You are on page 1of 5

Analysis of Impulse Noise Mitigation Techniques for Digital Television Systems

Himal A. Suraweera, Colin Chai, Jinwen Shentu and Jean Armstrong Department of Electronic Engineering, La Trobe University, Victoria 3086, Australia. j.armstrong@latrobe.edu.au
Abstract-- Although multicarrier modulation techniques are inherently more resistant to impulse noise than single carrier systems, impulse noise can still be a significant problem in many applications including digital television. A number of impulse noise mitigation algorithms are investigated. These use non linear processing on the digital input signals to limit the size of impulses reaching the demodulator. It is shown that when the noise at the input is highly impulsive, significant reduction in the signal to noise ratio can be achieved. The most effective non-linearity is clipping of input signal samples with large amplitudes.

on Reed Solomon coding. The mitigation strategy described in [4] is a clipping method with nulling strategy. At the receiver the occurrence of an impulse is determined with a set threshold and is corrected by replacing it with a null in the time domain which is then subjected to FFT processing. A modified method of determining the presence impulse noise is also described where initially the high peaks are ignored and if subsequent peaks occur they are treated as perturbations and nulled. II. THE EFFECT OF IMPULSE NOISE ON OFDM OFDM is more resistant to the effects of impulse noise than single carrier systems because of the spreading effect of the receiver DFT operation. A single impulse can completely mask a symbol in a single carrier system, whereas in an OFDM system the energy of each impulse is spread evenly across all of the subcarriers in that symbol. When there is more than one impulse in a received symbol period, T, the contributions combine linearly in each subcarrier. When there are enough impulses during T for the central limit theorem to apply, the impulse noise at the output of the DFT becomes Gaussian. The overall effect of impulsive noise on the BER of an OFDM system depends on both the design of the system and the characteristics of the noise. In general, the larger the number of subcarriers, N, the more robust the system. The digital video broadcasting (DVB) standard allows a range of signal constellation and error coding combinations, so that data rate can be traded against reliability. Smaller constellations and more powerful codes will increase the resistance to impulse noise. In the DVB standard, there is no interleaving and error correcting across symbols, so the BER for a given symbol depends on the characteristics of the impulse noise present during that symbol period. Thus, in the absence of any impulse mitigation techniques, and when the central limit theorem applies, the BER for a given symbol depends on the total noise energy input during that symbol period.

I. INTRODUCTION Orthogonal frequency division multiplexing (OFDM) technology is used for digital television transmission in many countries including Australia and much of Europe [1]. OFDM was chosen for this application partly because it is more resistant to impulse noise than single carrier systems. However impulse noise can cause problems in digital television reception. Interference caused by turning on electrical appliances, or from the ignitions of passing vehicles, can cause the television picture to freeze. This paper considers the effect of impulse noise present in the digital samples at the input to the receiver Discrete Fourier Transform (DFT). Techniques for reducing the impact of impulse noise by clipping or nulling these digital samples are investigated. Impulses can also affect other receiver functions such as the AGC. These effects are beyond the scope of this paper. The effects of impulse noise in multicarrier systems have previously been analysed in the literature. For example, in [2] bit error rates (BERs) for OFDM QAM systems gave been analysed by varying impulse power and probability. Techniques for mitigating the effects of impulse noise include signal processing algorithms, coding and interleaving and clipping [3-5]. In [3] several impulse mitigation techniques are presented based
This work was supported by an Australian Research Council Large Grant

III. IMPULSE NOISE MODELS A number of models for impulse noise have been presented in the literature [6-7]. Some characterize only the probability density function (PDF) of the amplitude of the noise, whereas others also consider the time correlation of impulse events. Two noise models are used in this paper, the well known Middleton Class A model and a model which we have called a gated Gaussian model. Middleton Class A is a statistical-physical impulse noise model. In general noise is represented by its envelope (t ) and phase (t ) as (1) Let the in-phase and quadrature components of n(t ) be x(t ) and y (t ) , n(t ) = x (t ) + jy (t ) (2)

