International Workshop on Cognitive and Cooperative Wireless Communication Systems 2012

ADAPTIVE MULTICHANNEL SEQUENTIAL LATTICE PREDICTION FILTERING METHOD FOR COGNITIVE RADIO SPECTRUM SENSING IN SUBBANDS

Mehmet Tahir Ozden, Senior Member, IEEE Piri Reis University, Tuzla, Istanbul, 34940 Turkey. mtozden@pirireis.edu.tr
is, it can reduce variance of the spectral estimate without compromising the bias of the estimate. Additionally, the multitaper method outperformed the multiple signal classication (MUSIC) algorithm and the modied forward-backward linear prediction (MFBLP) algorithms in so far as the continuous parts of the spectrum are concerned, but did not perform well in dealing with the line spectral components. Filter bank methods were compared against the lter bank reinterpretation of the multitaper method with applications to spectrum sensing by Farhang-Boroujeny in [4], and it was noted that, if a lter bank multicarrier technique is exploited as the physical layer of a CR network, then the same lter bank can also be used for channel sensing at virtually no cost. On the contrary to nonparametric methods, parametric methods for spectrum sensing applications have not attracted much attention even though they deserve consideration particularly when they are implemented in subbands due to Rao and Pearlmans ndings presented in [6], which showed that the well-known autoregressive(AR) modeling was a promising method for spectrum estimation in subbands, and that pthorder prediction from subbands was superior to pth-order prediction in the fullband when p was nite, and subband decomposition of a source resulted in a whitening of the composite subband spectrum. The equivalence of linear prediction and AR spectrum estimation was then exploited by Rao and Pearlman [6] to show that AR spectrum from subbands offers a gain over fullband AR spectrum estimation due to SNR and frequency spacing improvement as well as whitening effects. Furthermore, processes that are purely autoregressive are often transformed into ARMA processes by addition of measurement noise, and especially sinusoids in noise are known to obey the degenerate ARMA equation in fullband as well as in subbands [7]. Consequently, we think that developing new methods of subband spectrum estimation based on ARMA modeling is particularly relevant for spectrum sensing of multicarrier cognitive radio signals [8]. The concept of software radio on the otherhand, as CRs backbone, relies on the development of digital signal processing (DSP) technology, that is exible, recongurable, and reprogrammable by software to adapt to an environment where there are multiple services, standards, and frequency bands [9]. As such, the infrastructure in a software radio system is generally required to use recongurable VLSI hardware components such as DSP chip sets, eld programmable gate arrays (FPGAs), embedded processors, and even general purpose processors [10][12]. Under these considerations, we propose a new subband

AbstractA multichannel characterization for autoregressive moving average(ARMA) spectrum estimation in subbands with applications to cognitive radio spectrum sensing is considered in this presentation. The fullband ARMA spectrum estimation can be realized in two-channels as a special form of this characterization. A complete orthogonalization of input multichannel data is accomplished using a modied form of sequential processing multichannel lattice stages. Matrix operations are avoided, only scalar operations are used, and a multichannel ARMA prediction lter with a highly modular and suitable structure for VLSI implementations is achieved. The lter is therefore considered as a good candidate for FPGA and DSP chip based spectrum sensing functions in software dened radio implementations. Lattice reection coefcients for autoregressive and moving average parts are simultaneously computed. These coefcients are then converted to process parameters using a newly developed Levinson-Durbin type multichannel conversion algorithm, so that they can be utilized in the classication of the spectrum. Coarse and ne spectrum sensing can be accomplished using different prediction lter congurations. The computational complexity is given in terms of model order parameters. The performance of the proposed method is compared visually and statistically to that of the multitaper method.

I. I NTRODUCTION An important recent development in the design of wireless communication systems is the cognitive radio (CR), built on a software dened radio, functions as an intelligent system that is aware of its environment and uses the methodology of understanding-by-building to learn from the environment and adapt to statistical variations in the input stimuli in order to establish reliable communication by efcient utilization of the radio spectrum [1]. In CR systems, secondary users must have the capability to detect and opportunistically exploit under utilized spectral resources allocated to primary users. As a consequence of this fact, spectrum sensing functionality has emerged as an integral part of CR systems. Different approaches have been proposed to perform spectrum sensing function in CR systems such as matched ltering, the use of special signal features to detect and classify primary signal, energy detection, and spectral analysis by lter bank and multitaper methods [2][5]. Spectral analysis methods are preferable if the spectral properties of the signal to be detected are known, but the signal has otherwise no usable features that can be efciently exploited. Accordingly, the desirable attributes of spectral analysis methods with application to spectrum sensing are their estimation accuracies and resolution characteristics. It has been reported in [1] that the multitaper method can produce an accurate estimate of the desired power spectrum by trading spectral resolution for improved spectral characteristics, that

