A USRP based scheme for cooperative sensing networks

Transcription

1 A USRP based scheme for cooperative sensing networks Ricardo S. Yoshimura 1, Fabiano S. Mathilde 1, João P. M. Dantas 2, Vicente A. de S. Jr. 2, José H. da Cruz Jr. 3, Juliano J. Bazzo 1, Dick C. Melgarejo 1 1 Centro de Pesquisa e Desenvolvimento em Telecomunicações (CPqD) Campinas, SP Brazil 2 Universidade Federal do Rio Grande do Norte (UFRN) Natal, RN - Brazil 3 Universidade Estadual de Campinas (Unicamp) Campinas, SP Brazil {rseiti, fabianom, jbazzo, dickm}@cpqd.com.br, {jpaulo.gppcom, vicente.gppcom}@gmail.com, jrdecom@decom.fee.unicamp.br Abstract. In this contribution, a hardware platform for cooperative sensing networks is presented with the use of Roy s Largest Root Test, also known as a blindly combined energy detection scheme, such that the system is configured with three ETTUS USRPs N210 and one computer for processing. The hardware experiments were validated with simulations, and the results show that the cooperative scheme outperforms the case when sensing is taken individually, keeping in mind that the number of samples collected is a key performance indicator. 1. Introduction The growing demand for higher data rates in the unlicensed spectrum frequencies have attracted the attention of regulatory agencies such as FCC to explore unused portions of the spectrum in the licensed bands. In this way, cognitive radio (CR) emerged as a possible solution to overcome this spectrum scarcity, so that these devices (also called secondary users) can communicate each other while the licensed user is not transmitting or not detected over the covered region. As defined in [ETSI 2013], a CR is capable to obtain the knowledge of radio operational environment and established policies and to monitor usage patterns and users needs. Furthermore, it dynamically and autonomously adjust its operational parameters and protocols. The studies started with J. Mitola and G. Maguire (1999) and it has increased exponentially in the last decade. Since these radios should not be able to interfere while a primary user is covering a given area, spectrum sensing [Sahai et al 2004] is a key functionality that has to be researched. In this context, the authors in [Cabric et al 2004] described three sensing techniques: Matched Filter, Energy Detector and Cyclostationary Feature Detection. By combining CRs with one of these techniques, initial analysis on cooperative sensing was presented and some challenges were pointed out, like the different sensitivities and sensing times, and the need of a control channel. As main conclusion, they stated that cooperative sensing outperforms the other mentioned techniques in terms of the quality of final sensing decision. The last publications about spectrum sensing and cognitive radio explore many statistical strategies [Do et al 2013], channel types [Huang 2013][Yufan 2012] and even user reputation [u-van and Koo 2012]. The need to optimize spectrum capability of wireless communication systems and actual software-hardware capabilities to implement flexible radio functionalities 67

2 create common opportunities to Cognitive Radio and Software Defined Radio (SDR) technologies to work together. SDR is defined as a radio in which the radio frequency (RF) operating parameters including, but not limited to, frequency range, modulation type, or output power can be set or altered by software, and/or the technique by which this is achieved [ETSI 2013]. Another alternative definition is provided by the Wireless Innovation Forum [Wireless Innovation Forum 2013]: Radio in which some or all of the physical layer functions are software defined. These devices include field programmable gate array (FPGA), digital signal processors (DSP), general purpose processor (GPP), programmable system on chip (SoC) or other application specific programmable processors. The use of these technologies allows new wireless features and capabilities to be added to existing radio systems without requiring new hardware. The development of CR systems requires some algorithms that could be implemented in SDRs. The Universal Software Radio Peripheral (USRP) was developed by Ettus Research LLC that provides low cost radio systems for commercial and research applications [Open Source gnu-radio 2013]. USRP provides digital baseband and IF section within the hardware, which aids to use general purpose computers to operate as high bandwidth software radio. The so called daughterboard, which is attached to USRP main board, provides the RF front-end so that it can operate over a variety of spectrum ranges. A common USRP complement is GNU Radio, which is a free software development kit that provides signal processing modules to build an SDR in a real-time environment using low cost and reconfigurable radios. It is based on block architecture and involves hybrid Python/C++ programming [Open Source gnu-radio 2013]. This paper addresses the implementation of a sensing algorithm on USRP N210 using GNU Radio when two secondary users are cooperating. Sensing decisions are based on Roy s Largest Root Test (RLRT) algorithm, which provides a blind combined energy detection scheme through the evaluation of eigenvalues of received signal covariance matrix. Results from experiment are compared to simulation in two cases: (i) cooperation by RLRT; and (ii) without cooperation (individual taken decision). To our knowledge, some experimental studies have been conducted considering energy detection [Bielefeld et al 2010][Nir and Scheers 2012][Aftab and Mufti 2010] and simple eigenvalue-based algorithm [Buucardo 2010]. However, a comparison of simulation and experimental studies involving cooperative RLRT has never been put forward in the literature before. This paper is organized as follows: Section 2 briefly describes cooperative sensing paradigm; Section 3 presents the implemented cooperative detection algorithm as well as our system modeling; Section 4 explains the USRP implementation and discusses implementation related issues; Section 5 presents our simulation models and assumptions. Simulation and measurement results are presented in Section 6. Finally, a conclusion is given in Section Cooperative Sensing The According to [Ghasemi 2005][Visotsky et al 2005], cognitive radios are able to reduce uncertainties and relax individual sensing requirements with the use of cooperative sensing. In this approach, each cognitive radio performs spectrum sensing 68

