Sie sind auf Seite 1von 8

Blind Signal-to-Noise Ratio Estimation Algorithm with Small Samples for Wireless Digital Communications

Dan Wu, Xuemai Gu, and Qing Guo


Communication Research Center of Harbin Institute of Technology, Harbin, China {wudan, xmgu, qguo}@hit.edu.cn

Abstract. To extend the range of blind signal-to-noise ratio (SNR) estimation and reduce complexity at the same time, a new algorithm is presented based on a signal subspace approach using the sample covariance matrix of the received signal and combined information criterion (CIC) in information theory. CIC overcomes the disadvantages of both Akaike information criterions (AIC) under penalization and minimum description lengths (MDL) over penalization and its likelihood form is deduced. The algorithm needs no prior knowledge of modulation types, baud rate or carrier frequency of the signals. Computer simulation shows that the algorithm can blindly estimate the SNR of digital modulation signals commonly used in additional white Gaussian noise (AWGN) channels and Rayleigh fading channels with small samples, and the mean estimation error is less than 1dB for SNR ranging from -15dB to 25dB. The accuracy and simplicity of this method make it more adapt to engineering applications.

1 Introduction
Signal-to-noise ratio (SNR) is defined as a measure of signal strength relative to background noise and it is one of the most important criterions for information transformation quality. In modern wireless communication systems, the precise knowledge of SNR is often required by many algorithms for their optimal performance. For example, SNR estimates are typically employed in power control, mobile assisted hand-off, adaptive modulation schemes, as well as soft decoding [1,2] procedures, etc. . Estimating SNR and providing this estimate to the data detector are essential to the successful functioning of any communications receiver. SNR estimators can be divided into two classes. One class is the data-aided estimator for which known (or pilot) data is transmitted and the SNR estimator at the receiver uses known data to estimate the SNR. The other class is the non-data-aided estimator. For this class of estimator, no known data is transmitted, and therefore the SNR estimator at the receiver has to blindly estimate the SNR. Although the dataaided estimator performs better than the non-data-aided estimator, it is not suitable for non-cooperative situations. In this paper, non-data-aided or blind SNR estimator is considered. Some methods have been proposed recently. In [3], SNR estimation in

This work is supported by National 863 Projects of China. Item number is 2004AA001210.

D.-S. Huang, K. Li, and G.W. Irwin (Eds.): ICIC 2006, LNCIS 345, pp. 741 748, 2006. Springer-Verlag Berlin Heidelberg 2006

742

D. Wu, X. Gu, and Q. Guo

frequency domain was introduced, using circular correlation for M-ary phase shift keying (MPSK), but the method is not suitable for other modulation types. Fourthorder moments method was applied in [4] for constant envelope modulations, and in [5], an envelope-based estimation method is proposed for nonconstant envelope modulations. Both of these two methods need the prior knowledge of envelope. In [6], an interative SNR estimation for negative SNRs algorithm was developed. However, the method has relatively high bias for low SNR (When SNR is below -10dB, bias is over 3dB). Blind SNR estimation can be employed in many hot fields of Information War, such as threat analysis, electronic surveillance system, etc. These applications have a high demand of estimation speed and range of SNR. However, performances of the methods mentioned above decrease when the number of samples is not big enough. Even the number of samples is appropriate, performances are not desirable when SNR is below zero. In this paper, a new blind SNR estimation algorithm is presented based on eigenvalue decomposition of the correlation matrix of the received signals and the principle of combined information criterion (CIC)[7]. Compared with using Akaike information criterion (AIC) and minimum description length (MDL), algorithm using CIC gives more accurate results in additional white Gaussian noise (AWGN) channels at low SNR with small samples. When applied to Rayleigh fading channels, the performance is also acceptable. This paper is organized as follows. After the statement and formulation of the problem in Section II, the blind SNR estimation algorithm is introduced in Section III. In Section IV the computer simulation results are presented. Section V draws the conclusions.

2 Problem Formulation
Assume y(t) is a received signal in AWGN channels. Then the signal model after sampled is
y ( k ) = s ( k ) + n( k ) .

(1)

where s(k) is a digitally modulated signal with unknown modulation type and n(k) is an independent zero-mean, Gaussian random process with second order moments
2 E n(k )n H (l ) = N I kl .

(2) (3)

E nT ( k ) n(l ) = 0 .

2 where xH denotes the Hermitian transpose of x, xT denotes the transpose of x, N is the noise power, kl represents the Kronecker delta function, and I is the identity

matrix. Let Y(k)=[y(k) y(k+1) y(k+L-1)], then


Y (k ) = S (k ) + N (k ) .

(4)

Blind SNR Estimation Algorithm with Small Samples

743

where S(k)=[s(k) s(k+1) s(k+L-1)], and N(k)=[n(k) n(k+1) n(k+L-1)]. The L order covariance matrix of the received signal is
2 2 Ryy = E (YY H ) = E (( S + N )( S + N ) H ) = E ( SS H ) + N I = Rss + N I .