978-1-4673-2015-3/12/$31.00 2012 IEEE

961

spectrum estimation method based on ARMA modeling, which can also be viewed as a parametric lter bank method, and can meet the requirements for spectrum sensing in cognitive radios because of its highly modular, regular, and order recursive prediction lter structure, its suitability for recongurable VLSI implementations, and its spectrum estimation accuracy and resolution characteristics that improves with increasing number of subbands. The proposed method relies on estimation of the driving noise in subbands, such that the subbands ARMA ltering problem is at rst transformed into multichannel AR ltering problem by embedding subband ARMA processes into multichannel AR processes, and then a complete modied Gram-Schmidt orthogonalization of multichannel input signal is accomplished using a modied version of the sequential processing multichannel lattice stages (SPMLSs) [13]. Thus, the complete orthogonalization of multichannel input data and sequential nature of the modied SPMLSs make it possible to feed back the delayed forward prediction error signals to represent the unknown input noise signals of original ARMA processes, and lends itself to a novel ARMA prediction lter structure, which to the best of our knowledge does not exist in the literature and can be recongured as an equalizer during signal reception [14], and to the development of a new Levinson-Durbin type conversion algorithm for the modied SPMLSs in order to compute ARMA process parameters from lattice reection coefcients. A two-subband ARMA spectrum estimation problem is considered due to the ease of explanation and space limitations in developing the method. However, it is considered straightforward to apply the method to any number of subbands, and to AR spectrum estimation in subbands. The method is also appropriate for uniform as well as non-uniform lter bank realizations. The paper is organized as follows. In Section 2, we present the development of the new multichannel ARMA lattice prediction lter using the modied SPMLSs. In Section 3, we develop the new Levinson-Durbin type multichannel conversion algorithm for the modied SPMLS, and relate lattice parameters to process parameters. Spectrum estimation expression in two subbands is given in Section 4. The computational complexity computations are treated in Section 5. Section 6 is concerned with the experimental results. Finally, Section 7 is about the discussions of results and conclusions. The following notations are used in this presentation. () represents the complex conjugate of (). ()T and ()H stand for the transpose and the Hermitian transpose of () respectively. The variables m, i and n are global while all other variables are local. The variable m represents the stage number while n and i are the time indexes related to data and coefcients respectively till we equate them in Section 3 to have a single time index.

II. ADAPTIVE MULTICHANNEL ARMA LATTICE PREDICTION FILTERING A. Multichannel Prediction Problem An illustration of the adaptive multichannel ARMA prediction ltering in subbands for two-subbands case is presented in gure 1. Therein, y (n) represents the input fullband signal while y1 (n) and y2 (n) stand for the input subband signals. In adaptive multichannel ARMA prediction ltering, the objective is to nd an exponentially windowed, LS solution for the AR and MA coefcients of the k th forward prediction lter that minimizes each of the two cost functions,
i

J (i) =
n=0

k in | fp (n) |2 k

(1)

at each time instant i, and k = 1, 2. The forward prediction k error fp (n) in this expression is dened as, k
k k (n), (n) = dk (n) d fp i k

(2)

k (n), is an and the k th forward prediction lter output, d i k estimate of the k th desired signal, d (n) = yk (n), is given by
pk qk a k 1,j (i)yk (n j ) + j =1 l=0 a k uk (n l). 2,l (i)

k (n) = d i

(3)

Herein, pk and qk denote the order of the (pk , qk ) prediction error lter associated with the k th subband, and u k (n) is the estimate of the k th ARMA process input signal. The estimated k th ARMA process input signal, u k (n), is obtained by delaying and feeding back the pk th order forward prediction k error, u k (n) = fp (n 1). Hence, the input vector to the k th k ARMA lter at time instant n, y k (n), and the corresponding coefcient vector ak (i), at time instant i, are dened as y k (n) = [yk (n 1), . . . , yk (n pk ), u k (n), u k (n 1), ...,u k (n qk )]T and kT (i) = [ a ak k k k k 1,1 (i), . . . , a 1,pk (i), a 2,0 (i), a 2,1 (i), . . . , a 2,qk (i)] (5) respectively. Herein, a k k 1,j (i) and a 2,j (i) respectively represent the j th coefcient related to the AR and MA parts of the forward prediction lter for the k th subband at time instant i. It is assumed, without loss of generality, that pk qk . pk = qk case corresponds to the prediction lter for an ARMA(pk , pk ) process, while pk > qk prediction lter is for a general ARMA(pk , qk ) process. Note that an ARMA backward prediction can be performed for the desired signal, dk (n) = yk (n pk ), and the prediction lter in that case would use the reversed and conjugated forward prediction lter coefcients, which are dened in the backward prediction error coefcient vector as ckT (i) = [ ck k k k k 1,pk (i), . . . , c 1,1 (i), c 2,qk (i), . . . , c 2,1 (i), c 2,0 (i)] (6) (4)

