Magnitude and Phase Characteristics. H [db] Frequency [MHz] x Phase [rad] Frequency [MHz]

Transcription

1 EQUALISATION AND RADIO FREQUENCY INTERFERENCE CANCELLATION IN BROADBAND TWISTED PAIR RECEIVERS Dimitrios I. Pazaitis, Jan Maris, Serge Vernalde, Marc Engels, Ivo Bolsens Digital Broadband Transceivers Domain, VLSI System Design Methodologies Division, IMEC, Kapeldreef 75, 3001 Leuven, Belgium, Tel.: +32 (0) Fax: +32 (0) f pazaitis, maris, vernalde, engelsm, bolsensg@imec.be ABSTRACT In this contribution a hybrid equalisation concept for a single carrier very high bit rate twisted pair downstream receiver is proposed. The equaliser's requirements and performance are investigated and a technique aiming to combat radio frequency interference (RFI) digitally is proposed. The proposed method employs an adaptive notch lter to trace and reduce the RFI levels and - consequently - reduce the equaliser implementation complexity. Simulation results illustrate the desirable characteristics of the proposed scheme and conrm its ability to eciently cope with narrowband interference. 1 INTRODUCTION Recently, there has been an increasing interest in providing high speed communication services over residential access networks which has led to the development of high speed technologies [1, 4, 7, 8, 9, 12]. Although the concept of utilising the higher frequencies available on a telephone twisted pair for providing broadband access is not a new one, it has only become feasible very recently due to the advancements in digital signal processing (DSP) technology. Among the most promising broadband access technology options for both residential and business users is the xdsl (Digital Subscriber Line) family of technologies. These broadband copper technologies employ highly sophisticated techniques to limit the distortion eects, greatly expanding, therefore, the bandwidth potential over a single pair of copper wires. A single copper wire is used to carry both the forward and the reverse channel data streams. Depending on the service, the available bandwidth is symmetrically or asymmetrically allocated to the downstream (data trac from the service provider to the subscriber) and upstream (trac from the subscriber to the service provider) channels. In this contribution, equalisation aspects of a single carrier, 26 Mbits=s VDSL (Very high rate Digital Subscriber Line) downstream receiver are addressed. Single carrier modulation is one of the two main line encoding technologies - the other one is discrete multitone modulation - envisaged for VDSL modems. Aiming at a cost ecient ASIC implementation our study focuses on a single carrier receiver employing a QAM modulation scheme. The investigation is part of a broader project targeting a fully digital copper cable downstream receiver. Among the major challenges that a broadband receiver has to face, Radio Frequency Interference (RFI) is the most crucial one. Since VDSL invades the frequency ranges of AM and amateur radio, every above-ground telephone wire acts as an antenna that both radiates and attracts energy in these radio bands. The presence of strong RFI interference signicantly constrains both the data rates and the loop length and necessitates the use of long and complex equalisers. Furthermore, there are cases where even long equalisers fail to guarantee high transmission rates. To overcome this problem, a new RFI suppression solution is presented, based on the use of a constrained adaptive digital notch lter. The proposed technique signicantly enhances the receiver's performance while reducing, at the same time, the complexity of the equaliser. Although the paper discusses equalisation of VDSL channels, the results could be easily extended to other single carrier xdsl technologies. The structure of the paper has the following form: In Section 2 the VDSL environment is introduced and the proposed RFI suppression technique is explained in Section 3. In Section 4 the system and simulation defaults are provided and in Sections 5 and 6 equalisation and RFI suppression results are presented and discussed. Finally, the conclusions are summarised in Section 7. 2 VERY HIGH BIT-RATE DIGITAL SUBSCRIBER LINE (VDSL) VDSL is the broadband technology that converts the existing copper wire telephone infrastructure into very high-speed network access by transmitting very high speed data over short reaches of twisted-pair

2 telephone lines. The range of data rates depends on the actual line length and the downstream rate under consideration ranges from 13 Mbits/s to 52 Mbits/s. Upstream rates in early models will be asymmetric, just like in ADSL (Asymmetric Digital Subscriber Line), ranging from 1.6 to 2.3 Mbits/s. Although still in the denition stage, standardisation eorts are undertaken by various organisations. In our work the spectral allocation was carried out according to the ANSI and DAVIC standards [2, 11] and the main downstream modem parameters are presented in Table 1. Modulation QAM-16 Baud-rate (R) 6.48 Mbaud Bit-rate (R b ) Mbits/s Carrier Frequency f c 4.86 MHz Roll-o factor 0.2 Table 1: Spectrum allocation according to ANSI and DAVIC (prole C) The signal to noise ratio (SNR) at the output of the