OFDM symbol and which have variance i2 during this time. (i.e. the variance is calculated over only 2 T not over T ). In general i2 >>> n . This model is used because it gives a better indication of the performance of DVB systems, where the critical factor is whether the SNR for each symbol is above a given threshold, rather than the SNR averaged over the entire received signal. Three parameters describe the gated Gaussian noise: 2 , the average variance of the total noise, 2 i2 n , and the fraction of time impulse noise is present. measures how structured the impulse noise and is analogous to A of the Middleton model, while is analogous to 1 . Fig. 1 shows a typical section of the gated Gaussian noise. The impulse noise samples are much larger than the background white noise samples, but are present for only a fraction of the time.
Impulse Noise

n(t ) = (t ) cos(2ft + (t ) )

m/ A+ 1+ (3) Integrating with respect to x or y , gives the PDF for the real or imaginary part of the noise. x2 Am (4) P ( x) = exp ( A ) exp 2 2 m = 0 m ! 2 2 m m where
2 m = 2

Real Part

Then for a Middleton Class A PDF for x and y is given as, x2 + y 2 Am P ( x, y ) = exp ( A) exp 2 2 m = 0 m ! 2 2 m m

i2

Channel Noise

Time 2 n

Fig. 1. Gated Gaussian noise samples

y2 (5) exp 2 2 m = 0 m ! 2 2 m m From the above PDFs it is clear that although the real and imaginary components are uncorrelated they are not independent: if a high peak real component occurs, it is highly likely that its imaginary counterpart will also have a high peak. Three parameters are used in the Class A noise model. The first, 2 , signifies the variance of the Class A noise. The next, A , is termed the impulsive index. Small A indicates highly structured interference, with large noise values for a small fraction of the time. is the mean power ratio between the channel noise and the impulse noise (non Gaussian component). Small and small A gives a predominantly impulse type noise. The second model used in this paper is gated Gaussian, in which the noise is the sum of AWGN 2 of variance n and Gaussian noise impulses which last for a fraction, , of the time duration of each P ( y ) = exp ( A )

Am

IV. IMPULSE NOISE MITIGATION When an OFDM signal is subject to impulse noise, the probability of large signal samples increases. One way to reduce the effect of impulse noise is to identify these large signal samples and reduce their effect by clipping or nulling them. Fig. 2 shows the block diagram of part of the OFDM receiver model considered. The received OFDM baseband signal samples are given by s(l ) = r (l ) + ng (l ) + ni (l ) (6) is the Gaussian noise and ni (l ) is the impulse where r (l ) is the wanted OFDM signal, ng (l )

noise. The samples s (l ) are input to a memoryless non-linearity designed to reduce the effect of impulse noise. The samples at the output of the non-linearity, z (l ) , are serial-to-parallel converted before being input to the N-point DFT. The samples at the output of the DFT are Z (l ) . In this paper we present results for the SNR of Z (l ) .

OFDM Signal

z (l ) s (l )
Memory less Non Linearity

r (l )
Gaussian Impulse Noise Noise

S/P

N Point DFT

Z(l )

ng (l ) ni (l )

z (l ) = ks(l ) + nc (t )

Fig. 2. Block diagram of receiver

Due to the central limit theorem, the complex baseband signal of an OFDM system with large N can be modelled as a complex Gaussian process. Thus the effect a non-linearity on an OFDM signal can be analysed using Bussgangs theorem, which states that a zero mean stationary Gaussian process distorted by a memoryless non-linearity can be expressed as the sum of an attenuated input replica and an uncorrelated distortion component [8]. Bussgangs theorem applies exactly to the effect of a nonlinearity on an OFDM signal plus Gaussian noise and approximately when impulse noise is present. Thus the signal at the output of the nonlinearity in Fig. 2 is given approximately by z (l ) = ks(l ) + nc (l ) (7) where nc (l ) is the clipping noise and k is a constant that depends on the nonlinearity. Expanding s (l ) gives z (l ) = kr (l ) + kng (l ) + kni (l ) + nc (l ) (8)

