A self-adaptive universal receiver

pp. 421-456

421

A self-adaptive universal receiver Christian ROLAND*,

Jacques PALICOT**

Abstract

In this paper we describe a new Self-Adaptive Re-Configurable Terminal, based on blind recognition o f the system in use. This new terminal is operates thanks to a new radio interface receiver concept comprising two functional phases. These two phases are mapped on a specific architecture previously presented in [1]. The first one is devoted to a Wide Band Analysis (WBA) o f the received signal in order to f i n d which standard is being used. The second one being the demodulation itself. In this paper we focus on this WBA. It includes an iterative process in order to find the right bandwidth which has to be compatible with the Signal Processing techniques we use. During the last step o f the process, blind standard recognition is performed by means o f Radial Basis Function Neural Networks, which allow making full use of the analogy between our problem and conventional pattern recognition problems. Extensive simulation with true data of signals received in our lab has been performed and confirms the interest and efficiency o f this type o f recognition. Key words: Radiocommunication, Mobile station, Adaptive system, Reconfigurablecircuit, Software radio, System architecture, Spectral analysis, Power spectrum, Neural network, Error function, GSM, Automatic recognition, Patternrecognition.

UN RI~CEPTEUR UNIVERSEL AUTO-ADAPTATIF R~sum~

Un terminal reconfigurable de maniOre auto-adaptative est dgcrit dans cet article. Ce terminal fonctionne grace gt un nouveau concept d'interface radio comprenant deux phases fonctionnelles distinctes. Cette interface s'appuie elle-mgme sur une architecture prdsentde prdcddemment dans [1]. La premiOre phase fonctionnelle correspond gt une analyse d large bande du signal R F res La seconde gtant la ddmodulation elle-m~me. Nous nous intgressons dans cet article d la premiOre phase fonctionnelle d'analyse gt large bande (ALB). Cette analyse comprend, en premier lieu, une adaptation itgrative de la largeur de bande du signal RF compatible avec le pouvoir de rgsolution des techniques d'analyse spectrale utilisges. Cette analyse se poursuit par la reconnaissance autodidacte des normes des signaux regus au moyen de rdseaux neuronaux utilisant des fonctions de base radiale, qui permet-

* Universit6 de Rennes 1, IUT GEII ; 3, rue du Clos Courtel BP90422, 35704 RENNES ch.roland @ wanadoo.fr. ** IRISA Campus Universitaire de Beaulieu 35042 Rennes [email protected] 1/36

Cedex 7,

ANN. TI~LI~COMMUN.,57, n ~ 5-6, 2002

422

c. ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

tent de profiter au maximum de l'analogie entre notre probl&me et les probldmes de reconnaissance de formes. De nombreuses simulations avec des signaux rdels refus au laboratoire confirment l'intdrdt et la performance de ce type de reconnaissance. Mots cl~s : Radiocommunication, Station mobile, Syst~me adaptatif, Circuit reconfigurable, Radio logicielle, Architecture syst~me, Analyse spectrale, Spectre puissance, R6seau neuronal, Fonction erreur, GSM,Reconnaissance automatique, Reconnaissance forme.

Contents I. Introduction II. Our proposal for the Self-Adaptive Universal Receiver (SAUR) III. Blind recognition

IV. Results V. Conclusion References (26 ref )

I. I N T R O D U C T I O N

1.1. The Universal Terminal Wireless c o m m u n i c a t i o n has regained considerable interest o v e r the last few years and n o w a d a y s is one o f the fastest growing sectors o f the t e l e c o m m u n i c a t i o n industry. The step forward not only expands the market for wireless c o m m u n i c a t i o n s but also creates opportunities for newer products. The number o f services on heterogeneous wireless n e t w o r k s (GSM 3, IS954, PDC 5, DECT6, PHS 7 and future 3G standards like the UMTS8 proposal in Europe) is dramatically increasing. In European countries there is an explosion in the n u m b e r o f c o n s u m e r s for mobile c o m m u n i cations. M o r e o v e r RLAN 9 and Hiperlan 1~ also contribute to an increase in the number of wireless services. In the broadcasting area, DAB11 and DVB-T12 are additional services. A m o n g all the k e y - w o r d s b e h i n d these services the " m o b i l i t y " k e y w o r d is one o f the most important as presented in Figure 1. The main d r a w b a c k of this e x p l o s i o n o f services is that consumers need more and more terminals. Furthermore, the down-compatibility with pre-existing 2G and 2.5G networks should be insured. In addition, the interconnection ( " c o n v e r g e n c e " concept, m a y be the second k e y word) among all the networks increases the c o m p l e x i t y o f the u s e r ' s environment. Consequently there is a current and growing interest in universal terminals (multi-services, multi-networks) (see Figure 2). The technical a p p r o a c h for these universal terminals consists in d e v e l o p i n g re-configurable terminals, w h i c h c o u l d b e performed with software radio (SWR) techniques.

