isobe, T. Yamaki, Y. Yamanaka: “Network Information Modeling for Network Management”, IEEE Network Operations and Management Symposium, 1992 S. Bapat: “OSI Management Information Base Implementation”, Proceedings Second International Symposium on Integrated Network Management, 1991 J. Vucetic, P. Kline, J. Plaschke: “Implementation and Performance Analysis of MultiAlgorithm Dynamic Channel Allocation in a Wideband Cellular Network”, Proceedings of the IEEE ICC ’96 Conference, 1996 J. Vucetic, P. Kline: “Integrated Network Management for Rural Networks with Fixed Wireless Access”, Proceedings of the Wireless ’97, the 9th International Conference on Wireless Communications, 1997

234

11 COMPUTER-AIDED DESIGNING OF LAND MOBILE RADIO COMMUNICATION SYSTEMS , TAKING INTO CONSIDERATION INTERFERING STATIONS. Marek Amanowicz, Piotr Gajewski,

, Marian Wnuk

Military University of Technology, Electronics Faculty

.Kaliskiego 2 str, 01-489 Warsaw, Poland phone: (+48 22) 685-9228

fax: (+48 22) 685-9038

E-mail: [email protected].

ABSTRACT The electromagnetic environment is made up of signals emitted by mobile and stationary radio stations

in radio communications networks, transmitting stations in the radiolinie system as well as radar, navigation, jamming and other systems to be found in a given area. Formulas for coverage and minimum distance analysis, taking into account mutual or intentional interference, are presented.

1. INTRODUCTION The evaluation of the power of the received signal requires the knowledge of the models of the elements

in the system, i.e.: the transmitter, transmitting antenna, the propagation environment of electromagnetic waves, the receiving antenna and the receiver. The power of the signal (useful or interfering) at the receiver input is calculated with the use of the probability theory methods. The average value exceeding the admissible threshold signal value

standard deviation

and the probability

of

for example the minimal detectable signal

are defined. When

is known, it is possible to define the average value of propagation losses, which hinges on

the distance R between the devices. In the coverage analysis it is an operational range and in the interference analysis - an admissible distance. It is possible to calculate the range between devices with an eye to the terrain profile, which brings us close to the real propagation model.

2. SIGNAL POWER AT RECEIVER INPUT The average value of signal power

The minimal detectable signal

at the demodulator input, expressed in decibels, is:

is the value of the minimal signal power

(demodulator) input which meets the signal power requirements at the receiver output:

235

at the receiver

When a radiocommunication network comprising „n” transmitters is analysed,

is expressed by

the following dependence:

where:

I - interference energy sum

Ps - noise power of the receiving system at the IF amplifier output [W] at the receiver (demodulator) input

3. DISTANCE BETWEEN DEVICES FOR GIVEN PROBABILITY OF SIGNAL DETECTION The distance between devices for a given probability of signal detection is calculated to define

either the operational range or the admissible (minimal) distance between devices above which EMC conditions are preserved. In both cases, appropriate threshold values are assumed for signal power

and probability

of exceeding the threshold by the useful signal

as well as by the

disturbing signal

When the operational range is calculated, the probability for the useful signal

case of the admissible distance, the probability.

of the interfering signal

and in

exceeding the

assumed threshold is within the range , depending on the useful signal requirements:

The threshold value of signal

is expressed by formula (2), which applies both to useful signals

S in the analysis of operational range and interference I when the admissible distance is calculated. Appropriate values of the signal to noise ratio (S/N), interference to noise ratio (I/N) and signal to interference ratio (S/I) at the receiver (demodulator) input are calculated according to ITU-R and ITU-T

recommendations of the International Telecommunications Union (ITU) [5]. Though the admissible distance between devices is usually defined for out band transmission of the

interfering signal, cases of in band transmission should also be considered.

