X[k] = x[n] e j2π k /17/$ IEEE 278. n=0 F 1. N F n. (1)

Transcription

1 WIDEBAND SPECTRUM HOLES DETECTION IMPLEMENTATION FOR COGNITIVE RADIOS Ian Frasch and Andres Kwasinski Department of Computer Engineering, Rochester Institute of Technology, NY, USA. ABSTRACT The ability to dynamically discover portions of unused radio spectrum is a central ability of cognitive radios as it allows for dynamic access and sharing of the spectrum. Focusing on implementation lessons, this work steps through the design, implementation, and analysis of a wideband spectrum holes detector. Energy detection and cyclostationary detection algorithms for detecting spectrum holes are compared and the former is seen to perform better in applications with a constrained sensing time. A multiband energy detection algorithm able to scan a gigahertz-order range of frequencies and discover spectrum holes in real time is implemented on a software-defined radio. Resource utilization and requirements of the implementation are analyzed. Experiments are performed on the implementation to measure performance, study practical issues, and demonstrate its detection ability over a wide bandwidth with reasonable latency. Index Terms Cognitive radio, spectrum hole, energy detection, hardware, and real-time implementation. 1. INTRODUCTION The ability to dynamically discover and utilize unused portions of the radio spectrum (holes) leads to more reliable and more efficient uses of the radio spectrum. While regulation of the radio spectrum has caused an apparent spectrum scarcity, this is of an artificial nature since significant portions of licensed spectrum are unoccupied most of the time. Cognitive radios (CRs), which are wireless devices built on softwaredefined radio (SDR) with the ability to implement learning and adaptive algorithms that perform dynamic spectrum access (DSA), are able to find vacant frequency bands and avoid interfering with other users, thus addressing spectrum scarcity while making more efficient use of the radio spectrum [1]. While research that covers theoretical methods for sensing the radio spectrum and detecting holes is plentiful, there is a need for works that focus on the actual implementation and its related issues (especially wideband implementations). This work aims at addressing some of these needs by presenting the design, simulation, implementation, and analysis of a multiband blind spectrum hole detector for wideband CRs. Through simulations, we conclude that an energy detection algorithm is better suited for wideband applications than a cyclostationary algorithm. Following this, we propose a multiband energy detection algorithm and implement it in an SDR. A battery of experiment demonstrates the efficient application of the algorithm for an extremely wide bandwidth and with reasonable latency while also yielding valuable practical implementation and hardware/resource utilization lessons. 2. RELATION TO PRIOR WORK Theoretical research on cognitive radio is abundant, [1], SDRs are commercially available, white space frequencies are legally available, and there is even an IEEE standard [2]. For spectrum holes detection energy detection algorithms for cognitive radios are proposed in [3] and [4]. Cyclostationary detection is studied in [5, 6, 7], to name a few works. The works in [8] and [9] discussed the use of both cyclostationary detection (for fine sensing with a long sensing time) and energy detection algorithms (for coarse sensing with a short sensing time). Despite the large body of work, cognitive radios are still not yet part of everyday mainstream commercial use. We argue that the missing piece of the existing research puzzle is implementation studies. This paper aims at contributing to this incomplete part of the puzzle. 3. BLIND SPECTRUM HOLE DETECTION We focus on the practical implementation of blind spectrum holes detection, this is, which is performed without any a- priori knowledge of the signals to be detected. For this, we first consider the two primary techniques in the literature: energy detection and cyclostationary feature detection. [10] Energy Detection A single-band energy detection (ED) algorithm detects the presence of a signal using a power spectral density (PSD) estimate and does not require a-priori knowledge of the noise power. Instead, assuming that at least one frequency bin is unoccupied by signals during the sensing time, the noise power is estimated as the minimum value of the PSD. The assumption is expected to hold true in implementations where the receiver bandwidth is large enough that typical transmissions would not occupy the entire band. In this work, the target implementation platform has a 28 MHz bandwidth, which was deemed to be wide enough to support this algorithm because it is larger than typical transmissions (e.g. 20 MHz in Wi-Fi or 6-8 MHz in digital TV). The algorithm is as follows: 1. Compute successive N F -point FFTs of the input signal with no windowing functions or overlap between FFTs, X[k] = N F 1 x[n] e j2π k N F n. (1) 2. Compute the power spectral density (PSD) using the average squared magnitude of the FFTs: S xx [k] = E[ X[k] 2 ]/N F. 3. Find the minimum PSD value and use it as an estimate of the noise floor /17/$ IEEE 278 GlobalSIP 2017