slicer should be sucient to provide the targeted bit error rate (BER) of 10?7 plus a 6 db margin. To achieve this, the VDSL equaliser has to successfully face many challenges. Twisted copper lines exhibit signicant propagation loss (especially in high frequencies) as the cable length increases. Moreover, unterminated or incorrectly terminated lines, known as bridged taps, may have very detrimental eects on VDSL in certain congurations. The spectral characteristics of a channel with a bridged tap are presented in Figure 1. The channel corresponds to the VDSL6 test-loop as dened by ANSI [2]. H [db] Magnitude and Phase Characteristics Frequency [MHz] x 10 6 Phase [rad] Frequency [MHz] x 10 6 Figure 1: Spectral characteristics of a 1km twisted pair channel with one bridged tap (ANSI VDSL6) Other impairments manifested in the form of noise include additive white Gaussian noise (AWGN), impulse noise, Near (NEXT) and Far End Crosstalk Interference (FEXT) from other xdsl services. By employing frequency division multiplexing, NEXT inuence is signicantly reduced and we can therefore concentrate on FEXT, whose power spectral density for N disturbers can be modeled by [2] P SD f ext = P SD dist jh(f)j 2 N 49 :6 910?20 df 3 2 ; (1) where P SD dist is the power spectral density of the disturber, H(f) is the channel frequency response and d is the loop distance in feet. However, as mentioned earlier, the main challenge that a VDSL receiver has to overcome is the radio frequency interference. This type of interference, mainly due to AM and amateur radio signals, appears in certain frequency bands and severely limits the single carrier receiver's performance. Compared to the radio amateur signals, AM radio signals pose a lesser threat to the modem due to their xed spectral location. On the other side, the random characteristics of the amateur radio signals, render necessary the ability of the receiver to quickly trace and respond to signal changes with minimal performance degradation. The introduction of the adaptive notch lter, described in the following paragraph, aims to facilitate the role of the equaliser in suppressing the RFI interference and, consequently, reduce its complexity requirements. 3 RFI SUPPRESSION USING ADAPTIVE NOTCH FILTERS Adaptive notch lters have been widely used in many applications including telecommunications and control to extract, eliminate or trace narrowband/sinusoidal signals embedded in broadband white signals [3, 5, 10]. Although originally aimed at signals embedded in broadband white noise, they can be eciently used to cancel RFI superimposed on coloured signals, including xdsl type signals. In our work, we concentrate on the use of a simple constrained IIR adaptive notch lters for RFI reduction. The general form of the constrained notch lter is given by H(z?1 ) = ny k=1 1? 2cos! k z?1 + z?2 ; (2) 1? 2 k cos! k z?1 + 2 kz?2 where! k are the notch frequencies and k are the corresponding pole contraction factors, which dene the bandwidths of the notches. It is easily seen that the above notch lter consists of a series of second order notch lters in cascade and its transfer function has its zeros on the unit circle, resulting in a zero gain

3 at the notch frequencies. For the k th zero-pole pair, the bandwidth of the notch is approximately equal to BW (1? k ): (3) Thus, when the parameter is very close to one, the function H(z?1 ) behaves like an ideal notch lter. Using a recursive prediction error (RPE) adaptive algorithm, the update of the parameter ^ n =?2cos! n of each second order adaptive notch lter in (2), at time instant n, is given by the following set of equations [5, 10] ^ n+1 = ^ n + R?1 n n e n ; (4) R n = n R n?1 + 2 n ; (5) e n = 1 + ^ nq?1 + q?2 1 + ^ n q?1 + 2 q?2 y n; (6) where n is an approximation of the negative prediction error gradient n ^ ^n =?y n?1 + e n?1 : (7) 1 + ^ n q?1 + 2 q?2 In the above formulas, e n is the output of the notch lter, y n denotes the input to the second order notch lter and q?1 stands for the unit delay operator. If desired, a data adaptive pole contraction factor can be derived by applying the same gradient type optimisation [5]. The convergence behaviour was studied and stability conditions were obtained by applying the ordinary dierential equation (ODE) methodology [5, 10]. The tracking properties of the above adaptive notch lter are very sensitive to the notch bandwidth () and -especially- to the forgetting factor ( n ) choices. These choices together with the order of the notch lter, which denes the number of notches, will be discussed in detail later. 