(5)

where Rss is the covariance matrix of the original signal. Based on the properties of covariance, Rss is a positive semi-definite matrix and its rank is assumed q (q<L), or equivalently, the L-q smallest eigenvalues of Rss are equal to zero. Denoting the eigenvalues of Ryy by 12L, then the smallest L-q eigenvalues of Ryy are all 2 equal to n . SNR is
2 SNR = 10 log( E ( S H S ) / N ) .

(6)

From (5-6), SNR can be easily estimated if q is achieved.

3 SNR Estimation Algorithm


Rank q can be determined from the smallest eigenvalue of Ryy. However, the problem is that when estimated from a finite sample size, the resulting eigenvalues are all different with probability one, thus making it difficult to determine q by merely observing the eigenvalues. The problem of achieving q can be seen as AR model selection process. Assuming Y=[Y(1), Y(2),Y(N)], the most popular criteria for model selection is AIC and MDL
AIC = 2 log( f (Y | )) + 2w . MDL = log( f (Y | )) + 0.5w log( N ) .

(7a) (7b)

where f (Y | ) is a parameterized family of probability densities, is the maximum likelihood estimate of the parameters vector , and w is the number of free adjusted parameters in . AIC and MDL have been derived assuming that the number of samples is much larger than the number of estimated model parameters. For small samples, they have the tendency to select overly complex models. AIC tends to underpenalize whereas MDL tends to overpenalize. As a result, AIC yields a relatively large probability of overestimation of the rank q even when SNR is high while MDL underestimates the rank q when the SNR is low. To give a good tradeoff of over penalization and under penalization, CIC proposed by Broersen is employed here. CIC is a criteria for small sample model selection problems, and it combines the theoretical asymptotical preference for penalty 3 with the good finite sample characteristics of FSIC[7]. CIC for Burg in residual variance form is
1 k 1 . N + 1 i 1,3 CIC = log{res (k )} + max 1 i=0 N + 1 i i =0 1 N +1 i
k

1+

(8)

744

D. Wu, X. Gu, and Q. Guo

where k{1, 2, L} ranges over the set of possible ranks. Because likelihood estimator is needed in this paper, (8) should be transformed into likelihood form. Since AIC has the equation
AIC = 2 log( f (Y | )) + 2 w = N log(res(k )) + 2w

(9)

Then CIC in likelihood form is


2 CIC = log( f (Y | )) + max N 1 k 1 N + 1 i 1,3 1 i =0 N + 1 i i=0 1 N +1 i
k

1+

(10)

To apply (10) to determine the rank of Ryy, f (Y | ) should be described firstly. In [8],

log f (Y | ) = log

i = k +1

( pk ) N

1/( L k ) i L

1 i L k i = k +1 L . w can be estimated by

(11)

where i is the estimate of i, and satisfies 1 2


w = k (2 L k )

(12)

From (11) and (12), CIC can be rewritten as

2 CIC (k ) = log N

i = k +1

i1/( Lk )

( p k ) N

L 1 i L k i = k +1

1 k . 1 N + 1 i 1, 3 + max 1 N +1 i i =0 i =0 1 N +1 i
k

1+

(13)

Then q is estimated by
q = arg min CIC (k ) k=1, 2, L .
k

(14)

From (13) and (14), the steps of blind SNR estimation proposed here are as follows: (1) Make an eigenvalue decomposition of the sample covariance matrix Ryy ,
Ryy = 1 N
N i =1

Y (i )Y (i ) H = U

UH .

(15)

where

= diag (i ) , and U consists of the orthonormal eigenvectors of Ryy .

(2) Estimate the rank q, using (13) and (14). (3) Estimate the noise power according to
2 N = L 1 i . L q i = q +1

(16)

Blind SNR Estimation Algorithm with Small Samples

745

and the signal power according to

s2 = (
(4)

q i =1

2 i ) q N .

(17)

The estimate of the received signal SNR is obtained as


SNR = 10 log
2 S . 2 L N

(18)

4 Simulations Results
In this section, the performance of the proposed algorithm is investigated in MATLAB environment. Computer simulations are performed in Monte Carlo way and 200 experiments are used. Seven digital modulation types are selected for testing the estimation performance of the algorithm, i.e. BPSK, 4PSK, BFSK, 4FSK, TFM, /4QPSK and 16QAM. The receiver for SNR estimation is a blind receiver, that is, it has no knowledge of carrier frequency, modulation type and baud rate. First it should catch the locations of blind signals through frequency scanning, and then sample signals with the sample frequency 4 times of carrier frequency. The dimension of covariance matrix is set to L=40 and the range of SNR is -15dB to 25dB. The standard deviation can exhibit the performance more directly. In this simulation, the standard deviation (SD) is defined as SD = sqrt ( E ( SNR SNR ) 2 ) .

