Hybrid Spectrum Sensing Experimental Analysis Using GNU radio and USRP for Cognitive Radio Amor Nafkha∗ , Malek Naoues∗ , Krzysztof Cichon† , Adrian Kliks † and Babar Aziz ‡ ∗ CentraleSup´elec/IETR, Avenue de la Boulaie, 35576 Cesson S´evign´e, France, email: {amor.nafkha malek.naoues}@centralesupelec.fr † Poznan University of Technology, Polanka 3, 60-965 Poznan, Poland, email: {kcichon, akliks}@et.put.poznan.pl ‡ IFSTTAR,LEOST, F-59650 Villeneuve d’Ascq, France, email: [email protected] Abstract—The dynamic spectrum access is a concept opposite to static frequency spectrum assignment and one of the key elements of cognitive radio (CR). It allows secondary users to access unused licensed frequency bands. In order to detect the presence of primary users, one of the fundamental tasks that a CR equipment has to perform is spectrum sensing process. In this paper, a hybrid spectrum sensing architecture is proposed, which consists of the concatenation of a sequential energy detection (SED) and a cyclostationary-based (CB) detector. The former sensing method is performed to reduce spectrum sensing time, when the SNR is very high (resp. very low), while the latter sensing detector is performed whenever the SED does not guarantee the required quality of detection (i.e. moderate SNR regime). The efficiency of the proposed solution has been verified experimentally using USRP N210 boards under GNURadio environment.

I. I NTRODUCTION Spectrum sensing is a fundamental and basic functionality for the realization of the cognitive radio technology. It performs the role of monitoring and detection of spectrum holes. In a cognitive radio system, the first challenge of the unlicensed secondary user (SU) is to properly detect, share the spectrum with licensed primary users (PU), and avoid the harmful interference with them. In literature, several spectrum sensing techniques are developed [1]. In general, these techniques can be classified into three categories: (i) methods requiring both PU signal and noise variance information, (ii) methods requiring only noise variance information (also called semi-blind methods), and (iii) methods not requiring any information on PU signal or noise variance (known as blind methods). Unfortunately, it can be noticed that the more reliable an algorithm is, the more complicated is its hardware implementation. The energy detector (ED) [2] is a semi-blind sensing technique as no prior information about the PU is needed. However, it is known by its dependency on noise uncertainty [3], which is one of its major drawbacks. On the other hand, advanced cyclostationary-based (CB) techniques, such as [4], offer high reliability at the price of high complexity and relatively long processing time. The main contribution of this paper is the performance investigation of the three different spectrum sensing techniques to detect the presence of the PU in real experimental environment. The first two techniques (SED [5], CB [4]) are well-known and effective, but include significant shortcomings This research is supported by the European Commission in the framework of the FP7 Network of Excellence in Wireless COMmunications NEWCOM# (FP7 Contract Number: 318306)

(i.e. high complexity, noise uncertainty,..), while the third technique is based on a hybrid architecture of the SED and CB algorithms. The main motivation behind the hybrid architecture is twofold: first, the reliability of the detection can be improved due to the logical analysis of the decisions generated by SED and CB blocks connected in serial manner. In such a case the decision will be made after gathering S samples of the analyzed signal. Second, in the case of a very high or extremely low SNR of the analyzed signal, the sensing phase can be shortened, since the reliable decision can be made by the SED block just after collecting a fraction Ki (i ∈ {1 . . . K}) of the normally assumed S samples where K is the total number of the processed samples segments of S where each segment contains N samples. Through the reduction of the sensing time, the SU can either start its transmission earlier or detect the presence of the PU signal reducing the interference induced to it. The rest of this paper is organized as follows. Section II introduces the basics of the spectrum sensing. Moreover, the proposed hybrid architecture is analyzed providing pros and cons of each possible scheme. An overview of our experimental setup is presented in Section III. In Section IV, we evaluate the performance of the different presented algorithms in a real environment, while conclusions are drawn in Section V. II. S PECTRUM S ENSING F UNDAMENTALS In CR systems, a SU needs to make periodical decisions on spectrum occupancy; for the first time, such operation has to be performed when the CR terminal expresses its needs for data transmission, then it has to cyclically analyze the spectrum in order to vacate the used frequency band as soon as the presence of the PU signal is detected. The concept of spectrum sensing can be mathematically treated as a classical detection problem related to hypothesis verification [1]. Two binary hypotheses H0 and H1 can be defined to indicate the absence or presence of a PU in the environment. The received signal at a SU, r(k), can be expressed as:  H0 : n(k) r(k) = r(k.Ts ) = (1) H1 : h(k) · s(k) + n(k) where s(k) and h(k) stand for the PU signal and channel impulse response, respectively, n(k) is an additive white Gaussian noise (AWGN); when treating interference as noise, n(k) can represent the ambient noise and interference as well. n(k) is assumed to be circularly symmetric complete