2 4. Set the detection threshold T h ED to be the minimum PSD value multiplied by fixed scaling factor SF, chosen based on desired false alarm probability, P F A : T h ED = min{s xx [k]} SF. 5. A hole is detected when max{s xx [k]} < T h ED. For best performance in practice, E[ X[k] 2 ] should be estimated from the average of as many Fourier transforms as possible, but sensing time increases with the number of averages. Also, while energy detection has the advantage of being low in complexity, it presents the drawback that performance is degraded by inaccurate estimates of the in-band noise power, [11, 12]. For multiband spectrum hole detection the algorithm considers individual PSD bins instead of the maximum value of all PSD bins. This means a hole is detected in frequency bin k when S xx [k] < T h ED. Moreover, the single-band algorithm is modified by adding a metric to rank the strength of each spectrum hole or how likely it is to truly be a hole. Then, the multiband algorithm is implemented with the same steps as above with the addition of one extra step at the end, where for each frequency bin that is a hole, the corresponding rank is calculated as hole rank[k] = T h ED /S xx [k] Cyclostationary Detection A cyclostationary detection algorithm was based on [5], which uses the squared magnitude of the spectral correlation density (SCD) as a detection statistic followed by a threshold test to determine if a signal is present. For blind detection, the work in [6] suggests using a crest factor that is calculated from the spectral coherence magnitude. The work in [7] suggests integrating the SCD over a feature mask to detect signals, where the feature mask restricts the integration to the range of values for (α, f) where peaks/features are expected. The considered single-band cyclostationary detection (CD) algorithm is described as follows: 1. Compute successive N F -point FFTs of the input signal with no windowing functions or overlap between FFTs, X NF [n, f]. 2. Compute the scaled PSD (Note: N refers to the number of successive FFTs computed), 3. Compute the SCD, N 1 xx [f] = X NF [n, f] 2. (2) N 1 x α [f] = X NF [n, f + α/2]xn F [n, f α/2]. (3) 4. Compute the squared magnitude of the spectral coherence. S x α [f] 2 C x α [f] 2 = xx [f + α/2] S xx [f α/2]. (4) 5. A hole is detected when max{ C α x [f] 2 } < T h CD Energy Detection versus Cyclostationary Detection We compared single-band energy versus cyclostationary detection algorithms with the goal of selecting the best of the two for further study. The algorithms were compared through simulations when the input is a single baseband signal in AWGN. Single-band means that the only considered output is whether a signal is detected, rather than where the signal is inside the band. In the simulations, both QPSK and OFDM (with 16- QAM symbols) signals were tested. Both input signals have a length of 9216 samples, meaning 72 N F = 128-points FFTs are computed for each algorithm. Importantly, this sets the same sensing time for both algorithms. Other settings for the QPSK signal are 2304 symbols, 4 samples per symbol and an RRC roll-off of 0.5. For the OFDM signal, other settings are 128 symbols, 72 samples per symbol, 38 subcarriers and cyclic prefix length equal to 8 samples. The algorithms performance was evaluated in terms of the probability of detection (P D ) as a function of the Signal-to-Noise ratio (SNR) and the detector receiver operating characteristic (ROC) curve (P D as a function of P F A ). The value of P F A depends only on the detection threshold(s) chosen. Therefore, to measure P D as a function of SNR we set the detection threshold T h ED so that P F A = 0.1. This setting is in consideration that the IEEE standard for cognitive radio networks specifies that P F A should be less than or equal to 0.1 and P D should be greater than or equal to 0.9 [2]. In addition, we considered that the IEEE standard mandates a detection time of 2 seconds or less to meet P D and P F A requirements [2]. The