236

When

is known, the average value of propagation loss is calculated according to the following

formula:

Corresponding with this value is a certain distance probability

between the considered radio stations, at which the

that the received signal power exceeds the value

To fulfil condition (5) the

equals 0.5.

value should be selected in relation to the

value to meet the

following condition:

where:

X - normal distribution standardised variable of signal power corresponding to assumed probability

- signal power standard deviation

It is possible to fulfil condition (5) by correcting the propagation loss value

calculated

according to formula (6). In effect the appropriate average value of propagation loss

is expressed as follows:

The distance R between the devices, which corresponds with the average value of propagation loss is calculated using to make modifications the EPM-73 method, which takes into consideration the terrain profile, for the following parameters:

calculated according to formula (10) •

- antenna’s height above the surrounding terrain,

• f-frequency band

• H or V antenna polarisation, • average terrain height,

•

height of terrain below or above the average level (medians) for a transmitting and a receiving antenna.

Considering

the height of the antenna above the surrounding terrain (fig. 1) is:

237

Considering the effective height of the antenna for vertical and horizontal polarisation,

the height of the antenna is:

Direct transmission

- suspension height of the transmitting antenna - suspension height of the receiving antenna - average height of terrain surrounding the transmitting antenna within a 3-15 km distance - average height of terrain surrounding the receiving antenna within a 3-15 km distance

- terrain elevation near the transmitting antenna - terrain elevation near the receiving antenna Fig. 1. Terrain profile on a route between stationary radio station B and mobile station M as well as effective heights of antennas (average terrain cross-section for the given azimuth is shown). The values calculated from the dependencies (13,14) are taken into account while calculating propagation losses. It has to be remembered, however, that certain limits are imposed on values

determined by the dependence (11), which affect the value of propagation losses.

238

4. COMPUTER CALCULATIONS This software makes a visual presentation of the results of calculations possible. The necessary parameters are defined by the user. The screen is divided into three sections (Fig. 2):

. top menu bar

. bottom status bar . digital map with contour lines and coverage distribution

The status bar provides information about the map and the point of cursor of the mouse position: geographical and topographical coordinates, vegetation growth as well as the terrain height - fig.2.

The menu with dialogue boxes gives access to such commands as:

. change the parameters of the map (scale, centre screen situation, enlarge a section of the map), . present the terrain profile (display the profile, open a file with the profile shown before), . fix a network (create a new network, open an already existing network file, save the current network and

close), . fix data concerning the facilities (display the facility, create a new one, copy, delete one or all facilities), . select algorithm (calculations for a non-directional or a directional antenna).

239

The simplest operations involving the map are changing its scale, enlarging its section or refocusing the screen. Another useful function is a readout of the terrain profile along a given route (fig.4).

The situation on the map can be either fed by the user or a file with a network saved can be opened. The network fed can be modified, and the changes introduced saved.

240

Various kinds of broadcasting stations can be featured on the map. Each has a name and a set of parameters providing information about the transmitter power, the frequency on which it works, the height at which the antenna is suspended and many other data important for the algorithm (fig.5). Some data have an assumptive value, so the calculations can be made even if not all the values are given, but it has to be remembered that they cannot be left out as this would affect the final result. Since stations may have the

same parameters, it is possible to copy a station with a list of its parameters but without its name.

241

When the station’s parameters are known, the algorithm of calculations can be selected - for a nondirectional or a directional antenna. The next step is to select the type of calculation, which may concern: . the range,

. signal strength, . probability,

. power density, . field intensity.

There is also an option either to make the calculations considering the terrain profile or selection of an ,,average” profile (fig.6). The parameters option makes it possible to edit the parameters belonging to the

algorithm. In this way we can decide which probability levels or power levels the calculations will be made for.

When the calculations have been completed, isolines are displayed on the screen, their alignment depending on the levels on which the calculations have been made. Figs. 7,8,9,10 show the results of a series of calculations for various algorithms.

242

243

This software makes it possible to determine the quality of the useful signal against the background of interference. One of the quality measures in radio and telephone systems is the articulation (AS) or the

articulation index (AI) for voice transmission. The threshold signal is relation between of the ratio the useful signal and interference.

at which correct reception is possible. In case of digital systems it is

the bit error rate BER. To receive an undisturbed signal, the following inequality must be fulfilled:

Knowing the technical parameters of the facility and the interference level

we can calculate

propagation losses and the distance at which the received signal is jammed. If the interference comes from n sources, the resultant value of interference.

is

where:. disturbance from i-source at [W] n- number of interference sources

244

Fig. 11. shows a situation, where the receiving station with a round characteristic is within the range of two broadcasting stations, both with directional antennas, one emitting a useful and the other an interfering signal. When the useful signal and interference reach a given level, the receiving station will be jammed.