3. GSM: Global System for Mobile Communications 4. IS95: Interim Standard 95 (CDMA) 5. PDC: Personal Digital Cellular 6. DECT: Digital Enhanced Cordless Telecommunications 7. PHS: Personal Handy phone System 8. UMTS: Universal Mobile Telecommunication system 9. RLAN: Radio Local Area Network 10. Hiperlan: High Performance Radio Local Area Network l l. DAB: Digital Audio Broadcasting 12. DVB-T: Digital Video Broadcasting-Terrestrial ANN. TI~LI~COMMUN.,57, n ~ 5-6, 2002

2/36

423

c. ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

FIG. 1. - - W h y a U n i v e r s a l T e r m i n a l ?

Pourquoi un terminal universel ?

FIG. 2.

--

Universal Receiver Interfaces.

Interfagage du rgcepteur universel. 3/36

ANN. TI~LI~COMMUN.,57, n ~ 5-6, 2002

424

r

R O L A N D -- A S E L F - A D A P T I V E U N I V E R S A L R E C E I V E R

The idea of a re-configurable receiver first appeared in the military area [2]. In fact, the need for re-configuration appeared very soon in the 70's, with for example the first equipment called "SPEAKeasy". This concept became popular in the civil telecommunications area in the 90's, thanks mainly to the work of J. Mitola [3]. This induced the first magazine special issue in 95 [4]. Two very recent events prove the telecommunications community's interest in SWR techniques. The first is the creation by the major European telecommunications industrials of the Wireless Word Research Forum [5] which has defined SWR activities. The second one is the creation, in the free software world, of a new project: the GNU Radio [6]. All previous examples prove that there is now a growing interest in re-configurable receivers. This sector, whatever the technique use is to offer this re-configurability, is one of the fastest growing sectors in the telecommunications industry. It is in this context that we started this study.

1.2. Why a Self-Adaptive Terminal? As explained in the previous section, it seems clear that universal (re-configurable) terminals will be the solution in a near future, and also that these terminals will provide more and more "intelligence" to be user-, network- independent. In fact, we can imagine that the user's environment in a broad sense should be taken into account by the networks and obviously by the terminal. In this context, we propose a new vision for an universal terminal. It will be a self-adaptive terminal, in the sense that it will recognize the transmission system in use in a "blind" manner and consequently will re-configure all its architecture with the adequate software. This reconfigurable receiver will become very common in a near future, but in addition this reconfiguration will be performed in a blind manner, which is the new idea. This proposal is one possible solution for the physical layer evolution towards "cognitive" radio. As said by J. Mitola in [7], "this type of learning technique makes the Software Radio trainable in a broad sense instead of just re-configurable". This paper comprises three main parts. In the first one we describe our proposal : the SelfAdaptive Universal Receiver (SAUR). This part is a functional description (section II. 1), based on a new radio interface that includes two functional phases, the first one being called the "Wide Band Analysis" (WBA).The second section (II.2) presents the possible architectures on which this new radio interface could be mapped. In this section we present our two path architecture which could easily offer the new radio interface. This architecture could be carried out with today's technologies and will also be compatible with the future generic SWR architecture when the new Analog to Digital Converter (ADO generation is available. The following section (II.3) fully describes the Wide-Band Analysis (WBA), which is performed using powerful Signal Processing Algorithms. Actually it is the "heart" of our proposal. This description deals first (sub-section II.3.1) with the necessary adaptation of the width of the analyzed band, performed by means of an iterative process, and then (sub-section II.3.2) with the standard parameter study by analyzing the parameter discrimination power of almost all the existing systems and their accessibility. Next we will discuss the blind recognition itself using the previously selected parameter. This last print is precisely the object of the second part of this paper. We describe our blind recognition proposal based on Radial Basis Function Neural NetANN. Ts

57, n~ 5-6, 2002

4/36

425

C. ROLAND -- A S E L F - A D A P T I V E UNIVERSAL RECEIVER

works (RBF NN) recognition, carried out during the last iteration of the adaptation procedure, in that chapter we will explain, how the RBF NN is carried out. In particular, the pre-treatment function as well as the synchronization procedure are detailed. Finally, in the third part of this paper, we will give some results in terms of Good Recognition Rate (GRR), False Recognition Rate (FRR), both for "theoretical results" (section IV. 1) and for "live results" (section IV.2).