4 SYSTEM - SIMULATION SETUP The block diagram of the simulated QAM-16 VDSL system is shown in Figure 2. The receiver consists The use of a fractionally spaced equaliser removes the receiver's sensitivity to the sampler's phase. Equalisation is carried out in baseband whereas RFI cancellation is performed in passband. RFI cancellation could be also carried in baseband and this topic will be addressed at a later point. As a transmission medium, the 1km length channel depicted in Figure 1 is used. As specied in the ANSI standard [2], a power spectral limit of?60 dbm=hz is applied to the information signal resulting to an approximately 8:1 dbm transmission power level. The AWGN added to the system follows the ANSI model and has a at spectral power density of?140 dbm=hz (total power of?71:9 dbm). FEXT interference is modeled according to (1) with N = 20 and the average FEXT power was measured to be around?66 dbm. AM broadcast sources are modeled by a xed frequency carrier 30 % AM modulated with a at Gaussian noise source band limited to 5 KHz. Ten AM stations are simulated and the average power levels of the AM signals are specied in Table 2. Frequency Power Frequency Power [KHz] [dbm] [KHz] [dbm] Table 2: AM noise. Model 1 : High density urban with co-located transmitters (ANSI and DAVIC standards) From Tables 1 and 2 it can be seen that three AM signals are out of the signal band whereas seven of them are in the lower part of the signal band. Amateur radio signals are modeled as single sideband signals, band-limited to 5 KHz but unlike AM signals, they appear randomly within the specied radio amateur bands [2] (Table 3). The rst three RFI bands cos ω c t + Channel -sin ω c t Notch filter cos ω c t AGC -sin ω c t Complex Feedforward Equaliser - Decision Feedback Equaliser error - Decision device Frequencies MHz MHz MHz MHz MHz MHz MHz Figure 2: The simulated QAM-16 VDSL system of an RFI canceler, a simple automatic gain control (AGC), a down-converter, a fractionally spaced feedforward (FFE), a symbol spaced decision feedback (DFE) equaliser and a slicer. Table 3: Radio Amateur Bands overlap with the transmitted spectrum, severely distorting the received signal. We can now proceed to the presentation of equalisation and RFI suppression results.

4 5 EQUALISATION Adaptive equalisation is necessary in order to ensure high data rates on twisted pair loops. This and the following section address the equaliser requirements for the successful operation of a VDSL receiver. This includes the lengths of the equalisers, convergence and step sizes and initial training schemes. In this section, we focus on the equaliser's performance in the absence of radio amateur interference. In order to avoid the training sequence and the extra complexity that accompanies it, blind training methods can be applied. The constant modulus algorithm [6] was proven successful in opening the eye. However, motivated by the desire to further simplify and at the same time speed up the initial convergence, the following training procedure was developed, which relies only on the use of decision directed Least Mean Squares (LMS) to open the eye. The whole procedure is as follows. During the initial phase of adaptation, the DFE is switched o and random data samples obtained from a QAM-4 (the four corner points of a QAM-16) constellation are transmitted. As soon as the eye is opened, the DFE is switched on. At the end of this stage the 16-point QAM constellation can be switched to. This technique proved successful under a variety of conditions and all the simulation results presented in this paper were obtained by using it. The lter's length determines the degrees of freedom available to the equaliser. However, the selection of the equaliser's length is dictated by a tradeo between performance and implementational complexity. In the VDSL channel case, the eect of the equaliser length is shown in Table 4, where the receiver's performance is presented for various FFE and DFE lengths. During that experiment only one crosstalk channel (FEXT interference) was present. As a performance measure, the SNR at the output of the decision device (slicer) was taken and the results correspond to steady state operation. DFE taps FFE taps Table 4: System performance as a function of the equaliser lengths The update of the equalisers' coecients was carried out using the complex LMS adaptive algorithm [13]. The convergence factors F F E and DF E govern the stability of the feedforward and decision feedback algorithms, as well as the rate of convergence and the steady state excess mean square error in relation to the optimal Wiener solution. They should be therefore cautiously selected. The larger the step size values, the faster the convergence and the response of the equaliser to signal or system changes but also the larger the steady state excess error. The relative sizes of the convergence factors also determine the interaction between the two equaliser types. Extensive simulation results have shown that, in general, adaptation of the DFE coecients should follow the adjustment of the FFE weights. This is not so important in equilibrium state as during convergence to a new set of lter weights. The step size values used in the simulation were chosen equal to F F E = 1e?4 and DF E = 7e?5 and were shown to eectively combine fast response with low excess error. As it can be easily observed from Table 4, the existence of a decision feedback equaliser signicantly enhances the performance of the system both in terms of convergence speed and steady state error. Allowing more freedom in the choice of the feed-forward coecients also results in increased robustness against radio frequency interference. Table 4 also indicates that in the absence of radio amateur interference, an equaliser of a relatively short length could eectively compensate for the distortions due to ISI, FEXT and AM interference and there would be little to gain by employing longer lters. However, this does not imply that shorter equalisers tend to perform better. Longer equalisers require ner tuning and, consequently, smaller step size values to achieve maximum performance. The main advantage of longer equalisers lies in their improved ability to compensate for radio amateur signals distortion and this will be further investigated in the next section. 