simulation results showed that the energy detection algorithm outperforms the cyclostationary detection algorithm for both QPSK and OFDM for the same fixed signal length. This is because estimating the SCD with finite sensing time results in the noise SCD estimate to be small but not the theoretical zero. While increasing the sensing time improves performance of both cyclostationary and energy detection algorithms it was shown in [7] that cyclostationary feature detection requires longer sensing time to achieve the same performance as energy detection. The simulations corroborated this point. The longer sensing time makes the CD algorithm less desirable for wideband implementation, where the algorithm is applied repeatedly across a range of frequencies to cover a very wide bandwidth. Since further study focuses on multiband spectrum hole detection for wideband applications, in the sequel we consider only the energy detection algorithm. 4. IMPLEMENTATION To study practical issues, the multiband energy detection algorithm was implemented in a Nuand bladerf x115 platform. The bladerf is a SDR that can be tuned from 0.3 to 3.8 GHz and can transmit/receive 28 MHz of instantaneous bandwidth with a 40 Msps sampling rate and 12-bit sampling resolution. The bladerf contains an Altera Cyclone IV E FPGA with 115k logic elements which sits between a Lime Micro LMS6002D field-programmable RF transceiver and a Cypress FX3 USB 3.0 controller. The board s FPGA allows signal processing to be performed in hardware in addition to 279

3 software on the host PC. The detector architecture consists of a software program (running on the Nios II soft processor) for frequency tuning and control, and a hardware datapath for signal processing and hole detection (with all processing on the FPGA as shown in Fig. 1). For each band, the hardware datapath receives IQ samples at 40 Msps (for a sensing time of µs) and performs successive 128-point FFTs. Although sampling at 40 Msps could allow for a larger bandwidth, a lowpass filter in the RF front end reduces the passband bandwidth to 28 MHz (-14 MHz to +14 MHz). Then, the FFT frequency bins outside of the 28 MHz lowpass filter bandwidth are discarded, reducing the number of frequency bins from 128 to 89. The signal is still oversampled at 40 Msps to reduce aliasing. The squared magnitude of the FFT results are averaged (with 256 FFT results) to produce a PSD estimate. The FPGA logic was designed with VHDL and includes two Altera IP-cores (a 128-point FFT and a pipelined divider for calculating the hole ranking). The datapath processes samples in pipelined fashion with one sample per clock cycle. Fixed-point arithmetic was used for all calculations. In order to reduce precision loss, the implementation performs minimal bit reduction on fixed-point values. 39-bit precision was used for each calculated PSD value, and the scaling factor is stored as a 13-bit Q1.12 number. Fig. 1. Implementation Architecture For detection over GHz ranges, the detector partitions the bladerf s entire 0.3 to 3.8 GHz range into 125 frequency bands and sequentially performs the multiband energy detection algorithm at each band. One important measure of performance is the total sensing and processing time for each full scan, called the scan latency, which is calculated as t scan = (t s + t r ) N freq, where N freq is the number of frequency bands, t s is the sensing time per band, t r is the retune time to change from one band to the next, and where additional delay at the end of a full scan due to pipelining in the FPGA is negligible. We determined from the implementation that t r = µs and t s = µs. With N freq = 125, t scan = 169 ms approximately. The implementation detects holes with a resolution of khz. One issue to address is that direct conversion receivers suffer from unwanted baseband DC offset due to imperfections in the analog signal chain. Instead of compensating for DC offset, the detector estimates the PSD value of the DC bin with the average of PSD values at the two adjacent bins. Compensation methods proved ineffective at completed eliminating the DC spike. Another practical issue to consider is that the antenna frequency response is not flat over a 28 MHz band at those bands at the edges of its operating frequency range, thus, the detector should not be applied over these bands. As part of the study of practical implementation issues, Table 1 shows the measured FPGA resource utilization. For the Altera Cyclone IV E FPGA in the bladerf x115, the resource use was 8.8% of the total available. Table 1. FPGA Resource Utilization Usage Resource Nios + Datapath Total Interfaces 