imaginary components separately or based on the amplitude of the complex signal. Which of these can more easily be implemented in practice depends on the design of the front end of the receiver. Two types of non-linearity are considered. Clipping - that is the maximum signal value is limited, or nulling - that is the signal is set to zero. It is not intuitively obvious whether clipping or nulling is likely to be the better strategy. If a sample is correctly identified as being subject to impulse noise, then nulling is likely to be the better technique, because the noise component of the sample is likely to be larger than the signal component, so setting the sample to zero will give more overall reduction in noise. However nulling will cause more signal distortion when a sample is incorrectly identified as impulse noise. Mathematical expressions for the non-linear functions applied to the samples s (l ) are given below. Amplitude clipping:

s (l ) s (l ) AClip f {s (l )} = AClip exp( j arg{s (l )}) s (l ) AClip


Real and imaginary clipping:

where kr (l ) is the wanted OFDM component. When the input signal is Gaussian, k is a constant. In the presence of impulse noise, k may vary slightly. Assuming constant k, the signal at the output of the DFT is given by Z (l ) = kR(l ) + kN g (l ) + kNi (l ) + N c (l ) (9) If k is not constant, intercarrier interference (ICI) will be caused. However even in the most extreme cases it can be shown that this ICI will be 35dB or more below the signal level. These calculations are independent of the channel response, they assume only that the OFDM samples r (l ) are Gaussian. However the relationship between the SNR and the overall BER depends on the channel coding and on the fading characteristics of the channel. A number of techniques for reducing the effect of impulse noise are considered. They vary in the way that large signal samples are identified and the non linear function that is applied. Signals can be identified as large either based on the real and

where / {} denotes the real part or imaginary part of the input signal. Amplitude nulling:

/ {s (l )} AClip AClip / {s (l )} s (l ) A / [ f {s (l )}] = Clip AClip / {s(l )} AClip

/ {s (l )} AClip 0 / {s (l )} s (l ) A / [ f {s (l )}] = Clip 0 / {s (l )} AClip


V. SIMULATION RESULTS Matlab simulations were performed to examine the effects of these four non-linearities. Both Middleton and gated Gaussian noise models were used. All of the figures below are for 64QAM constellations, with 2048 subcarriers, and 100 symbols simulated. But changing the number of subcarriers to 8192 or the constellations to 4QAM or 16QAM does not alter the results.

Real an imaginary nulling:

s (l ) f {s (l )} = 0

s (l ) AClip s (l ) AClip

In each simulation, the average power in each of the real and imaginary components of the wanted OFDM signal is unity. Similarly, the average noise (Gaussian plus impulse) power in either the real or imaginary component is 2 . The value of 2 was adjusted to give varying input signal to noise power. The clipping level, AClip , is standardised in terms of the standard deviation of the wanted OFDM signal. The output SNR was measured at the output of the DFT. As noted earlier, the nonlinearity causes a shrinking of the constellation by a factor k which depends on the non-linearity. Before calculating the output SNR the constellation was adjusted by the value of k measured over the entire length of the input signal. Figs. 3 to 6 show the results using Middleton Class A noise, while Figs. 7 to 10 are for the gated Gaussian case. For Figs. 3 to 5 and 7 to 9 the real and imaginary components were separately clipped. Fig. 3. shows the effect of varying A for a received SNR of 13dB. For small AClip , the SNR is reduced because the increase in clipping noise more than offsets the reduction in impulse noise. At clipping levels between 1 and 4, the SNR depends on the value of A. For small A, there is a maximum improvement of SNR of about 3dB. As AClip increases, the SNR for each case plateaus at the input SNR level because no samples are clipped. In Fig. 4, is varied. for a received SNR of 13dB. As the total noise power is constant, an increase in represents an increase in the impulse noise component. The graphs have the same general form as Fig. 3. Fig. 5 shows the variation with 2 the total noise power at the input. As expected, the graphs plateau at levels depending on 2 . The noise mitigation technique is more effective for the higher noise environment. This is because high levels of noise are more likely to result in instantaneous signal plus noise levels greater than AClip . The results for different non-linearities are shown in Fig. 6. All four graphs show peaking, indicating that for these parameters all four techniques mitigate the impulse noise to some extent. When the non-linearity is operating on the amplitude, the peak occurs at a larger value of AClip . The nulling techniques give much lower SNRs for low values of AClip because incorrect nulling results in much greater signal distortion. For this case amplitude clipping gives the greatest impulse noise mitigation.
Fig. 6. SNR versus AClip , for different non linearities.