3 and the results are combined in such a way that more accurate detection can be achieved concerning the primary user activity. When signals are subjected to small fading or shadowing, a cognitive radio requires higher detection sensitivity in order to overcome the uncertainty caused by this channel s randomness. However, sensors or secondary users placed more than a few wavelengths from each other may experience uncorrelated small fading effects. Therefore, the uncertainty due to small fading effects may be mitigated when different users share their sensing results and cooperatively decide if the channel is occupied or not by a primary user. In this way, such diversity gain can improve the accuracy regarding the detection sensitivity without the use of severe sensitivity requirements on individual cognitive radios [Fatemieh et al 2010]. There are basically two ways to model a cooperative sensing: In centralized collaborative fashion, the cognitive radios report their results to a centralized data base in certain periods of time or when requested. Softcombining techniques can be applied in such a way that raw signal power measurements from cognitive radios are combined. Otherwise, hardcombining techniques considers a 0/1 decision from each cognitive radio. However, the centralized scheme adds some extra overhead on the communication system and a control channel is needed to enable the exchange of information between the cooperating cognitive radios and the data base. This kind of scheme is included, for example, in the IEEE standard draft (2013); In distributed collaborative fashion, each individual sensing measurement is exchanged with the neighbors, so that the primary user presence is determined by the network with no support of a base station. 3. Eigenvalue-based scheme for cooperative spectrum sensing and system modelling In this section we present the strategy for a centralized cooperative spectrum sensing, based on the eigenvalues of the received signal covariance matrix [Zeng 2009][Kortun et al 2010][Neto and Guimarães 2012]. This strategy is considered blind, because it does not need any previous knowledge of the primary user (such as energy detection). For this reason, this scheme has shown to be relevant for practical purposes. Consider a model described by p primary users and m cognitive radios (sensors or secondary users). The channel between the primary user j and the cognitive radio i is represented by the coefficients h ij, i = 1, 2,, m and j = 1, 2,, p. Each cognitive radio collects n samples of incoming signal and sends them to a remote data base (following a centralized approach). After that, the combination of these collected signals will help decide the occupation of the required frequency channel. This model is illustrated in Figure 1. 69

4 Figure 1: Illustration of centralized cooperative sensing model. During each sensing period there are two hypotheses: (i) hypothesis H 0 considers that there is no primary user in the required frequency channel; and (ii) hypothesis H 1 considers that there is a primary user at the sensed frequency bandwidth. Under these assumptions, a resultant Y mxn matrix among all cognitive radios can be used to build the following binary hypotheses test H 0 1 : Y= V H : Y= HX+ V (1) where X pxn is the matrix corresponding to all the samples sent by different primary users and V mxn comes from the noise matrix. H is the channel matrix between the primary user j and the cognitive radio i, written as h11 h12 L h1p h21 h22 L h2p H = (2) M M O M hm 1 hm2 L hmp Each element of matrix H, given by h ij, i = 1, 2,...m and j = 1, 2,..., p, represents the complex gain coefficients of the channel. In the cooperative sensing based on eigenvalues, vacant spectrum are detected with the use of a statistical test based on the eigenvalues of the covariance matrix R y, which can be estimated through Rˆ y= 1 H (3) n YY where the operator (.) H means the Hermitian transpose (complex conjugate and transpose). From the eigenvalues of such matrix it is possible to define the decision variable such that T RLRT λ = (4) σ represents the quotient between the maximum eigenvalue of correlation covariance matrix and noise variance (noise power). Equation 4 defines Roy s Largest Root Test [Nadler et al 2011] and the RLRT algorithm, also known as a blindly combined energy detection scheme (BCED) [Zeng et al 2008][Kortun et al 2010] or a maximum eigenvalue detection (MED) [Kortun et al 2010]. This decision variable is then compared to the decision threshold γ. If T RLRT > γ, then it is considered that the primary user is present in the sensed frequency bandwidth; otherwise it is considered that the primary user is not present. max 2 v 70