II. O U R P R O P O S A L FOR THE SELF- ADAPTIVE U N I V E R S A L R E C E I V E R (SAUR)

II.1. Functional description Figure 3 presents the new radio interface concept.

FIG. 3. -- Our proposal for a Self Adaptive Universal Receiver. Notre proposition de Terminal Universel Auto-adaptatif

The solution proposed herein has two functional phases. This solution is fully described in [8]. We have called the first one "Wide Band Analysis" because it consists in the study of the complete multistandard received signal. It is represented by equation (4) and illustrated in Figure 4. This signal is the result of the summation of many standards, each standard being itself the summation of several modulated carriers (channels). The analysis of this signal should give the important information that is useful for the receiver. It could, for example, be the modulated carrier existence, the type and the position (in the frequency domain) of the standard(s) in use in the analyzed band. In some particular cases (not detailed in this article), 5/36

ANN. TI~LI~COMMUN.,

57, n~ 5-6, 2002

C. R O L A N D -- A S E L F - A D A P T I V E

426

UNIVERSAL RECEIVER

this information could also be the symbol frequency value, the mapping of the modulation, etc. After having recognized the standard, the receiver will check if the right demodulation software is available in the local database9 If so, the second phase could use the software directly and start. Otherwise, there are many possibilities such as a download of it through wire connection (service provider on the Internet) or directly Over The Air. This download problem, which is of course of great importance for SWR is not addressed in this paper. The reader may find it useful to refer to article [9]. Another function at that level in the receiver is the management of the user's rights (i.e. the operator rental, the user's requirements given by parameters such as price/bit, spectrum efficiency, bit rate, overall QoS). This management problem also is not addressed in this paper. // 9 ............ i

Cod ng ~ ~ /

[ Coding ~

o

L

-.,

~

...............................

7

!

Modulation ~ Modulation ~

- -

.............

~

FDMA ~

I

~

........... : .................~ . . . . i ,,, 1' Coding i ~ : ..................................... Moduaton ~ ~[

~ /

I Coding ~

Modulation ~ ~

Coding ~

Modulation ~

~

F

.............

q

Channel i

Channel ~ / / \ , _ ~: Universal ~ - T - r - - - ~ Receiver ................

............................. Channe

J,/

',

TDMA ~

N

/

Channei ]

/

/

C~ C

FIG. 4. -- Composition of the received signal. Composition du signal refu.

Whatever the technique used for getting the right software, the second phase shall start just after the standard has been recognized and selected. Then the receiver will try to communicate. This second functional phase is so obvious that there is no need to describe it any further.

II.2.

II.2.1.

Possible

architectures

Generic

SWR a r c h i t e c t u r e

for

this

SAUR

As presented in Figure 5, this ideal architecture is the best for offering the SAUR. Concerning the WBA phase, this architecture is well adapted. In fact, under the assumption that we have sufficient time to perform powerful signal processing techniques, it is totally clear that with existing ADC,even with a low number of bits, we will reach our objective9 The only restriction is when a complete standard is under the noise quantization. ANN. TI~LI~COMMUN.,57, n~ 5-6, 2002

6/36

427

C. ROLAND -- A S E L F - A D A P T I V E U N I V E R S A L RECEIVER

FIG. 5. - - Ideal

Software Radio Architecture.

Architecture ggndrique de la radio logicielle iddale.

Unfortunately, it is not the same "story" for the demodulation phase. As is well-known, and presented in the F i g u r e 6, this architecture could not be i m p l e m e n t e d with t o d a y ' s technology. In fact, for the d e m o d u l a t i o n phase, the ADC r e q u i r e m e n t s are the s a m e as the conventional SWR p r o b l e m and could not be fulfilled currently. That was the reason for our two-path architecture p r o p o s a l , which is described below.

FIG. 6. - - ADC requirements for ideal

SWR.

Performances du CANpour une radio logicielle idgale.

II.2.2. Two-path architecture

Two years ago w e p r o p o s e d [1] an architecture that was w e l l - a d a p t e d for this new radio interface. That new r e c e i v e r architecture was called "two-path architecture". It is a mix o f