962

where c k k 1,j (i) and c 2,j (i) are respectively dened as the j th coefcient related to the AR and MA parts of the backward prediction lter for the k th subband at time instant i. Consequently, the main concern of the exponentially weighted LS problem under consideration is to nd, at each time i, the k th optimal coefcient vector, ak (i) that would minimize the cost function,
i

and input to two-channel stages for which the stage number (m) has a range of values given by 0 < m (p1 q1 ). Accordingly, we redene equations (9) and (10) using this new data vector as follows,
i

(i) =
n=0

in y
i

+1 (n)

y H +1 (n)

(12)

and in | dk (n) akH (i)y k (n) |2 . (7)

,k (i) =

J k (i) =
n=0

in y
n=0

+1 (n)

dk (n)

(13)

The k th optimal coefcient vector related to the k th subband lter, 1 ak (8) opt (i) = Rk (i)Pk (i) is found by differentiating J (i) with respect to a (i), setting the derivative to zero, and solving for ak (i), where
i k k

where k = 1, 2. The orthogonalization of data using SPMLSs corresponds to the transformation of (12) and (13) into
i

Df+1 (i) =
n=0

in f (i) y

+1 (n)

fH y H (i) (14) +1 (n)

and
in H y k (n)y k (n)

Rk (i) =
n=0

(9)

Zf+1,k (i) =
n=0

in f (i 1) y

+1 (n

1) dk (n) (15)

and Pk (i) =

in y k (n) dk (n).
n=0

(10)

B. Sequential Lattice Orthogonalization In order to nd a modular, regular, and simple solution to the two subbands ARMA prediction problem, we would like to use a single multichannel lattice lter as depicted in gure 2, instead of using two separate transversal lters and solving two separate optimization problems as in gure 1. We would also like to avoid direct evaluations as in (8), and achieve good numerical properties. As the number of channels at different sections of the proposed multichannel lattice lter is different due to the sequential processing nature of SPMLSs, we carry out the exponentially weighted LS optimization problem by taking into consideration each of these sections separately, and therefore we assume that the lter is comprised of three cascaded lters, which are two-channel, three-channel, and four-channel lattice sections ; and we use a different index for each section while using m to indicate a stage in the whole lter. We also assume p1 = p2 for the ease of explanation without loss of generality. In order to sequentially solve the exponentially weighted LS optimization problem under consideration, we rst organize the elements of input signal vectors y1 (n) = [y1 (n), . . . , y1 (n )]T , and y2 (n) = [y2 (n), . . . , y2 (n )]T according to the natural ordering of SPMLSs as : y1 (n) y2 (n) y1 (n 1) (11) y +1 (n) = y2 (n 1) y1 (n ) y2 (n )

respectively. Here, f (i) is the 2 2 lower triangular transformation matrix for forward prediction, and is sequentially realized stage-by-stage using 2 2 lower triangular transformation matrices, Lf (i) = 1
f

(i 1) 1

(16)

whose diagonal elements are all equal to unity at time instant i, and f (i) is the reection coefcient computed at the single circular cell in the triangular shaped self-orthogonalization processor of the th two-channel SPMLS. Then, the forward lattice predictor coefcients are computed using
f f f,k (i) = D +1 (i 1) Z +1,k (i)

(17)

where f,k (i) represents the k th row of the 2 2 lattice forward prediction reection coefcient matrix f (i), and is also sequentially implemented stage-by-stage by means of 2 2 forward prediction reection coefcient matrices, f (i) = f,1,1 (i) f,1,2 (i) f,2,1 (i) f,2,2 (i) (18)

in which f,k,j (i) is the j th reection coefcient related to the forward prediction of the k th channel signal, and it is computed at the (k, j )th single circular cell of the square shaped reference-orthogonalization processor related to forward prediction at the th two-channel SPMLS. Note that the matrix inversion operation in equation (8) is now transformed into a simple scalar inversion operation in (17) due to the diagonal nature of Df+1 (i). The backward prediction counterpart of this optimization problem is similarly solved using 2 2 lower triangular transformation matrices, Lb (i), and 2 2 lattice backward prediction reection coefcient matrices, b (i). After the processing of input signals by two-channel lattice stages, the delayed and fed back forward prediction error