4-input LUT bit Register x18 Multiplier Kb Memory Block Implementation source code, including both VHDL hardware description and C software, has been provided online at 5. EXPERIMENTS AND RESULTS In this Section we discuss practical issues and results from experiments performed to assess performance. When running the detector with no antenna connected (only noise), we noted spurious tones (one or two for every 28 MHz band) in the PSD (with values 5-15 db above the noise floor) that were separate from the expected DC offset spikes at each frequency band. This issue was also seen in [13]. While the origin of the tones is unknown, it is expected to come from imperfections in the analog circuitry in or around the RF front end. Although the number of tones that could be discerned from background noise decreased drastically after attaching an antenna, each tone could be detected as a signal, adding to the number of false alarms. Also, testing revealed that the magnitude of the frequency response of the lowpass filter in the RF front end is around 1 db higher at the cutoff frequency than at DC. This issue, also noted in [13], results in a higher detection probability for frequencies near the cutoff frequency, and lower detection probability for frequencies around DC. While this issue could be addressed by scaling the PSD values according to the filter response it is not included in the results reported herein. Because it affects the energy detection performance, an experiment was done to study how varying noise levels affect noise power estimation. With the receive antenna unattached and replaced with a 50-ohm terminator, internal noise power from the RF front end varied by 20 db between the lowest frequency (0.3 GHz) and the highest frequency (3.8 GHz). Simply attaching the antenna or adjusting its orientation changed the observed noise power by db at certain frequencies. Therefore, we concluded that a dynamic estimation of the noise floor is needed for realistic systems. Figure 2 illustrates the setup in the experiments discussed next. The setup, intended to reproduce wireless communications without actually generating radio emissions in nonpermitted bands, consisted of three bladerf SDR devices transmitting through SMA cables to one receiver bladerf performing hole detection. The three transmitted signals were added using a 4:1 RF combiner. After the combiner, two 280

4 30 db attenuators were applied to protect the receiving device and reduce SNR. To yield longer signals, the experiments used the same QPSK and OFDM test signals used earlier but the number of QPSK symbols increased to for a signal length of samples and the number of OFDM symbols was also increased to 896 for a signal length of samples. A sampling rate of 15 Msps was used for transmission of the QPSK signal, resulting in a signal bandwidth of 5 MHz and the lowpass filter in the transmit chain of the RF front end was set to a bandwidth of 12 MHz. For transmission of the OFDM signal, a sampling rate of 24 Msps was used, resulting in a signal bandwidth of 16.5 MHz. The bandwidth of the lowpass filter in the transmit chain of the RF front end was set to 20 MHz. Each signal transmission was repeated indefinitely to result in a continuous transmission. Fig. 3. Detector performance for a single continuous transmission and fixed P F A = 0.09 Fig. 2. Experimental Setup The detector was configured to scan frequencies from 1308 MHz to 2288 MHz for a total sensing bandwidth of 980 MHz over 35 bands. The scaling factor was set to 1.455, producing a measured P F A of Note that for a multiband detection algorithm, the P F A is defined on a per-bin basis, meaning P F A is the probability for an individual bin to be detected as a signal when no signal is present (e.g. when the input is pure noise, the average number of frequency bins per band incorrectly detected as signals would be N F P F A, where N F is the number of frequency bins per band). The latency of each full scan was measured equal to 47.3 ms, which yields an average retune time of µs. A first experiment considered transmission from a single bladerf device. To address the non-flat frequency response of the front end filter, two different carrier frequencies were used. One carrier frequency was placed at the center of one of the detector s frequency bands, and the other carrier frequency was placed at the edge of one of the detector s frequency bands. The carrier frequencies used were 2134 MHz and 2148 MHz. Performance was measured in terms of the probability of detecting at least 90% of the power in the signal s bandwidth (P D,90 ). For each test signal (QPSK and