4.0 3.8 3.6 3.4 3.2 3.0 2.8 2.6 2.4 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 -15

The SD of the SNR estimators

CIC AIC MDL N=400

-10

-5

5 SNR(dB)

10

15

20

25

Fig. 1. The SD of SNR estimators in AWGN channel for BPSK (N=400)

746

D. Wu, X. Gu, and Q. Guo

4.0 3.8 3.6 3.4 3.2 3.0 2.8 2.6 2.4 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 -15

The SD of the SNR estimators

CIC AIC MDL N=600

-10

-5

5
SNR(dB)

10

15

20

25

Fig. 2. The SD comparison of SNR estimators in AWGN channel for BPSK (N=600)

3.0 2.8 2.6 2.4 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 -15

The SD of the SNR estimator

BPSK 4PSK BFSK 4FSK TFM PI/4QPSK 16QAM N=600

-10

-5

5 SNR(dB)

10

15

20

25

Fig. 3. The SD of the proposed SNR estimator in AWGN channel (N=600)

4.1 AWGN Channel The robustness of the blind estimation algorithm in AWGN channels is tested in this subsection. Comparisons are made between algorithms using CIC, AIC and MDL for BPSK. Fig.1-2 show the results of the SD of the estimators using CIC, AIC and MDL

Blind SNR Estimation Algorithm with Small Samples

747

and sample numbers are N=400 and N=600 respectively. In Fig.1, the SNR estimator using CIC has better performance than that using AIC and MDL, especially at the low SNR range. The average SD for CIC is 0.55dB, for AIC is 0.73dB, and for MDL is 0.84dB. In Fig. 2, the SD decreases with the increasing of the samples. The average SD for CIC is 0.43dB, for AIC is 0.51dB, and for MDL is 0.64dB. Fig. 3 illustrates the SNR estimation results for BPSK, 4PSK, BFSK, 4FSK, TFM, /4QPSK and 16QAM. It is seen that the proposed algorithm is not affected by modulation types. It is suitable for both constant envelope modulations and nonconstant envelope ones. The average SD is 0.38dB.

5.0 4.5 The SD of the SNR esitmator 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 -15 -10 -5 0 5 SNR(dB)
Fig. 4. The SD of the proposed SNR estimator in Rayleigh fading channel (N=600)

BPSK 4PSK BFSK 4FSK TFM PI/4QPSK 16QAM N=600 fdT=0.01

10

15

20

25

4.2 Rayleigh Fading Channel Although the proposed algorithm is deduced in AWGN channels, simulation results show that it can be applied in Rayleigh fading channel situations. Fig.4 illustrates the SNR estimation results for BPSK, 4PSK, BFSK, 4FSK, TFM and /4QPSK in Rayleigh fading channels when fdT is 0.01. It is seen that the SD of the SNR estimator is below 1dB when SNR is not lower than -5dB. Although the performance goes down in fading channels compared with that in AWGN channels, the results still have reference value and instructive significance.

5 Conclusion
In this paper, a new blind SNR estimation algorithm is proposed based on signal subspace approach and CIC. Computer simulation shows that the algorithm can

748

D. Wu, X. Gu, and Q. Guo

achieve accurate SNR estimations blindly with small samples at SNR range from 15dB to 25dB in AWGN channels and when applied in Rayleigh fading channels, it still has a good performance at SNR range from -5dB to 25dB. It exhibits a lower bias and better simplicity and more adapts to non-cooperative situations.

Acknowledgments
The authors are grateful to Piet M. T. Broerson of the Department of Applied Physics of the Delft University and Fjo De Ridder of the department ELEC of the Vrije Universiteit Brussel (VUB), for their patient explaining by email.

References
1. Balachandran, K., Kadaba, S.R., Nanda, S.: Channel Quality Estimation and Rate Adption for Cellular Mobile Radio. IEEE Journal on Selected Areas in Communications. (1999) 1244-1256 2. Summers, T. A., Wilson, S.G.: SNR Mismatch and Online Estimation in Terbo Decoding. IEEE Transactions on Communications. (1998) 421-423 3. Dea-Ki Hong, Cheol-Hee Park, Min-Chul Ju. :SNR Estimation in Frequency Domain Using Circular Correlation. Electronics Letters. (2002)1693-1694 4. Matzner, R., Englberger, F.: An SNR Estimation Algorithm using Fourth-order Moments. Proceedings of the 1994 IEEE Symposium on Information Theory. (1994) 119 5. Ping G., Cihan T.: SNR Estimation for Nonconstant Modulus Constellations. IEEE Transactions on Signal Processing. (2005) 865-871 6. Bin, L., Robert, D., Ariela, Z.: A Low Bias Algorithm to Estimate Negative SNRs in an AWGN Channel. IEEE Communications Letters. (2002) 469-472 7. Piet M.T. Broersen.: Finite Sample Criteria for Aautoregressive Order Selection. IEEE Transactions on Signal Procesing. (2000) 3550-3559 8. Mati Wax, Thomas Kailath.: Detection of Signals by Information Theoretic Criteria. IEEE Transaactions on Aacoustics , Speech, and Signal Processing. (1985) 387-392

Das könnte Ihnen auch gefallen