963

and input to three-channel lattice section, where the stage number (m) takes values in the range given by (p1 q1 ) < m (p2 q2 ). Subsequently, we solve the optimization problem in (17) once again with the new input vector, in which case f (i) and f (i) are the 3 3 lower triangular transformation and the 3 3 forward lattice prediction coefcient matrices respectively. f (i) is computed sequentially by means of 3 3 f lower triangular transformation matrices, Lf (i), and (i) is similarly realized stage-by-stage making use of 3 3 forward prediction coefcient matrices, f (i), at time instant i. Note that, since the delayed and fed back signal is considered to constitute a new channel in the multichannel sequential lattice ltering, we have three desired signals at this point, dk (n), where k = 1, 2, 3, one of which did not exist as a separate channel in the optimization problem stated in Subsection II.A, and this new desired signal, d3 (n), is related to the MA part of the rst subband ARMA modeling. Finally, the optimization problem is expanded one more time with the inclusion of the second delayed and fed back forward prediction error u 2 (n) = fp2 (n 1), and this time, the elements of input data vectors y1 (n) = [y1 (n), . . . , y1 (n )]T , y2 (n) = [y2 (n), . . . , y2 (n )]T , u 1 (n) = [ u1 (n), . . . , u 1 (n )]T , and u 2 (n) = [ u2 (n), . . . , u 2 (n )]T are organized similar to (11) and (19) with the stage number (m) is in the range given by (p2 q2 ) < m p2 due to four channel processing. Similar to two-channel and three-channel cases, we solve the optimization problem in (17) using the new data f vector, in which case f (i) and (i) are 4 4 lower triangular transformation, and 4 4 forward lattice prediction coefcient matrices at the time instant i respectively. As in previous sections, these matrices are computed stage-by-stage by use of 4 4 lower triangular transformation matrices, Lf (i), and 4 4 forward prediction coefcient matrices, f (i), at time instant i respectively. Since the second delayed and fed back signal is also considered as a new channel in the multichannel sequential lattice ltering, hereafter we have four desired signals, dk (n) ,where k = 1, 2, 3, 4, and this fourth desired signal, d4 (n), is associated with the MA part of the second subband ARMA modeling. Consequently, we present a diagram of the four channel lattice prediction lter structure under consideration in gure 3.

u 1 (n) = fp1 (n 1) is incorporated at the (p1 q1 + 1)th stage, as the third channel. Accordingly, we expand the optimization problem by organizing the elements of the input data vectors y1 (n) = [y1 (n), . . . , y1 (n )]T , y2 (n) = [y2 (n), . . . , y2 (n )]T , and u 1 (n) = [ u1 (n), . . . , u 1 (n )]T as follows : y1 (n) y2 (n) u 1 (n) (19) y +1 (n) = , y1 (n ) y2 (n ) u 1 (n )

III. CONVERSION OF LATTICE COEFFICIENTS TO PROCESS PARAMETERS Since the mathematical link between process parameters and reection coefcients of a lattice prediction lter is provided by the Levinson-Durbin algorithm [15], we develop a new Levinson-Durbin type conversion algorithm specically for SPMLSs in order to convert lattice reection coefcients to ARMA process parameters. Due to the sequential nature of the proposed lattice structure, we carry out the development of the new Levinson-Durbin type multichannel conversion algorithm by taking into consideration each of these sections separately, and therefore we assume that the lter is comprised of three cascaded lters as discussed in Subsection II.B. We rst consider the conversion algorithm for the twochannel section of lattice prediction lter, and we organize the input signal samples to two-channel section as y1 (n) y1 (n) y1 (n 1) y1 (n ) y +1 (n) = = y2 (n) y2 (n) y2 (n 1) y2 (n ) (20) where we dene the data vectors as y1 (n) = [y1 (n), . . . , y1 (n + 1)]T , y2 (n) = [y2 (n), . . . , y2 (n + 1)]T , and 0 < m (p1 q1 ). The corresponding forward and backward prediction error coefcient matrices for the th order transversal lter for the k th channel are dened as (i)], (21) kT k k k k k k c (i) = [ c2 (i), c 2 1 (i), c 2 2 (i), . . . , c 2 (i), c 1 (i), c 0 (i)] (22) where k = 1, 2 due to two-channel processing, and a k 0 (i) = c k 0 (i) = 1.0. Since the signal time shifting and ordering properties of SPMLSs when expressed in matrix form as in equation (11) are different than the organization of input signal samples in matrix form for transversal ltering as shown in equation (20), we use (2 + 1) (2 + 1) shufing matrices, J1+1 for the rst channel and J2+1 for the second channel, to reorder the elements of coefcient matrices, a1H (i), a2H (i) and c1H (i), c2H (i), according to the sample ordering of SPMLSs. Therefore, the forward and backward prediction errors for the end of the observation interval n = i at the output of the general th order lters with transversal structure can be stated as : f 1 (n) f 2 (n) b1 (n) b2 (n) = J1+1 a1H (n) 0 c1H (n) J1+1 0 0 a2H (n) J2+1 0 c2H (n) J2+1 y
+1 (n)

akT (i) = [ ak k k k 0 (i), a 1 (i), a 2 (i), . . . , a 2

k k 2 (i), a 2 1 (i), a 2

(23) = y
+1 (n)

(24)