5 However, due to the random characteristics of the signal, when a channel is classified as free or occupied, there is no guarantee of the results accuracy, but there is a probability associated to this classification. There are two parameters used to analyze the performance of any binary hypothesis test: (i) the false alarm probability (P FA ) and the detection probability (P D ). The false alarm probability represents the probability of deciding for a present primary user, while in fact it is not present in the sensed frequency bandwidth, such that [ T > γ H ] = P FA = Pr 0 f0( t) dt (5) where f 0 (t) represents the probability density function (PDF) of the used decision variable T RLRT, under hypothesis H 0. On the other hand, the detection probability is the probability of deciding for a present primary user, while it is really present in the sensed frequency bandwidth, such that γ [ T > γ H ] = P D = Pr 1 f1( t) dt (6) Differently from the P FA case, f 1 (t) is the PDF of the used decision variable T RLRT, under the hypothesis H Hardware Implementation Our SDR setup models a scenario under primary user activity with two cognitive radios acting as sensors for cooperative sensing. These primary user and sensors consist of three ETTUS USRP N210 with firmware version UHD_ g On software side, we implemented signal measurements using GNU Radio Companion version 3.6.1git-64-g23dd54bf. γ Figure 2: Block diagram of the USRP implementation. Figure 2 shows a block diagram of how our cooperative sensing scheme was implemented. Each sensor is a USRP based platform responsible for performing signal measurement through a scenario with a primary user transmitting continuously a 5 MHz OFDM signal. A signal generator (Agilent MXG Vector Signal Generator N5182B) provided a controllable noise like signal such that its power spectral density over the primary user bandwidth could be precisely adjusted, in order to compose the desired range of SNR values. In other words, for the sake of our reasons, the primary user signal was set on a fixed power level while the noise effects was emulated by the signal generator, over which its power level was adjusted for each particular SNR. By adding these signals, the resultant one is then divided through a RF splitter, where its output feeds each sensor. Next, the spectrum sensing data is collected separately by each sensor to produce samples for both H 0 and H 1 hypothesis, which are stored in a common data base with the use of a personal computer. More specifically, each USRP sensor collects n samples a 71

6 time at 770 MHz central frequency, and move forward them to the data base. After repeating this procedure a hundred times in a cyclic way, the personal computer evaluates the RLRT decision variable, yielding N e ( number of events ) values of T RLRT, so that the decision-taking rule can be applied over the calculated values of T RLRT. Finally, the channel status (occupied or free) is decided. Keep in mind that, considering equation 4, the noise power was measured before all the aforementioned procedure could take place. This measurement collected samples for each specific noise power, assuming the use of the noise like generator. Since we had knowledge about the primary user activity i.e., if it is transmitting a signal or not, detection and false alarm probabilities could be computed. In order to provide an environment free of undesirable signals, the connection among the devices was wired oriented. Undesirable signal is interpreted as all the signals besides the additive white Gaussian noise that are not generated by the aforementioned devices. General parameters of USRP setup are presented in Table 1. Table 1. General USRP implementation parameters. Name Value Primary user bandwidth 5 MHz Primary user modulation OFDM Primary user signal FFT size 512 Samples by T RLRT 1000 and Total of experiments (values of T RLRT ) 100 Central frequency 770 MHz GNU Radio Companion version 3.6.1git-64-g23dd54bf RF front-end ETTUS SBX daughterboard USRP version N210 with firmware version UHD_ g Simulation Modeling The simulator used for performance comparison is based on Monte Carlo approach [Kortun et al 2010] which evaluates the average behavior of the stochastic processes involved in cooperative sensing modeling described in Section 3. It calculates detection and false alarm rates for a range of detection thresholds (γ) equally distributed between the values of γ min and γ max. The value of γ min is assumed to be the minimum value of the calculated T RLRT metric under hypotheses H 0 whereas γ max is the maximum value of calculated T RLRT metric under hypotheses H 1. For controlling the accuracy of the results, a specified number of events (N e ) is defined. Thus, the simulator calculates N e values of T RLRT metric for hypotheses H 0 and H 1. Later, metrics P D and P FA are evaluated. The implemented cooperative sensing strategy depends on the noise power (σ 2 ), which, for simulation purposes, is calculated by the SNR, an input parameter of the simulation. The simulation input parameters and output metrics are summarized in Table 2. 72