6 RFI SUPPRESSION In this section, RFI cancellation results using adaptive notch lters are presented and discussed. The constrained digital notch lters are positioned in passband before the equaliser, as depicted in Figure 2. Placing the notch lter after the equaliser would result in contention between the two units and, more importantly, it would introduce extra complexity in the equaliser's update. Furthermore, the lower clock rates required for a baseband implementation are traded-o against the complex nature of baseband signals. However, the potential benets of a baseband implementation due to the whiteness of the equaliser's output are still under investigation. To evaluate the eciency of the notch lter, extensive simulations were carried out. Radio amateur signals of various levels were added to the output of the channel at random frequency locations and the receiver's performance was observed. We focused on the case specied in the standard [2], where one strong radio amateur signal is present together with ten other AM station signals. During the sim-

5 ulations, FEXT interference and AWGN were also present. Frequency location [MHz] SN R [db] Table 5: Receiver's performance as a function of the RFI location In Table 5 steady state performance results for various frequency locations of RFI are presented. There is crosstalk interference from one user and the lengths of the FFE and DFE equalisers are equal to 50 and 20 taps respectively and the power level of the radio amateur signals is equal to?30 dbm. Without the help of the notch lter, even longer equaliser were unable to combat radio amateur signals with power levels higher than?60 dbm. In the case of N = 20 crosstalk disturbers the SNR at the output of the receiver falls to 20:7 db, regardless of the RFI location. The notch lter used consists of two second order adaptive lters in cascade, enabling therefore suppression of two RFI signals. The rst second order adaptive notch lter was observed to converge to the most powerful AM station signal (at 1:13 MHz) whereas the second successfully traced the radio amateur signal. Although a single second order notch lter was proved sucient in canceling the radio amateur signal in most cases, the existence of a second adaptive notch lter signicantly enhanced the exibility and hence the performance of the equaliser. One notch lter was, therefore, allocated to the AM frequency band and was, like the AM stations, permanently switched on. The second adaptive notch lter was switched on whenever a severe drop in the output SNR was observed. In our simulations, whenever the average SNR dropped below 10 db, the second notch lter was switched on. The value of the parameter, was taken equal to 1 = 0:995 and 2 = 0:99 for the rst and second adaptive notch lters respectively. Due to the spectral shape of the signal at the input of the receiver, the second notch had to be initially placed to the right end of the higher radio amateur band and was left to converge to the interfering signal. In the absence of RFI, the notch tends to converge to the peak of the spectral magnitude which is located around 1:6 MHz. If the second notch is initially placed far to the left from a relatively weak RFI signal, the notch will fail to trace the RFI and will converge to the closest spectral peak. The initial values of R and were chosen equal to R 0 = 8 and 0 = 0 respectively and were found to work well under a variety of conditions. Although the initial value of R determines the convergence steps during the rst iterations, the algorithm exhibited a relative insensitivity to its choice, provided that the initial steps are small enough. Results in Table 5 conrm the eciency of the adaptive notch lters and support their use in RFIthreatened environments. Although notch lters do not fully cancel RFI in all cases, they signicantly facilitate the equaliser's job by reducing its power level. By comparing the steady state SNRs of Tables 4 and 5 it is seen that the performance degradation due to the existence of the notch lters is -in most casesalmost negligible. The extra ISI which the notch lters introduce can be eectively compensated by the equalisers. The low SNRs observed for the rst RFI band are due to the proximity of the rst radio amateur band to the AM band. The equalisers have to compensate for the interference of closely spaced narrowband signals and this reduces their eectiveness. Longer equalisers could, however, tackle the problem successfully. Magnitude [db] Frequency Response Channel Output Notch Filter Output Frequency [MHz] Figure 3: AM signal suppression using an adaptive notch lter In Figure 3 a detail of the spectral (magnitude) characteristics