7/36

ANN. TI~Lt~COMMUN.,

57, n~ 5-6, 2002

428

c. ROLAND

-- A SELF-ADAPTIVE

UNIVERSAL

RECEIVER

ideal SWR and SDR 13 techniques. In any case, the WBA phase is performed through a true SWR architecture path. Two variants of it are presented below. II.2.2.1. With one ADC The first a r c h i t e c t u r e uses only one ADC. T h e d i g i t i z e d b i t - s t r e a m is u s e d b y b o t h phases, as is p r e s e n t e d in F i g u r e 7. But the d i f f e r e n c e lies in the signal at the input o f the ADC. As e x p l a i n e d p r e v i o u s l y during the WBA p h a s e , the true 8WR architecture is p o s s i b l e and then the signal is directly the RF signal after the antenna. But for the d e m o d u l a t i o n phase this b e h a v i o r is p o s s i b l e only in some rare cases as IS95 or DEFT, but in general, this is not p o s s i b l e and we have to reduce the signal b a n d at the ADC input. This c o u l d be done b y one o f the n u m e r o u s m e t h o d s o f S o f t w a r e D e f i n e d R a d i o . In this p a p e r we will not d e s c r i b e these techniques, since, they are, for the b e s t k n o w n d e s c r i b e d in [10, 11] and in [12].

Path 1 J ADC L b J DSP Analog part SDR

"1

Path 2

I 1

w

FIG. 7. -- Two-path architecture with one ADC. A r c h i t e c t u r e gz d e u x c h e m i n s a v e c un s e u l CAN.

II.2.2.2. With two ADC'S This architecture differs from the previous one b y in that there are 2 AOC'S (see Figure 8). With this slight c o m p l e x i t y increase, we could obtain a very interesting improvement. This would consist in running o f both functional phases at the same time. That means we are able to analyze the r e c e i v e d signal and c o m m u n i c a t e simultaneously. In addition, this m e t h o d reduces the ADC requirements during the d e m o d u l a t i o n phase. In fact, the input signal c o m e s from the SDR interface (path 2) and therefore has a l i m i t e d band.

13. In this paper, SDR is a SWRapplied to restricted bandwidth. ANN. T~L~COMMUN.,57, n~ 5-6, 2002

8/36

429

C. ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

Path 1

J

ADC

DSP

J ADC "1

DSP

"1 Path 2

Analog part SDR

FIG. 8. - - T w o - p a t h a r c h i t e c t u r e w i t h t w o ADCS.

Architecture f deux chemins avec 2 CAN.

II.3. T h e WBA : T h e first f u n c t i o n a l p h a s e o f o u r SAUR

Now, in the rest of this paper, we will focus on the WBA which is the "heart" of our proposal. This description first comprises an explanation of the need of the iterative process to adapt the width of the analyzed band. Then we will describe in detail the Blind Analysis itself, performed during the last iteration of the iterative process. This blind identification of the standard in use will be shared into two sections. The first one (II.3.2) will study the main parameters of almost all the existing wireless standards, in order to find which one will be used in the second section. This second one (part III) describes the blind identification method itself, based on Radial Basis Function Neural Networks. I I . 3 . 1 . The iterative process for t h e WBA

The ratio between the global bandwidth to be analyzed and the smallest bandwidth parameter to be recognized is very high. It is around 105 as illustrated in Figure 9. No known Spectral Analysis techniques could performed in these conditions, because the resolution ration is too high. Consequently, we proposed an iterative adaptation of the Wide Band to be analyzed. Figure 10 presents this process. At each iteration, we research energy in the band with a conventional periodogram, then we filter and decimate the samples around this detected peak of energy.

>

<

10 kHz

1 GHz FIG. 9. - -

W h y a r e d u c t i o n o f the a n a l y s e d b a n d ?

Pourquoi une rdduction de la bande d'analyse? 9/36

ANN. TgLI~COMMUN.,57, n ~ 5-6, 2002

430

c . R O L A N D -- A S E L F - A D A P T I V E U N I V E R S A L R E C E I V E R

I

/

//

/,/"

4 GHz

400 MHz ./ 84184184 ~84

40 MHz

FIG. 10. -- The iterative process during the WBA. Le processus itgratif d'analyse spectrale.

11.3.2. Standard parameters study

During the WBA phase, the SAUR(defined previously) should recognize at least one standard with which it could be connected. Typically, system identification studies, tried to determine the type of modulation or other characteristics of the signal with, for example, the use of Second Order Statistics methods [13, 14, 15], Neural Networks or Monte Carlo Markov Chain methods, in order to determine the mapping of the modulation [16, 17, 18, 19]. These approaches, very interesting from the Signal Processing point of view, are nevertheless limited for determining the transmitted system. This is because, they cannot be performed directly on the digitized signal. In fact, it is necessary to determine many other parameters of the signal like Symbol Frequency, Channel Coding, as well as the parameters of the upper layers. In practice, this type of process can be very long and very computationally expensive. So we have chosen a different approach. This consists firstly in analyzing almost all the existing systems in use and to determine the discriminating parameters. Secondly, we recognize this discriminating parameter in a blind manner. In this section we will analyze almost all the existing radio standards in order to highlight the parameters that are able to identify the standards. We have voluntarily limited our study to the physical layer of the standards. In fact, the parameters of this layer are directly responsible for the shape of the transmitted signal, and further, they are easier to access than the upper layer parameters. Obviously, upper layer parameters are accessible only after channel filtering and demodulation, which indeed implies knowing the standard. Remark : this parameter discrimination study is useful only for existing standards and in a cooperative environment. II.3.2. I. Discriminating power of the parameters Every standard defines channel shape in order to optimize the spectrum sharing among operators and among users. Consequently the goal of a standard is to maximize, for all the users, the spectrum efficiency in terms of bits/(s.Hz.m2). In other words, we could say that a standard is defined in a 3 dimensional space. It includes the 3 following domains : the time domain, the ANN. TELI~COMMUN.,57, n~ 5-6, 2002