964

where 0 is a 1 ( + 1) zero matrix. We then express the ( 1)th prediction errors for transversal prediction ltering. Note that the size of each coefcient matrix increases by two when the order of prediction lter increases from ( 1) to , and 0 is a 1 ( + 1) zero matrix. Subsequently, we also dene the th order lattice prediction errors in terms of lattice parameters and the ( 1)th order forward and backward prediction errors. The lattice parameters appear in the lower coefcient triangular and square matrices, which are generated in triangular shaped self-orthogonalization, and square shaped reference-orthogonalization processors in twochannel SPMLSs, as previously dened in equations (16) and (18). Accordingly, the th and the ( 1)th order prediction error matrices for transversal prediction lter, are substituted in the th order lattice prediction error expressions so as to obtain the pairs of the Levinson-Durbin recursions for two-channel section. Note that we can not provide a complete treatment of the aforementioned matrix operations due to space limitations. The three-channel section starts with the incorporation of the third channel (u 1 (n)) as the new channel at the (p1 q1 + 1)th stage. The Levinson-Durbin algorithm for this section is developed similar to two-channel section by assuming that three-channel section is a separate lter, and thereby considering the input signal samples to the threechannel section as follows: y1 (n) y1 (n) y1 (n 1) ( n ) y 1 y ( n ) 2 y2 (n) y +1 (n) = = ( n 1) y 2 y2 (n ) u ( n ) u 1 ( n ) 1 u 1 (n 1) u 1 (n ) (25) where y1 (n) = [y1 (n), . . . , y1 (n + 1)]T , y2 (n) = [y2 (n), . . . , y2 (n + 1)]T , and u 1 (n) = [ u1 (n), . . . , u 1 (n + 1)]T . Correspondingly, the forward and backward prediction error coefcient matrices for the th order transversal ltering are dened as kT (n) = [ a ak k k k 0 (n), a 1 (n), a 2 (n), a 3 (n), . . . , k k ...,a k 32 (n), a 31 (n), a 3 (n)], ckT (n) = [ ck k k 3 (n), c 31 (n), c 32 (n), . . . , k k k ...,c k 3 (n), c 2 (n), c 1 (n), c 0 (n)] where k = 1, 2, 3 due to three-channel processing. Then, the prediction ltering continues with three-channel lattice stages for (p1 q1 ) < m (p2 q2 ). The Levinson-Durbin recursions for three-channel section can be developed similar to two-channel section by establishing the mathematical link (27) (26)

between transversal and lattice lter coefcients, that is, by substituting ( 1)th order error expressions for transversal ltering into the th order prediction error expressions for lattice ltering. Since the organization of signal samples in equation (25) is different than the ordering of signal samples entering into three-channel SPMLSs in (19), we use (3 +1) (3 +1) 2 shufing matrices, J1 +1 for the rst channel, J+1 for the 3 second channel, and J+1 for the third channel to reorder the H H H elements of coefcient matrices, a1 a2 a3 (n), (n), (n) and 1H 2H 3H c (n), c (n), c (n), according to the sample ordering of SPMLSs. The Levison-Durbin Recursions for three-channel section are subsequently produced similar to two-channel section. Finally, the fourth channel( u 2 (n)), which represents the fed back and delayed signal related to the second subband, is taken into the orthogonalization process at the (p2 q2 + 1)th stage, and the prediction ltering continues with four-channel lattice stages through (p2 q2 ) < m p2 . In order to develop the Levinson-Durbin recursions for this section, we dene the forward and backward prediction error coefcient matrices for the th order transversal ltering similar to the previous sections, and we also organize of the elements of input vectors y1 (n) = [y1 (n), . . . , y1 (n + 1)]T , y2 (n) = [y2 (n), . . . , y2 (n + 1)]T , u 1 (n) = [ u1 (n), . . . , u 1 (n + 1)]T , and u 2 (n) = [ u2 (n), . . . , u 2 (n + 1)]T column wise as we did in equations (20) and (25) for two-channel and three-channel sections respectively. Furthermore, the signal sample ordering for four channel transversal ltering is different than the ordering for lattice ltering, hence we use (4 +1)(4 +1) shufing matrices, J1 +1 for the rst channel, 3 J2 for the second channel, J +1 +1 for the third channel, and J4 +1 for the fourth channel to reorder the elements H H H H of coefcient matrices a1 a2 a3 a4 (n), (n), (n), (n) and 1H 2H 3H 4H c (n), c (n), c (n), c (n), according to the sample ordering of SPMLSs. The development of the Levinson-Durbin recursions for this section unfolds as in two and three channel sections. Firstly, the th and the ( 1)th order forward and backward prediction errors are stated as outputs of transversal lters. Secondly, the prediction order update equations for the the th order four-channel lattice section are given in terms of ( 1)th prediction errors and lattice parameters, and nally the th and the ( 1)th order forward and backward transversal lter prediction error expressions are substituted in the lattice prediction order update equations, such that the pairs of Levinson-Durbin order updates for fourchannel section are obtained. Eventually, the Levinson-Durbin recursions for two-channel, three-channel, and four-channel sections are concatenated to attain the complete algorithm for the proposed prediction lter. IV. SPECTRUM ESTIMATION FROM SUBBANDS By computing process parameters from lattice coefcients, we have established the link between multichannel prediction and spectrum estimation. Hence, the estimated spectrum in subbands can now be expressed in terms of the subband