of the input and output of the notch lter is displayed. The magnitude values are normalised and the shown part of the spectrum corresponds to the frequencies around 1:13 M Hz, where the most powerful AM station transmits. It is observed that the notch lter almost eliminates the narrowband AM signal facilitating therefore the equalisers' task. The convergence factors of the equalisers were chosen, as before, equal to F F E = 1e?4 and DF E = 7e?5. Freezing temporarily the update of the DFE coecients immediately after a radio amateur signal detection was found to reduce error propagation effects, accelerating the equalisers' response. However, a very long freezing period may delay the system's response. Although the optimal freezing interval is closely related to the RFI power levels and location, a freezing interval of 1000 symbols was empirically x 10 6

6 found to provide on average fast convergence. As mentioned earlier, the tracking properties of the adaptive notch lter are highly sensitive to the choice of the convergence factor. A large value increases the convergence rate of the notch lter but at the same it reduces its ability to nely track narrowband signals and may also result in suppression of information carrying frequencies. On the other hand, a very small value increases the required time to convergence and delays the receiver's response. In [5] it was shown that the asymptotically optimal value of is equal to the value of the pole contraction factor which determines the bandwidth of the notch. In our case, the bandwidth of the narrowband interference remains constant and constrained to a maximum of 5 KHz and we can, therefore, assign constant and equal values to and. In our simulations these values were taken equal to 1 = 0:995 and 2 = 0:99 for the rst and second notch lter respectively. The dierence in the two values reects the dierence nature of the AM and radio amateur signals. In AM signals the carrier is only 30% AM modulated, resulting in a high energy concentration on the carrier frequency. On the other hand, radio amateur signals are single sideband (suppressed carrier) signals and the energy is evenly spread over the entire single sideband. In the cases where multiple RFI interferences were present, the notch lter always converged to the stronger one, regardless of the frequency band. The ability to cancel higher level (more than?30 dbm) RFI signals was also observed. 7 CONCLUSIONS In this contribution, an equalisation solution for Very high bit rate DSL (VDSL) receivers was presented. The equaliser's requirements and performance were investigated and a new method to combat radio frequency interference (RFI) digitally was presented. The proposed method employs an adaptive notch lter to reduce the RFI levels and, consequently, relax the equaliser constraints, leading to a more ecient implementation. Simulation results under worst case conditions illustrated the desirable characteristics of the proposed scheme and conrmed its ability to eciently cope with narrowband interference, rendering it a very useful tool in high speed broadband communications. ACKNOWLEDGMENTS This work was carried out under the Flemisch Goverment Impulse Program for Information Technology. Marc Engels is a senior research assistant of the Belgium National Fund for Scientic Research References [1] Open Technical Issues in Provisioning High- Speed Interactive Data Services Over Residential Access Networks. IEEE Network, Special Issue, January [2] ANSI T1E1.4 :. Very-high-speed Digital Subscriber Lines, Draft Technical Document - Revision [3] B.-S. Chen, T.-Y. Yang, and B.-H. Lin. Adaptive notch lter by direct frequency estimation. Signal Processing, 27(2):161{176, May [4] B. Daneshrad and H. Samueli. A 1.6 Mbps Digital-QAM System for DSL Transmission. IEEE Journal Selected Areas Com., 13(9):1600{ 1610, December [5] M.V. Dragosevic and S.S. Stankovic. An Adaptive Notch Filter with Improved Tracking Properties. IEEE Trans. Signal Processing, 43(9):2068{2077, September [6] D. Godard. Self-Recovering Equalization and Carrier Tracking in Two-Dimensional Data Communication Systems. IEEE Trans. Commun., COM-28(11):1867{1875, November [7] G.-H. Im and J.-J. Werner. Bandwidth-Ecient Digital Transmission over Unshielded Twisted- Pair Wiring. IEEE Journal Selected Areas Com., 13(9):1643{1655, December [8] K.J. Kerpez and K. Sistanizadeh. High Bit Rate Asymmetric Digital Communications Over Telephone Loops. IEEE Trans. Communications, 43(6):2038{2049, June [9] J. Maris, P. Schaumont, S. Vernalde, M. Engels, and I. Bolsens. Dynamical Analysis of All Digital Symbol Timing Recovery in Twisted Pair Broadband Receivers. In Proc. 13 th Int. Conference on DSP, pages 1055{1058, Santorini, Greece,, July [10] D.V.B. Rao and S.Y. Kung. Adaptive Notch Filtering with Constrained Poles and Zeros. IEEE Trans. Acoustics, Speech and Signal Processing, ASSP-32:791{802, August [11] DAVIC Standard. [12] D. Tkoc, D.I. Pazaitis, and S. Vernalde. Adaptive Equalization of VDSL Channels. In Proc. Int. Workshop on Copper Wire Access Systems, pages 271{278, Budapest, Hungary, October [13] B. Widrow and S. D. Stearns. Adaptive Signal Processing. Prentice-Hall, New Jersey, 1985.