10/36

C. R O L A N D - A S E L F - A D A P T I V E

431

UNIVERSAL RECEIVER

frequency domain and the space domain, as is presented in Figure 11. Having said that, it is easy to understand that the frequency sharing, within one standard, is related to the parameters FDD, FDMA or FDM. There is also a frequency discontinuity between two different standards. If frequency and t i m e d o m a i n s are closely connected, a p a r a m e t e r has been defined for one domain and has " o n l y " influence on the other domain. In any case, we will be interested in the discontinuities both in the time and frequency domains because they are easier to measure.

PSD

f

/

/ / A

9

C h a n n e l BWe

9

Space I localization

~1

Iq

I

"-

Ap

I'

Standard B W

FIG. 11. -- Position of the parameters in the 3 domains (Space, Time and Frequency). Les param~tres dans les trois domaines (espace, temps et frdquence).

II.3.2.2. Accessibility of the parameters In order to find the most discriminating parameters, we also have to find their accessibility. In other words: " W h i c h operations are needed to access this parameter and which other parameter do w e also need to k n o w ?". In Figure 12, there is a conventional h a r d w a r e separation between cable and wireless standards. The other separation is more interesting for us. It is the typical " i n n e r channel" and "outer channel" separation. That means the "outer channel" parameters are m o r e closely related to the frequency d o m a i n and, conversely the " i n n e r channel" parameters are m o r e closely related to the time domain. 11.3.2.3. Conclusion and choice In this paper, w e will not describe the tedious w o r k consisting in classifying all the parameters. These w e r e classified using the classification of the three domains a b o v e presented. A f t e r numerous tables and cross-tables, we easily prove that two parameters are totally dis11/36

ANN.TI~LflCOMMUN.,57, n~ 5-6, 2002

432

c . R O L A N D -- A S E L F - A D A P T I V E

UNIVERSAL RECEIVER

FIG. 12. -- Accessibility of the parameters. Accessibility des paramOtres.

criminating, one in the space domain, the other in the frequency d o m a i n (the "outer channel" domain). This latter will be better adapted for blind recognition purposes. The first p a r a m e t e r corresponds to the location o f the user associated with the frequency plan. This p a r a m e t e r has already been studied in the context o f a universal receiver and we h a v e p r o p o s e d a solution w h i c h integrates a l o c a l i z a t i o n f u n c t i o n (GPS or GALILEO for e x a m p l e s ) associated with a database of the possible systems in use in the calculated location. This data base can obviously be updated. This solution is fully described in [20] and presented in Figure 13. To fulfill all the requirements, this p r o p o s a l should find an efficient solution to the i n d o o r localization problem. We shall now describe the second parameter, the one w h i c h will be retained for the SAUR. The Table I presents the channel bandwidth (swc) for m a n y wireless systems. This parameter is also the distance between two consecutive carriers o f the standard. Reading this table, it seems very clear that the 8wc parameter discriminates almost all the systems. In some rare cases, it is not discriminating as for DVB-T and some LMDS systems. In such case, we should

ANN. TI~LECOMMUN.,57, n~ 5-6, 2002

12/36

C. ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

433

Table I. - - C h a n n e l Bandwidth: the most discriminating parameter. Bande passante du canal: le paramdtre le plus discriminant.

Channel Bandwith 25 k Hz 30 kHz 100 kHz 200 kHz 300 kHz 1 MHz 1.25 M H z 1.712 M H z 1.728 M H z 5 MHz 7 et 8 M H z 10 M H z 20 M H z 32 et 36 M H z 50 M H z

Standard PDC 5 ADC 14 (D-AMPS) CT215 GSM 3 PHS 7, PACS 16 RLAN 9 Bluetooth IS954, Globalstar DAB 11 DECT 6 UMTS s (FDD) DVB_T 12, LMDS 17 RLAN, Hiperlan l~ I Hiperlan I et II DVB-S, LMDS Hiperlan II

FIG. 13. - - SAURwith a localization function. RUA avec une fonction de localisation.

use a secondary

level parameter

this secondary level parameter

to discriminate these two standards.