5. CONCLUSIONS Out of the many measures of signal quality in analogy and digital radio and other communication systems,

the most useful are:

• articulation ... or index for telephone radiocommunication systems, • bit error rate or code-word distortion degree - for digital transmission,

• resolution - for TV and facsimile,

• probability of detection and of false alarm - for radar systems,

• location error - for navigation systems. The software presented here makes it possible to define a probable power distribution around the

broadcasting station considered or the probability distribution for a given power. It is also possible to determine the strength of the electric field, power density and propagation losses. The useful range and the admissible distance from the source of interference can be calculated as well. In all those calculations the

signal quality measures listed above are taken into consideration.

245

6. REFERENCE: [l]Duff W.G., White D.R.J.: EMI Prediction and Analysis Techniques. A Handbook Series on Electromagnetic Interference and Compatibi-lity, vol. 5, Don White Consultans, germantown, Maryland, 1971-1974. [2]Lustgarten M.N., Madison J.A.: An Empirical Propagation Model (EPM-73), IEEE Trans on EMC, vol EMC -19, No 3, August 1977. [3]M. Amanowicz, P. Gajewski, W. Ko osowski,, M. Wnuk. ,,Land mobile communication systems engineering” - 4th Conference, AFRICON 1996 [4]Freeman R.L.: Telecommunication Transmission Handbook, John Willey & Sons, New York 1981. [5] International Telecommunication Union : Radio Regulations, edition 1990, Geneva. [6] M. Amanowicz, P. Gajewski, W. Ko osowski, M. Wnuk ,,Analiza komputerowa i do wiadczalna stref pokrycia sieci radiokomunikacji 1 dowej” KST97 Bydgoszcz wrzesie .

246

12 ADAPTIVE INTERFERENCE CANCELLATIONWITH NEURAL NETWORKS A.H.EL Zooghby, C. G. Christodoulou, and M. Georgiopoulos Electrical and Computer Engineering Department

University of Central Florida 4000 Central Florida Blvd, Orlando, Florida 32816

[email protected]

Abstract In modern cellular, satellite mobile communications systems, and in GPS systems, a fast tracking system is needed to constantly locate the users, and then adapt the radiation pattern of the antenna to direct multiple narrow beams to desired users and nulls to

sources of interference. In this paper, the computation of the optimum weight vector of the array is approached as a mapping problem which can be modeled using a suitable

artificial neural network trained with input output pairs, A three-layer radial basis function neural network (RBFNN) is used in the design of one and two-dimensional array antennas to perform beamforming and nulling. RBFNN's are used due to their ability to interpolate data in higher dimensions. Simulations results performed under different

scenarios of angular separations, and SNR are in excellent agreement with the Wiener solution. It was found that networks implementing these functions are successful in

tracking mobile users as they move across the antenna’s field of view.

1. Introduction

Multiple access techniques are often used to maximize the number of users a wireless communications system can accommodate. With frequency reuse, where the same frequency is

used in two different cells separated far enough so that users in one cell do not interfere with the users in the other cell. Further improvements in the system capacity can be achieved. Moreover,

adaptive arrays implemented in base stations allow for closer proximity of cofrequency cells or

beams providing additional frequency reuse by rejecting or minimizing cochannel and adjacent channel interference[l]. Motivated by the inherent advantages of neural networks([2]-[7]), this paper presents the development of a radial basis function neural network-based algorithm to compute the weights of an adaptive array antenna [8]. In this new approach, the adaptive array can detect and locate mobile users, track these mobiles as they move within or between cells, and allocate narrow beams in the directions of the desired users while simultaneously nulling unwanted sources of interference. This paper is organized as follows: In sections 2 an 3 a brief derivation of the optimum array weights in 1-D and 2-D adaptive beamforming is presented. In Section 4 the RBFNN approach for the computation of the adaptive array weights is introduced. Finally, Section 4 presents the simulation results and Section 6 offers some conclusive remarks.

2. Adaptive beamforming using 1-Dimensional linear arrays

Consider a linear array composed of M elements. Let K (K