965

prediction lter parameters as Sy (wk ) =

jwk jq1 wk 1+a 3 + ... + a 3 q1 e 1e 1 jw 1 k + ... + a 1+a 1 e p1 e jp1 wk 2 2 . x 1 2 2 . x 2

(28)

jwk jq2 wk 1+a 4 + ... + a 4 1e q2 e + jwk + ... + a jp2 wk 1+a 2 2 p2 e 1e

2 where x represent the prediction error variance for the the k k th subband ; and the coefcients, a 1 3 3 1 1, . . . , a q1 1, . . . , a p1 and a are related to the AR and MA parts of the rst subband ARMA spectrum while the coefcients a 2 2 4 4 1, . . . , a p2 and a 1, . . . , a q2 are associated with AR and MA parts of the second subband ARMA spectrum. Note that the coefcients related to the rst and second subbands in equation (28) are selected from the elements of coefcient vectors in equations (21), (26), and (??) using a coefcient selection rule, which we can not provide herein due to space limitations.

We used a data length of N = 128, and data was zeropadded to 32 times the data length in all simulations. The forgetting factor was = 0.995. We repeated simulation experiments one hundred times ; the results of these simulations are then ensemble-averaged. In subband decomposition of input signals, we used the Kaiser window based approach in [16] for designing cosine modulated lter banks. In two and four subband decompositions, the lter lengths are thirty and sixty respectively. In simulations involving the multitaper method, the number of tapers (K ) was chosen as K = 3 for both visual and statistical analysis plots. The case of two closely spaced cisoids embedded in white as well as colored case was considered in the simulations. The time series for cisoids in white noise are given by y (n) = A1 ej 2f1 n+1 + A2 ej 2f2 n+2 + u(n), 1 n N, (30)

V. COMPUTATIONAL COMPLEXITY The number of operations required for the two-channel ARMA lattice prediction lter for fullband spectrum estimation is calculated as 10p + 16q using the number of operations required for one-channel and two-channel sequential processing lattice stages [13], where one operation is considered as one multiplication(division) and one addition. The LevinsonDurbin recursion for one-channel lattice sections requires (p q )(p q + 1) operations, and 4q (q + 1) operations for two-channel lattice sections to compute the ARMA process parameters. Therefore, the number of operations required becomes p2 + 11p 2pq + 5q 2 + 19q , and then this expression can be extended to the total number of required operations for an M subband, multichannel implementation as
M

where u(n) is a white Gaussian complex time series with uncorrelated real and imaginary parts, each with a variance 2 u and zero mean, such that the SNR for the k th cisoid is dened as | Ak |2 SN Rk = 10 log10 ( ). (31) 2 2u The time series generated by two cisoids embedded in colored noise are similarly expressed by passing the white Gaussian complex noise u(n), with zero mean and unit variance, through a second order AR coloring lter given by v (n) = 0.1v (n 1) 0.3v (n 2) + u(n), 1 n N, (32) and the local SNR in this case for the k th cisoid is dened as | Ak |2 SN Rk = 10 log10 ( ) (33) V (ejwk ) where V (ejwk ) denotes the spectral density function of the AR process at the frequency of the k th cisoid. While the cisoidal frequencies were f1 = 0.05 and f2 = 0.0656 in the white noise case, they were f1 = 0.2106 and f2 = 0.2262 in the colored noise case. The initial phases 1 and 2 were zero in both cases. A. Visual Results We rst consider the frequency spacing improvement simulations in which case the input time series were generated using equation (30) with no noise. Hence, the individual SNRs are innite so that the only improvement can be due to frequency spacing. In gure 4, we show that two subband lattice implementation improves the frequency spacing between closely spaced cisoids, and that this improvement is even furthered by four subband lattice when compared to fullband lattice implementation. In the SNR improvement simulations, we again used the input time series generated by (30) with lower SNRs, SN R1 = SN R2 = 3 dB, so that the SNR improvement can be better displayed. Figure 5 illustrates that two subband lattice implementation is able to resolve cisoids better than fullband lattice implementation, and that four subband implementation

Total Complexity =
k=1

2 p2 k +11pk 2pk qk +5qk +19qk . (29)

This computational complexity can be compared to that of the multitaper method using a FFT algorithm, approximately given by (2K + 2)N log2 N , where K is the number of tapers used in the computation of power spectral estimate, and K = 2N W 1, where N W is the time-bandwidth product. N is the number of signal samples and is a power of 2. VI. EXPERIMENTAL RESULTS We focused on spectral estimation of cisoids in noise during simulation experiments due to its relevance in multicarrier cognitive radio applications. In order to visually demonstrate the frequency spacing, whitening, and SNR improvements achieved by subbands implementation, we present and compare the subband spectrum estimation results with the fullband results. We then present the statistical analysis of our method by performing variance and mean bias analysis in estimating the frequencies of cisoids embedded in white noise. In both cases, we compare the performance against that of the multitaper method [5].