In this particular case,

will be the Guard Interval which fully discriminates

between

14. ADC: American Digital Cordless. 15. CT2: Cordless Telephone of 2 nd generation. 16. PACS: Personal Access Communication System. 17. LMDS: Local Multipoint Distribution Services. 13/36

ANN. TI~LI~COMMUN.,57, n ~ 5-6, 2002

434

c. ROLAND

-- A S E L F - A D A P T I V E

UNIVERSAL

RECEIVER

mono and multi carrier modulations. The channel bandwidth is a very reliable parameter because when it is defined it supports the standard and it cannot be easily modified. If the mvc value is clearly an "outer channel" parameter, its shape (pattern) depends on several "inner channel" parameters, like the modulation, the channel coding and the channel filtering. This specificity will be used in our method for the recognition of the parameter. In fact we will not only measure the 8wc value, which could easily be done with a threshold [21] but we will also use all the information given by the shape of the spectrum in the bandwidth. In our approach to recognizing a mvc, it is exactly like a "pattern recognition". It is as though we were doing a recognition of several parameters in the same operation. For that purpose, we use Radial Basis Function Neural Networks. This will be described in the following sections.

III. B L I N D R E C O G N I T I O N

In the previous section, our choice was to recognize the Bwc shape. The question now remains, how to find this shape on the received signal. The answer is after all very simple. One only has to perform a Power Spectrum Density (PSD) o n this signal in order to obtain this Bwc shape. Section III. 1 details the key equations to obtain this PSD. Then we compare this shape with reference spectrum shapes presented in section III.2. This comparison is performed using RBF NN (section III.3).

III.1. E q u a t i o n s o f the received signal The received Signal has already been presented Figure 4. It is the result of the summation of many standards. Each standard being itself the summation of several modulated carriers (channels).

9 Modelization without channel transmission We keep in mind that it is the 8wc parameter which we are interested in. Consequently we will not retain, in this model, parameters which have a slight influence on Bwc. If x(t) is the received signal, using the well-known mono-carrier modulated signal equation, we obtain: (1)

s Ps x(t) : ~ ~ (fern s (t)*ms,p(C(t))) exp(2~jf~.,pt) s=l

p=l

where ms,p(C(t)) is the modulation of carrier p with c(t) the mapping function and ferns(t) its shape filter. This received signal comprises S standards and Ps modulated carriers for each standard. We voluntarily do not take into account the channel coding aspect in this formula, under the assumption that the codes have only a slight influence on the bandwidth shape.

ANN. TI~LI~COMMUN.,57, n~ 5-6, 2002

14/36

435

C. ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

Replacing

ms,p(t) by

gs(n)An,s,p(n) exp(jtp,,s,p(n)),

its expression ~,

(1) becomes (2) for a

QAM monocarrier modulation type:

x(t) =

(2)

• ~" (:ems(t)

)

gs(n)An,s,p(n) exp (jq)n,s,p(n))

s=l p=l

exp

(2lrjfs,pt)

with A and (p, the amplitude and phase components of each symbol, and filter, with the notation (n) equivalent to (t-nT).

g(n)

the impulse

Replacing the modulation ms,p(t) by its expression, equation (1) becomes (3) for a QAM multicarrier modulation type:

x(t) =

(3)

~,

s=l p=l

)

ms(t)

l=l

In the simple case where

Lp =

~, =gs(n) An,s,p,! (n) exp (JtDn,s,p,l(n))) exp(27cjlfspt) 1, equation (3) becomes again (2).

9 Modelization with the channel transmission

We add to equation (1) the channel impulse response hp(t) at frequency fp, and the noise

b(t) at the ADC input. (1) becomes (4): S

(4)

P

x(,) : Z Z hA.* (:em:*ms~ (c(,)) oxp(2~H:) + b(,) s=l p=l

At the receiver side: 9 Sampling x(t) at sample frequencYfe, we obtain:

s P x(kTe) = Z Z hs,p(kTe) * (fems(kTe)*ms,p(c(kTe))

(5)

) (2~jkfp/fe) + b(kT e)

exp

s=l p=l

9 Then after quantization (6)

x(kTe) = Z Z hs,p(kTe) *

ms(kTe)*ms,p(C(kTe)) exp (2~jkfp/fe) + b~ (kT )

s=l p=l

In this equation the total noise b~ is the summation of the input noise with the quantization noise. The Power Spectral Density [22] which is easily computed like the average of N square modulus EFT of the equation (6) becomes: (7)y(k) = N(~=I ~

15/36

[

Fems(~ofP

,~'~H~'P:(~e-k)=l

,

,

2

ANN.TI;LI~COMMUN.57, , n~ 5-6, 2002

436

c.