7 6. Results Table 2. Simulation parameters and output metrics. Name Input parameters Number of primary users Number of secondary users Input SNR Number of Monte Carlo events Number of samples for calculation of each T RLRT Detection rate False alarm rate Output metrics Symbol The ROC (Receiver Operating Characteristic) curve was used to compare the performance of implemented algorithms, allowing one to explore the relationship between the sensitivity of sensing through P D, and the specificity of sensing through P FA. The y-axis is represented by sensitivity (P D ) and the x-axis by the specificity (P FA ), both given as percentages. Another evaluation artifact is the plot of detection capability (P D ) for different values of SNR, given a fixed P FA. Following recommendations of [Stevenson et al 2005], we adopt 10% as P FA target. Table 3 shows the set of parameters used in this paper for the performance evaluation. p m SNR Ne n P D P FA Table 3. Performance evaluation parameters. Name Value No. of primary users (p) 1 Primary user modulation OFDM (for both simulation and experiment) No. of secondary users (m) 1 (individual sensing) and 2 (cooperative sensing) Input SNR range -20 to 0 (for both simulation and experiment) No. of samples for calculation of each 100 and 1000 for simulations T RLRT (n) 1000 and for USRP experiments Number of Monte Carlo events (Ne) 100 Figure 3 shows the ROC curve of simulation results for individual and cooperative sensing considering different values of samples by T RLRT (n). As expected, the higher the samples by T RLRT, the better the quality of detection. We also have a clear evidence that cooperative sensing (m=2) outperforms individual sensing (m=1). The value of n has a significant influence on detection performance. Cooperative sensing (m=2) with few samples by T RLRT (e.g. n=100) could have lower performance than individual sensing with some more samples by T RLRT (e.g. n=1000), at the cost of longer sensing time. 73

8 Figure 3: RoC curve of simulation results for individual (m=1) and coperative (m=2) sensing. Figure 4: PD vs SNR of simulation results for individual (m=1) and cooperative (m=2) sensing. Figure 4 confirms these findings by showing PD vs SNR curves for the same cases presented previously. Now, we can state that such conclusions are valid considering the tested range of noise power. Following the setup configuration shown in Section 4, Figure 5 presents the comparison between simulation and hardware implementation using USRP. As expected, there is a performance degradation comparing USRP implementation to simulation results. Considering n=1000 for example, our USRP implementation of cooperative sensing provides 100% of P D just over 3dB of SNR compared to simulation. We claim this is the consequence of analogue RF front-end of USRP. While the digital signal loaded to the SDR might be, in some sense, perfect, once it passes from the digital into the domain of practical devices, it is still vulnerable to non-linear effects and other such worldly imperfections. Such distortions may be the reason for performance gap, as shown in Figure 5. Other source of distortion is some imperfection of noise power estimation. Figure 5: Simulation vs USRP implementation performance of cooperative spectrum sensing based on RLRT. Figure 6: Simulation vs USRP implementation performance of individual spectrum sensing based on RLRT. Figure 6 presents USRP implementation for individual sensing. Also, there is a performance gap between USRP implementation and simulation, which is about 5dB for 74