Gaussian random variable with zero mean and σn2 variance and distributed as n(k) ∼ CN (0, σn2 ). The sampling frequency is defined by Fs = T1s . The objective of the spectrum sensing operation is to decide between H0 and H1 based on the observation of the received signal r(k). The detection performance is characterized by two probabilities: probability of detection, Pd , where the decision is H1 , while H1 is true; and probability of a false alarm, Pfa , which corresponds to the case where the decision is H1 , while H0 is valid. In our experiments, two algorithms have been tested, i.e., the sequential version of energy detection, and a method for cyclostationarity-based detection called Symmetry Property of Cyclic Autocorrelation Function (SPCAF). Let us briefly summarize the theoretical basis of these algorithms, which will be followed by a description of the hybrid solution. A. Sequential Energy-Based Spectrum Sensing The idea behind the energy-based spectrum sensing, being at the same time one of the simplest ways of PU detection, is to calculate the amount of the received power in the considered frequency fragment and compare this value with a decision threshold derived based on the noise variance [6]. In the case that the received power is smaller than the previously approximated threshold, the algorithm will make a decision on the spectrum vacancy. In turn, the channel will be assumed to be occupied by a PU if the computed signal power is much higher than the noise variance. The crucial part is played by a properly defined decision threshold, and in consequence, by the correctness of noise variance, and the duration of sensing time (expressed in seconds or - for discrete signals - in terms of the number of gathered samples). For the given values of probability of a false alarm, Pfa , number of collected samples S, and noise variance σn2 , the decision threshold can be defined as follows: i h √ (2) γthr = σn2 · Q−1 (Pfa ) · 2S + S where Q() represents the Gaussian Q-function and is defined R +∞ u2 as Q(x) = √12π x e− 2 du. Having in mind that the total power of S collected samples in the given frequency band can PS−1 2 be represented as the random variable ΓS = k=0 |r[k]| , then, based on (1), the generic decision rule DS can be modified to the considered case:  H0 : ΓS ≤ γthr DS = (3) H1 : ΓS > γthr The behavior of an energy detector can be improved in various ways, e.g., by extending the sensing time or by the application of an adaptively modified threshold. In our experiments, we have selected a double-threshold, sequential energy detector which has the same reliability as a classical one, but its application could reduce the sensing time. The main concept is based on the assumption that for a very strong PU signal, or - contrarily - in the presence of noise only (or a very weak signal), the number of samples that should be collected for decision making with a given reliability can be reduced. If this is the case, the sensing time is minimized, increasing,

Decision

Decision

H0

H1 No Decision

2 n

Fig. 1.

LO,i

HI,i

2 n (1+SNR)

Decision regions in the Sequential Energy-Based Spectrum Sensing

at the same time, the time devoted to data transmission and reducing the energy consumption. In order to achieve this goal, two decision thresholds have to be applied, γHI and γLO , which will be used for decision making if the signal is or is not present in the observed frequency fragment. In a nutshell, the procedure can be realized in an iterative way. The energy detector collects the signal samples in the shorter period N and tries to make the decision. If the computed power is higher than γHI , the decision on the PU signal presence can be made; if the received power is lower than γLO one can state that the considered channel is vacant. If the calculated value falls between these thresholds, the sequential energy detector collects the next block of N samples and repeats the procedure. When the total number of gathered samples reaches the allowed maximum S (i.e., the maximum sensing time is up), the decision is made as for traditional algorithms. The decision rule for i-th iteration is illustrated in Fig. 1 and can be written as follows:  ΓNi ≤ γLO,i  H0 : continue : ΓNi ∈ (γLO,i , γHI,i ) DNi = (4)  H1 : ΓNi ≥ γHI,i where ΓNi denotes the average power after collecting Ni = i × N samples, i ∈ {1 . . . K}. The γHI,i and γLO,i are defined in (5), where Pfa,LO,i and Pfa,HI,i denote the probability of a false alarm assumed for low and high thresholds in the ith iteration.    √ γLO,i = σ ˆn2 · Q−1 (Pfa,HI,i ) · √2Ni + Ni  (5) γHI,i = σ ˆn2 · Q−1 (Pfa,LO,i ) · 2Ni + Ni Where σ ˆn2 is the estimated noise variance and it is assumed to be in an interval support [3] [(1 + ξ)−1 σn2 , (1 + ξ)σn2 ], ξ > 0. Using the fact that the noise uncertainty has a bounded model which can be included in the no decision region, we can evaluate the two thresholds in Eq. 5 based on the estimated noise variance. B. Cyclostationarity Feature-Based Spectrum Sensing In wireless communications, the transmitted signals display very strong cyclostationary features as introduced by Gardner in [7]. Therefore, identifying a unique set a features of a particular radio signal can be used to detect its presence based on its cyclostationary features. In the context of spectrum

sensing, many works have been conducted on using the cyclostationary features to detect the presence of PU in the radio environment [1]. In general, this method can perform better than the energy based detector. However, its main drawbacks are the complexity associated with the detection technique and the need for some a-priori knowledge of the PU signal (e.g. cyclic frequency). The cyclostationarity feature detector can be realized by analyzing the Cyclic Autocorrelation Function (CAF) of a received signal r(k). The CAF of a received signal r(k) at the SU can be expressed as: Rr (k, τ ) =