ROLAND

-- A SELF-ADAPTIVE

UNIVERSAL

RECEIVER

111.2. T h e r e f e r e n c e signals The reference spectrum signals are given by the following equation:

f,

2

A k= fP )'ref( ) Fems(~e -- k)

(8)

)'mods(2

-- k )

This corresponds to the product of the modulation PSD by the transmitter filter modulus. With the modulation PSD given by:

fp (9)

~ n o d , ( ~ -e

1 --

k)=

N

~n~=l

M s , p,n(ff~P e

--

k)

2

It can be seen that equation (8) is for one standard (S = l) and for one carrier (P = 1), as well as for a perfect channel equivalent to equation (7). This is exactly the equivalence we would like to recognize. Figure 14 presents 3 examples of reference signals. One for each class of signal: 9 DAB for the multicarrier signal class 9 GSM for the GMSK modulation class 9 LMDS for the QAM Nyquist filtering modulation class

DAB

~/mod o dB

:/

0

GSM

: ---...:

.......

0

-5

-10

-10

-10

-20

-20

i -15 i J -20

'

'

'

i

~25 -3(

/

-i

j 100

200

300

-30

,

./ . . . . . . . . . . . . . .

,

,

\

.

.

~

...........

. . . . . . . . . . .

2\

-40

-50

Number of points of PSD

_

.......... ~

-30

/ i . . . . . . . . . . . . . . .

-60

LMDS 7--7

-50 20

40

60

Number of points of PSD

S0

-60

500

1000

1500

Number of points of PSD

FIG. 14. - - R e f e r e n c e signals.

Spectres de rdfdrence.

Remark: Multicarrier

(OFDM)

spectrums are designed by a rectangular window.

111.3. N e u r a l N e t w o r k D e s c r i p t i o n Several Neural Network classes are described in the literature [23, 24, 25] in which the reader could find all the basic information. Among them, the better-known are the multilayer Perceptron, Kohonen network, Z-FI (Sigma-Pi) network and the Radial Basis Function network. Each of them is better adapted for a particular application. For example, Multilayer ANN. TI~LECOMMUN., 57, n ~ 5-6, 2 0 0 2

16/36

437

C. ROLAND -- A S E L F - A D A P T I V E UNI VERSAL RECEIVER

Perceptron is well adapted for equalization purposes, and RBF well suited for pattern recognition. Because, as previously described, our problem looks like a pattern recognition, we decided to use RBF Neural Networks [24, 25, 26]. Therefore, there are no comparisons with others NN'S in this paper. The network will recognize the shape of the Bwc. One of the advantages of the RBF NN will be the intrinsic strength against channel perturbations. A conventional RBF NN features 3 layers (the input layer, the hidden layer and the output layer). In addition, some pre- and @r post-treatment can be used. Generally, the network output S(k) is given at step k by equation number 10 which corresponds to the weighted linear combination of the outputs of each neuron number i. L

(10)

S(k) = Z widi (Xk) i=1

where w i is the weighting coefficient of the considered neuron i. The output of each neuron is the result of a function which measures the distance d i between the input vector X and the center C i (or the kernel) of the neuron i. (11)

d i (S) : f(r) =

f(Plx -

ci[I)

Practically, the functions which compute the distance d(X) are the following: f ( r ) = r, f ( r ) = r 2,

and

f ( r ) = exp(-r2/o e)

The Gaussian function that is the most widely used for classification. It is not so easy to use in our problem because it implies knowing parameter r The behavior of the Network comprises two phases. The first is known as the training phase. This could be done in two different ways: one consisting in the adaptation of the weights of the neurons, the other in the addition or subtraction of references (neurons) in the network. The second phase corresponds to the recognition phase itself. In our problem, these two phases are not performed at the same time. Moreover, we will not implement a training phase, since the number of references is known and fixed and, under the assumption of cooperative transmission, their shapes are also known and defined precisely by the standards. Obviously, in case of standard evolution or new system definition, the SAUR software could be updated so that the RBF NN includes these new references. Nevertheless, the specific pretreatment we will describe could be considered as a type of training function. In this section we deal with the necessary adaptation of the RBF NN to our BWC recognition problem. The parameters (or functions) to be adapted are: the pretreatment (see 111.3.1), the Kernelfunction (see II1.3.2) and the neuron number as well as its parameters (references, threshold...). The received signal has already been presented (see III. 1). On this signal, we performed a Power Spectrum Density (PSD), and thanks to the RBF NN, as presented on Figure 15, we compare it with the reference signals. The NN input vector is composed by the K points of the PSD ~k) (cf. equation 7). To each neuron number i corresponds the standard number i. In fact, we consider that the DSP reference signal ~'ref(k) (see equation 8) is exactly the Kernel center Ci, k of the neuron.