OFDM), the detection probability P D,90 is the average of the two carrier frequencies. Figure 3 shows the resulting P D,90 as a function of in-band SNR. The higher P D,90 seen for the QPSK signal was attributed to the wider bandwidth of the OFDM signal (16.5 MHz) in comparison to QPSK signal (5 MHz), which leaves fewer free bands for a more accurate noise floor estimation. Figure 4 shows the resulting receiver operating characteristic (ROC) curve of the detector with inband SNR fixed at -6 db. The ROC curve was generated by Fig. 4. Detector ROC curve for a single continuous transmission and fixed in-band SNR = -6 db varying the detector s scaling factor and measuring each corresponding P D,90 and P F A value. A second experiment considered continuous transmissions from three bladerf devices. The three frequencies used were MHz, MHz, and MHz. Tests for QPSK signals and OFDM signals were kept separate. The transmit gain in each radio was adjusted so that all transmissions had the same in-band SNR (+/- 0.5 db). Performance was measured through the probability of detecting at least 90% of all ). Figure 5 shows the resulting PD,90 all as a function of in-band SNR. Interestingly, the performance of the detector with the OFDM signals is about the same as the performance with the QPSK signals. signals at any given time (P all D,90 6. CONCLUSION This paper contributes to a better understanding of implementation feasibility of spectrum hole detection in CRs by presenting the design, simulation, implementation, and analysis of a blind spectrum hole detector for cognitive radio applications. For wideband applications, simulations concluded that the performance of the energy detection algorithm was significantly higher than a cyclostationary algorithm. A multiband energy detection algorithm suitable for wideband implementation was proposed. The algorithm implementation demonstrated efficient application for an extremely wide bandwidth and with reasonable latency. Practical implementation issues and hardware/resource requirements were described in detail. Fig. 5. Detector performance for three continuous transmissions 281

5 7. REFERENCES [1] S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp , Feb [2] I. S. Association, Cognitive Wireless RAN Medium Access Control and Physical Layer Specifications: Policies and Procedures for Operation in TV bands, IEEE Standard , p. 447, Jun [3] M. López-Benítez and F. Casadevall, Improved energy detection spectrum sensing for cognitive radio, IET communications, vol. 6, no. 8, pp , [4] S. Atapattu, C. Tellambura, and H. Jiang, Energy Detection Based Cooperative Spectrum Sensing in Cognitive Radio Networks, IEEE Transactions on Wireless Communications, vol. 10, no. 4, pp , April [11] A. Mariani, A. Giorgetti, and M. Chiani, Effects of Noise Power Estimation on Energy Detection for Cognitive Radio Applications, IEEE Transactions on Communications, vol. 59, no. 12, pp , December [12] B. Shent, L. Huang, C. Zhao, Z. Zhou, and K. Kwak, Energy Detection Based Spectrum Sensing for Cognitive Radios in Noise of Uncertain Power, in 2008 International Symposium on Communications and Information Technologies, Oct 2008, pp [13] S. M. Mishra, S. ten Brink, R. Mahadevappa, and R. W. Brodersen, Cognitive Technology for Ultra- Wideband/WiMax Coexistence, in nd IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, April 2007, pp [5] S. Enserink and D. Cochran, A cyclostationary feature detector, in Proceedings of th Asilomar Conference on Signals, Systems and Computers, vol. 2, Oct 1994, pp vol.2. [6] K. Kim, I. A. Akbar, K. K. Bae, J. S. Um, C. M. Spooner, and J. H. Reed, Cyclostationary Approaches to Signal Detection and Classification in Cognitive Radio, in nd IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, April 2007, pp [7] A. Tkachenko, D. Cabric, and R. W. Brodersen, Cyclostationary Feature Detector Experiments Using Reconfigurable BEE2, in nd IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, April 2007, pp [8] S. Maleki, A. Pandharipande, and G. Leus, Two-stage spectrum sensing for cognitive radios, in 2010 IEEE International Conference on Acoustics, Speech and Signal Processing, March 2010, pp [9] W. Yue and B. Zheng, A two-stage spectrum sensing technique in cognitive radio systems based on combining energy detection and one-order cyclostationary feature detection, in Proceedings of the 2009 International Symposium on Web Information Systems and Applications (WISA09), 2009, pp [10] T. Yucek and H. Arslan, A survey of spectrum sensing algorithms for cognitive radio applications, IEEE Communications Surveys Tutorials, vol. 11, no. 1, pp , First