9 n = Those gaps reduces meaningfully for higher number of samples (e.g. n=10000). 7. Conclusions This paper addresses the implementation of a sensing algorithm on USRP N210, with the use of GNU Radio and contextualized in two scenarios: in one of them two users are cooperating and in another one it is an individual taken decision. Sensing decisions are based on Roy s Largest Root Test (RLRT) algorithm, which provides a blind combined energy detection scheme through the evaluation of eigenvalues of the received signal covariance matrix. The considered scenario was carefully set such that the SNR could be precisely measured in order to obtain the desired ROC curves. We compared the hardware experiments to simulation results ranging important parameters of the considered detection algorithm. In this way, we validated the effectiveness of the cooperative sensing, which outperforms the case when sensing is taken individually. A performance gap between USRP implementation and simulations was observed, especially if few samples are taken to calculate the sensing decision variable. High number of samples decreases significantly this gap. As a motivation for future work, we intend to improve the algorithm and compare the implementation with other eigenvalue-based detection schemes using USRP N210 and GNU Radio, testing the algorithms in real environment. Moreover, it is desirable to increase the number of cooperative sensors and evaluate the performance not only in flat fading scenarios but also in multi-path environments, since these ones could bring the true benefits of a cooperative sensing, despite of individual sensing. Acknowledgment UFRN team would like to thank the Brazilian Research Agencies CNPq and FAPERN for partial financial support. From CPqD side, this work was funded by Brazilian Fund for Telecommunications Development (FUNTTEL) as a scope of the Advanced Wireless Access Network (RASFA) Project. References Mitola J. and Maguire G. Q. (1999), "Cognitive radio: making software radios more personal", IEEE Personal Communications, Vol. 6, No. 4, Aug., pp Sahai A., Hoven, Tandra R. (2004), Some Fundamental Limits on Cognitive Radio, Proc. of Allerton Conference, Monticello, Oct. Cabric D., Mishra S. M. and Brodersen R. W. (2004), Implementation Issues in Spectrum Sensing for Cognitive Radios, Proceedings of 38th Asilomar Conference on Signals, Systems and Computers, Pacific Grove, November, pp Do, C.; Tran, N.; Hong, C.; Lee, S.; Lee, J.; Lee, W. (2013), "A Lightweight Algorithm for Probability-Based Spectrum Decision Scheme in Multiple Channels Cognitive Radio Networks", Communications Letters, IEEE. Huang, Y.; Huang, X. (2013), "Detection of Temporally Correlated Signals over Multipath Fading Channels," Wireless Communications, IEEE Trans. on, pp Yufan W. (2012), "Cooperative Spectrum Sensing for Cognitive Radio Networks under Nakagami-M Fading Channels", Industrial Control and Electronics Engineering (ICICEE), International Conference on, vol., no., pp , Aug. 75

10 u-van, H.; Koo, I. (2012), "A sequential cooperative spectrum sensing scheme based on cognitive user reputation", Consumer Electronics, IEEE Transactions on, vol.58, no.4, pp , November. ETSI Terms and Definitions Database Interactive (TEDDI) (2013), February. Wireless Innovation Forum, (2013), February. Open Source gnu-radio (2013), Last Accessed, February. Bielefeld D., Fabeck G., Zivkovic M. and Mathar R. (2010), Optimization of Cooperative Spectrum Sensing and Implementation on Software Defined Radios, 3rd Int. Symp. on Applied Sciences in Biomedical and Comm. Tech. Nir V. and Scheers B. (2012), Implementation of an adaptive OFDMA PHY/MAC on USRP platforms for a cognitive tactical radio network, MCC. Aftab A. and Mufti M. N. (2010), Spectrum Sensing Through Implementation of USRP2, Master Thesis, School of Computing, Blekinge Institute of Technology. Buucardo A. (2010), A Signal Detector for Cognitive Radio System, Master Thesis, University of Gavle. Ghasemi A. and Sousa E. (2005), Collaborative spectrum sensing for opportunistic access in fading environments, IEEE DySPAN 05, Nov. Visotsky E., Kuffner S., and Peterson R. (2005), On collaborative detection of tv transmissions in support of dynamic spectrum sharing, IEEE DySPAN 05. Fatemieh O., Chandra R. and Gunter C. (2010), Secure Collaborative Sensing for Crowdsourcing Spectrum Data in White Space Networks, IEEE Symposium on New Frontiers on Dynamic Spectrum. IEEE Working Group Website (2013), February. Zeng, Y.; Liang, Y. C. (2009), Eingenvalue-based spectrum sensing algorithms for cognitive radio, IEEE Transactions on Communications, v. 57, p Kortun, A. et al. (2010), On the performance of eigenvalue-based cooperative spectrum sensing for cognitive radio, IEEE Symp. New Frontiers in Dyn. Spectrum. Neto, J. S.; Guimarães, D. A. (2012), Sensoriamento espectral cooperativo baseado em autovalores para rádios cognitivos, Revista Telecomunicações, v. 14. Nadler, B.; Penna, F.; Garello, R. (2011), Performance of eigenvalue-based signal detectors with known and unknown noise level, IEEE Int. Conf. on Comm. (ICC). Zeng, Y.; Liang, Y. C.; Zhang, R. (2008), Blindly combined energy detection for spectrum sensing in cognitive radio, IEEE Signal Processing Letters, v. 15. Kortun, A.; Ratnarajah, T.; Sellathurai, M. (2010), Exact performance analysis of blindly combined energy detection for spectrum sensing, IEEE 21 st PIMRC. Hayes, J. F.; Babu, T. V. J. G. (2004), Modeling and Analysis of Telecommunication Networks-Monte Carlo Simulation: Wiley-Interscience, ISBN Stevenson, C. R., Cordeiro, C., Sofer, E., and Chouinard, G. (2005), Functional requirements for the WRAN standard. IEEE Technical Report. 76