17/36

ANN. TI~LECOMMUN.,

57, n~ 5-6, 2002

438

C . R O L A N D -- A S E L F - A D A P T I V E U N I V E R S A L R E C E I V E R

.....'" H i d d e n l a y e r s ...... ._ de (I) R B F ........../

...................

J

: .......... '

J

......~

l

. . . .

.....

{w,

.....

V / / ~ '

, "Output

' ~"

'

layer.,

(d)

..,

4 ..............

.--"

~

200kHz

......................

x(t+mL):

~ PSD ! MJ- 1

':~, .L

! :

;~t(7, c)',---~,

' '

8

J , ........ '

-<

FIG. 15. - -

'

.

'-'

;

Complete Radial Basis Function Neural Network.

Le rdseau neuronal dtfonctions de base radiales compldtes.

The output neuron will choose which standard is active thanks to the information given by the outputs o f all the previous neurons. This output neuron is the object of section III.3.5,

III.3.1. The pretreatment

Figure 16 presents the pretreatment which c o u l d itself be shared in three different main functions (sharing the analyzed bandwidth in cuts function, suppression o f some cuts, normalization o f these cuts). 9 The sharing function B e c a u s e each reference signal is obtained with a PSD o f different length, it is n e c e s s a r y to cut the a n a l y z e d PSD signal in m a n y cuts o f the s a m e l e n g t h in order to c o m p a r e these cuts with the reference signal. The number o f cuts N i for one reference is given b y the ratio M / L i , where M is equal to LFFT/2 half the length o f the PSD) and L i the number o f points o f the ith PSD o f the reference C r 9 Suppression o f cuts We compute the minimum, m a x i m u m and m e a n p o w e r s o f all the N/cuts. For each cut, if the ratio b e t w e e n the mean power and the m a x i m u m p o w e r is less than a predefined threshold, then this cut is not used. This is the case w h e n the f o l l o w i n g equation is verified:

ANN. TI~LI~COMMUN.,57, n~ 5-6, 2002

18/36

439

C . R O L A N D -- A S E L F - A D A P T I V E U N I V E R S A L R E C E I V E R

PSD

v

v

V Cup up into BW~band

I

v

comou,a,,o, I oo.;s l~ ........ , F'rn!.,pma•

Suppression of band < 5 dB

v

V

:

| /

AGC of each band

Radial Basic

Function

output neuron

FIO. 16. -- Pretreatment functions. Fonctions de prEtraitement.

(12)

101og[

?~m.x.t~ < 5 d B

1 x~,

this operation is twofold. Firstly, it decreases the number of comparisons and then decreases notably the computation burden; secondly, it also decreases the number of "false detections". In fact after AGC (see following function), the noise could, in certain cases, be recognized as true signals. 19/36

ANN. TI~LI~COMMUN.,57,

n~ 5-6, 2002

440

c . ROLAND -- A SELF-ADAPTIVE UNIVERSAL RECEIVER

9 The normalization function We have to perform a normalization function (like an AGC) in order to have the same mean power in the analyzed signal and in the reference signal. Then it is possible to perform a normalized error function computation. Then the modified spectrum becomes:

1 Li

-Zi 2ci, (13)

~ k m o d = ~/k

1

Li

I11.3.2. The Errors function of the neuron

We have compared three different error functions. First, two conventional Mean Square Error functions on the differences between the two PSD on both on the linear (equation (14)) and log scales (equation (15)). EQM lin = - Li

(14)

(,,-it)C

2

1 Li EQM log = ~ii t~=~ ( l o g ~ - l o g C i , l ) 2

(15)

The first error function is well adapted for measuring smooth transition as well the second one is better adapted for measuring abrupt transition (like the PSD in dB scale). Because we will benefit both effects we defined a new error function. This third error function is given by equation (16). It corresponds to a combination of the two previous functions. It is with this last one that we obtain the best results, particularly on GSM signals. 1 Li

EOM C~

(16)

=

-Lii l~=l((']Zl-

C

i,I)2 X ]log(~//Ci,/)l)

III.3.3. Threshold of a neuron

The threshold T/is the value which separates the activation or not of the output Si( ^ k ) of the neuron i. That means for the ith n e u r o n : If

EQM Comb