966

can resolve them much better due to SNR improvement effect of subband ltering. The next property to be considered is the whitening property of subband ltering. In this experiment, we once again generated the input time series using equation (30), however the white noise u(n) was replaced with the colored noise v (n) in equation (32). The local SNRs for the cisoids were SN R1 = SN R2 = 0 dB in this case. In gure 6, we compare the estimated spectrums based on fullband, two and four subband ARMA lattice ltering. Evidently, the cisoids are better resolved in two and four subband spectrums than fullband spectrum. Also, they are more clearly resolvable in four subbands than in two subbands. Note that in all of the three visual plots, the spectral estimates with the multitaper method have deeper nulls, wider main lobes, and higher sidelobes comparing to the four subband implementation of the proposed method.

VII. C ONCLUSIONS A novel lattice method for adaptive ARMA spectrum estimation in subbands with applications to spectrum sensing in cognitive radio systems has been presented. The proposed method relies on a particular lattice lter structure, which is suitable for reprogrammable VLSI implementations in software dened radios, and can be recongured as an equalizer during signal reception. In order to convert lattice reection coefcients to process parameters, the method includes a new Levinson-Durbin type multichannel conversion algorithm specially developed for SPMLSs. These parameters can therefore be utilized in the classication of the spectrum. Coarse and ne spectrum sensing can be accomplished by using different prediction lter congurations. The computational complexity of the proposed lattice method for ARM A(p, p) spectrum estimation compared to the multitaper method (N = 128,K=3) is low as long as lter order (p) is smaller than 39 in fullband, 27 in two subbands, and 18 in four subbands. The simulation results demonstrate that the performance of the proposed method improve with increasing number of subbands, and that it provides more accurate frequency estimates than the multitaper method particularly in the low SNR region. R EFERENCES
[1] S. Haykin, Cognitive radio : brain-empowered wireless communications, IEEE J. Select. Areas Commun., vol. 23, pp. 201-220, February 2005. [2] T. Yucek and H. Arslan, A survey of spectrum sensing algorithms for cognitive radio applications IEEE Commun. Surv. Tutor., vol. 11, no. 1, pp. 116-130, Mar. 2009. [3] E. Axell, G. Leus, E.G. Larsson, and H.V. Poor, Spectrum Sensing for Cognitive Radio, IEEE Signal Process. Mag., vol. 29, no. 3, pp. 101116, 2012. [4] B. Farhang-Boroujeny, Filter bank spectrum sensing for cognitive radios, IEEE Trans. Signal Process., vol. 56, no. 5, pp. 1801-1811, May 2008. [5] S. Haykin, D.J. Thomson, and J.H. Reed, Spectrum sensing for cognitive radio, Proc. IEEE, vol. 97, no. 5, 2009. [6] S. Rao and W. A. Pearlman, Analysis of linear prediction, coding, and spectral estimation from subbands, IEEE Trans. Inform. Theory, vol. 142, no. 4 , pp. 1160-1178, 1996. [7] K.B. Eom and R. Chellappa, Noncooperative target classication using hierarchical modeling of high-range resolution radar signatures, IEEE Trans. Signal Process., vol. 45, no. 9, pp. 2318-2327, 1997. [8] B. Farhang-Boroujeny, and R. Kempter, Multicarrier communication techniques for spectrum sensing and communication in cognitive radios, IEEE Commun. Mag., vol. 46, no. 4, pp. 80-85, Apr.2008. [9] M. Milliger et al., Software Dened Radio : Architectures, Systems and Functions, Wiley, New York 2003. [10] O. Anjum, T. Ahonen, F. Garzia, J. Nurmi, C. Brunelli, and H. Berg, State of the art baseband DSP platforms for software dened radio : a survey, EURASIP J. Wirel. Commun. and Netw., vol. 5, pp. 1-19, 2011. [11] R. Woods, J. McAllister, G. Lightbody, and Y. Yi, FPGA-based Implementation of Signal Processing Systems, John Wiley and Sons, 2008. [12] J. Ma, G.Y. Li, and B.H. Juang, Signal processing in cognitive radio, Proc. IEEE, vol. 97, no. 5, pp. 805-823, 2009. [13] F. Ling and J.G. Proakis, A generalized multichannel least squares lattice algorithm based on sequential processing stages, IEEE Trans. Acoustics, Speech, Signal Process., vol. ASSP-32, no. 2, pp. 381-389, 1984. [14] M.T. Ozden and A.H. Kayran, Adaptive multichannel decision feedback equalization for volterra type nonlinear communication channels, Int. , vol. 62, no. 6, pp. 430-437, 2008. J. Electron. Commun. (AEU) [15] S. Haykin, Adaptive Filter Theory, 4/E, Prentice Hall, 2004.