+∞ X

Rrα (τ )e2πjαk

(6)

where τ is a lag associated with the autocorrelation function, α the cyclic frequency, S the number of samples and Rrα (τ ) is given by (7). S−1 1 X Rr (k, τ )e−2πjαk = lim S→∞ S

(7)

k=0

1) Classical Cyclostationary Feature-Based Detector: The classical approach to realizing the cyclostationary detector is based on the Cyclic Spectrum Density (CSD) or the spectral correlation function of the received signal r(k). Υα r (f ) =

S−1 1 X α R (τ )e−2jπf τ S τ =0 r

(8)

The CSD function presented in (8) exhibits peaks when the cyclic frequency α equals the fundamental frequencies of s(k) the transmitted signal. Under the H0 hypothesis, the CSD function does not have peaks, since the noise is generally non-cyclostationary. Using this technique, it is possible to distinguish even weak PU signals from the noise at a very low SNR, where the energy detector is not applicable. 2) SPCAF Detector: The discrete-time consistent and unbiased estimation of the CAF of a random process is given as: M −1 X ˜ α ∗ (τ ) = 1 R r(k)r∗ (k + τ )e−2jπαk (9) rr M k=0

For a given lag parameter τ ∈ {1, 2, . . . , L}, the cyclic autocorrelation function (CAF) can be seen as a Fourier transform of V = [r(0)r∗ (0+τ ), . . . , r(M −1)r∗ (M −1+τ )], where M is FFT size. As shown in the work of Khalaf et al. [4], the CAF is an M -dimensional sparse vector in a cyclic frequency domain for a fixed lag parameter τ . Moreover, it presents a symmetry property as illustrated in (10). α ˜ rr ˜ −α ||R ∗ (τ )||2 = ||Rrr ∗ (τ )||2

N×K Samples

SPCAF

N×K Samples

Decision

Decision

Fig. 2.

Schematic hybrid detector diagram

verifies the symmetry. However, when using a small number of samples, the probability of obtaining symmetrical CAF under H0 is very low [4]. This SPCAF technique can perform with a limited number of samples and consequently, with lower complexity and a shorter observation time as compared to the classical approach. C. Proposed Hybrid Approach

α=−∞

Rrα (τ )

No Decision

SED

(10)

Using a compressed sensing (CS) recovery technique like the Orthogonal Matching Pursuit (OMP) algorithm , we can accurately estimate the CAF using a limited and small number of received samples S γHI ) (12) K The sensing time of the SED can be approximated by tS ≈ f (ρ) =

The paper analyzes the performance of the hybrid spectrum sensing algorithm, focusing mainly on the probability of detection as the function of averaged SNR in the system. For the proposed two-phase, non-iterative solution, the CDF that lies between the curves of the pure SPCAF and energy detection, has been achieved. It means that there is a trade-off between the reliability of the proposed spectrum algorithm and its processing time. The more restrictive decision thresholds applied in the first sensing phase, the better the reliability of the solution. The true benefit of the considered solution is the reduction of the sensing time. Some initial results have been achieved, which show the probability that no decision is made after the collection of N = 128 and N = 512 received samples. This results in a significant reduction of the sensing time. This observation will be the subject of a future research study. R EFERENCES [1] L. Lu, X. Zhou, U. Onunkwo, and G. Y. Li, “Ten years of research in spectrum sensing and sharing in cognitive radio,” EURASIP Journal on Wireless Communications and Networking, vol. 2012, no. 1, pp. 1–16, 2012. [2] Y. Zeng and Y.-C. Liang, “Maximum-minimum eigenvalue detection for cognitive radio,” in Proc. IEEE PIMRC, vol. 7, 2007, pp. 1–5. [3] S. Bahamou and A. Nafkha, “Noise uncertainty analysis of energy detector: Bounded and unbounded approximation relationship,” in Signal Processing Conference (EUSIPCO), 2013 Proceedings of the 21st European, Sept 2013, pp. 1–4. [4] Z. Khalaf, A. Nafkha, and J. Palicot, “Blind spectrum detector for cognitive radio using compressed sensing and symmetry property of the second order cyclic autocorrelation,” in 7th International ICST Conference on Cognitive Radio Oriented Wireless Networks and Communications (CROWNCOM), 2012. IEEE, 2012, pp. 291–296. [5] N. Kundargi and A. Tewfik, “Doubly sequential energy detection for distributed dynamic spectrum access,” in IEEE International Conference on Communications (ICC), 2010, May 2010, pp. 1–5. [6] H. Urkowitz, “Energy detection of unknown deterministic signals,” Proceedings of the IEEE, vol. 55, no. 4, pp. 523–531, April 1967. [7] W. A. Gardner, “Exploitation of spectral redundancy in cyclostationary signals,” IEEE Signal Processing Magazine, vol. 8, no. 2, pp. 14–36, 1991.