Fig. 3. SNR versus AClip , for varying A.

Fig. 4. SNR versus AClip for varying .

Fig. 5. SNR versus AClip for varying 2 .

Figs. 7 to 10 show the results for the gated Gaussian noise model. The same characteristics are found. When the noise is highly impulsive, all of the non-linearities give some impulse noise mitigation if AClip is correctly chosen. VI. CONCLUSIONS The use of non linear processing at the input of an OFDM receiver to reduce the effect of impulse noise has been investigated. The effect of non linear processing has been analysed using Bussgangs theorem. Simulations have been performed for two different forms of impulse noise: Class A Middleton and gated Gaussian and several different non-linearities. The results show that for highly impulsive noise an improvement of around 3dB in SNR can be achieved. For the situations simulated, the optimum non-linearity was a clipping operation on the amplitude of the input signal samples. Clipping based on the real and imaginary components gave about 1dB less improvement, but may be easier to implement in practice. The clipping level should be set at about 2.6 for amplitude clipping and about 2.4 for real and imaginary clipping. VII. REFERENCES
[1] [2] [3] ETSI EN 300 744 Digital Video Broadcasting (DVB); Framing Structure, channel coding and modulation for digital terrestrial television. M. Ghosh, Analysis of the Effect of Impulse Noise on Multicarrier and Single Carrier QAM Systems, IEEE Trans. on Comms., Vol. 44, No. 2, February 1996, pp.145-147. T.N Zogakis, P.S. Chow, J.T Aslanis and J.M Cioffi, Impulse noise mitigation strategies for multicarrier modulation, IEEE International Conference on Communications (ICC 93),Vol. 2, May 1993, pp. 784-788 J.H Stott, Detection and removal of Clipping in Multi-Carrier Receivers, European Patent Office, EP 1043874, October 2000. M. Sliskovic, Signal Processing Algorithm for OFDM Channel with Impulse Noise, 7th IEEE Conference on Electronics, Circuits and Systems 2000 (ICECS 2000), Vol. 1, December 2000, pp 222-225. A. D. Spaulding and D. Middleton, Optimum Reception in an Impulsive Interference Environment-Part I: Coherent Detection, IEEE Transactions on Communications, Vol. Com-25, No. 9, September 1977, pp.910-923. I. Mann, S. McLaughlin, W. Henkel , R. Kirkby and T. Kessler, Impulse Generation With Appropriate Amplitude,Length,InterArrival and Spectral Characteristics, IEEE Selected Journal on Selected Areas in Communications, Vol. 20, No. 5, June 2002, pp.901-912. D. Dardari, V. Tralli and A. Vaccari, A Theoretical Characterization of Nonlinear Distortion Effects in OFDM Systems, IEEE Transactions on Communications, Vol. 48, No. 10, October 2000, pp. 1755-1764.

Fig. 7. SNR versus clipping level, for varying .

Fig. 8. SNR versus clipping level, for varying .

[4]

Fig. 9. SNR versus clipping level, for varying .

[5]

[6]

[7]

[8]

Fig. 10. SNR versus AClip , for different nonlinearities.

You might also like