B. Statistical Analysis We investigate the frequency estimation accuracy performance of the proposed method by performing variance and mean bias simulations for SNRs ranging from 20 dB to 20 dB with 1 dB increments. We assumed that the number of cisoids are known in these simulations, and then compared the variance performance of the proposed method against the Cramer-Rao Bound (CRB) in [17]. The variance and mean bias plots related to these simulations are presented only for the rst frequency in order to make it easier for the reader to separate the individual plots. The variance and mean bias curves of the proposed lattice method when estimating the spectrum of two cisoids in white noise are presented in gure 7 and 8 respectively. It can be observed in gure 7 that subband ltering improvement effects are more pronounced in the low SNR region (20 dB SN R -10 dB), and the threshold SNR, below which the estimation accuracy degrades rapidly, is about 9 dB. In the higher SNR region (10 dB < SN R < 10 dB), the subband estimation variance curves cross over the CRB while saturating earlier (at approximately SN R = 7 dB and SNR= 9 dB in two and four subbands respectively ) than the fullband estimation variance curve due to the shorter data lengths in subbands and the quantization effects induced by the sampling of the frequency variable. The saturation point can be moved to higher SNRs and accordingly the estimation accuracy can be improved by using longer data and more zeropadding at the cost of more computation time. In gure 8, we present mean bias curves for fullband as well as subband implementations. Similar to the variance curves, relatively higher mean bias behavior is observed in the low SNR region while lower values are observed after the threshold SNR. Note the poor variance and mean bias performance of the multitaper method in the low SNR region, which is considered because of its wide main lobes and comparatively higher sidelobes.

967

Normalized power spectrum (dB)

[16] Y.-P. Lin and V.D.D. Vaidyanathan, A Kaiser window approach for the design of prototype lters of cosine modulated lterbanks, IEEE Signal Process. Lett., vol. 5, no. 6, pp. 132-134, 1998. [17] S.M. Kay, Modern Spectral Estimation : Theory and Application, Prentice Hall, 1987.

SNR = 3 dB
1 5

SNR = 3 dB

(1)

10

(2) (3)

15

20

(4)
25

(1) Lattice fullband (2) Lattice twosubbands (3) Lattice foursubbands (4) Multitaper with K=3

30 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095 0.1 0.105 0.11 0.115 0.12

Fraction of sampling frequency

Fig. 1 : Block diagram of the adaptive multichannel ARMA prediction ltering in subbands.
0 2
Normalized power spectrum (dB)

Fig. 5 : SNR improvement performance.

SNR =0 dB
1

SNR =0 dB
2

4 6 8 10 12

(1) Lattice fullband

(1) (2)

(2) Lattice twosubbands

(3) Lattice foursubbands (4) Multitaper with K=3

(3)
14 16

Fig. 2 : Block diagram of the adaptive multichannel ARMA lattice prediction ltering in subbands.

(4)
18 20 0.17 0.175 0.18 0.185 0.19 0.195 0.2 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 0.245 0.25 0.255 0.26 0.265 0.27

Fraction of sampling frequency

Fig. 6 : Whitening effect performance.

100 90 80 70 60 50

10*log10(Variance(fest))

(5) (4) (2) (3) (1) Lattice fullband (2) Lattice twosubbands (3) Lattice foursubbands (4) Multitaper with K=3 (5) CRB for white noise
5 0 5 10 15 20

40

(1)
30

Fig. 3 : Diagram of the four-channel ARMA lattice lter structure for two-subband spectrum estimation.
0 5 10 15 20 25 30 35 40

20 20

15

(1) 10

SNR(dB)

Fig. 7: Variance vs. SNR plots.

(1) 10 SNR1 =
(1) (2)
10*log10(Mean Bias(fest))

SNR = 2 15

(1) (1) Lattice fullband (2) Multitaper with K=3 (3) Lattice twosubbands (4) Lattice foursubbands

Normalized power spectrum (dB)

(3)

20

(2) (3)

25 (4) 30

(1) Lattice fullband (2) Lattice twosubbands (3) Lattice foursubbands (4) Multitaper with K=3
(4)

35

45 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095 0.1 0.105 0.11 0.115 0.12

Fraction of sampling frequency

40 20

15

10

0 SNR(dB)

10

15

20

Fig. 4 : Frequency spacing improvement performance.

Fig. 8 : Mean bias vs. SNR plots.

968