Results and future prospects of ENCOR. Enabling Methods for Dynamic Spectrum Access and Cognitive Radio

Transcription

1 Trial Program Results and future prospects of ENCOR Enabling Methods for Dynamic Spectrum Access and Cognitive Radio Lauri Anttila, Ali Shahed, Markus Allen, Jaakko Marttila, Mikko Valkama Jarmo Lunden, Sachin Chaudari, Jan Oksanen, Visa Koivunen Marko Kosunen, Vesa Turunen, Jerry Lemberg, Semu Mäkinen,Jussi Ryynänen Ahmet Gokceoglu, Sener Dikmese, Markku Renfors

2 Outline ENCOR Overview Technical Progress in short International Collaboration National Collaboration Examples of Technical Work and Results Example Publications Project Continuation and Future Prospects Project Website and Contact Information

3 ENCOR- project overview Focus on concept design, signal processing methods and radio circuits in dynamic spectrum access and cognitive radio networks with strong demonstration emphasis. The goal of the project is to develop new hardware and algorithm solutions for cognitive radios and verify their functionality in real radio environments. I AGC LPF A / D BPF LNA LO frf DSP Q AGC LPF A / D

4 ENCOR- project overview I AGC LPF A / D BPF LNA LO frf DSP Q AGC LPF A / D Harware nonideality compensation with DSP Algorithm development and research Algorithm implementations Field tests/ applications

5 ENCOR results in short Spectrum exploration and exploitation Developed new decentralized/distributed detection methods for reliably identifying available spectrum and analyzed their performance. Developed new collaborative learning methods for efficiently exploiting the identified free spectrum Developed jointly optimized sensing and access policies based on restless multiarm bandit (RMAB) model and proofed that they have logarithmic regret Developed wideband radio scene analysis tools and circuits for wideband sensing and detection of reappearing primary users Developed Bayesian detection based spectrum sensing algorithms

6 ENCOR results in short Energy-efficient spectrum sensing implementations Development of power efficient detector implementations Analysis of the benefits from using multiple detectors simultaneously Adaptive energy detection in coarse-fine spectrum sensing scheme Hardware originated interference and distortion problems Modeled and analyzed hardware originated interference and nonidealities for state-of-the-art spectrum sensors Developed simple but efficient hardware and algorithm solutions to mitigate hardware originated interferer and distortion problems

7 ENCOR results in short (cont d) Reconfigurable building blocks for Cognitive Radio receivers Developed new multiband quadrature sigma-delta converter solutions for cognitive radio A/D interfance Developed new algorithm solutions for linearization of cognitive radio receiver chain, including nonlinearities of A/D interface Developed concept solutions for simultaneous reception and sensing Some conrete examples are given in the following slides Only selected snapshots without any details For more details, please see contact information at the end of this slide set Also massive set of publications available

8 National Collaboration ENCOR is implemented together with Aalto and TUT Industrial members in the steering board are Nokia and PV Project is also linked directly with Nokia s CRT project There is an explicit signed consortium and dissemination agreement between Aalto, TUT and Nokia This enables ENCOR to use, e.g., Nokia s spectrum sensor platform for demonstration and verification purposes (true radio HW) ENCOR has also collaboration agreement with Elektrobit (EB) Demonstrations and radio HW studies in the EB s RACE environment and EB-Trial Contentwise, we also have direct linking with WISE project, e.g. in wireless microphones use case The work is also be complemented by the NSF-Tekes funded project Reconfigurable Antenna-based Enhancement of Dynamic Spectrum Access Algorithms (READS/RADSA) under WiFiUS umbrella (in collaboration with Drexel, VTT and Univ. Oulu)

9 International Collaboration The key international partners are UCLA, UC Berkeley, Princeton University, and TU Karlsruhe Builds on extensive long-term researher exchange and joint research Examples: UCLA: Dr. Ali shahed (TUT) 18 months with Prof. Cabric, ongoing UC Berkeley: Dr. Marko Kosunen (Aalto) 6 months with Prof. Nikolic Princeton: M.Sc. Jan Oksanen (Aalto), 12 months with Prof. Poor, ongoing TUT: MSc. Georg Vallant (TU Karlsruhe) 3 months with Prof. Valkama TU Karlsruhe: MSc. Markus Allen (TUT) 5 months with Prof. Jondral Plus national researcher exchange, e.g. Jaakko Marttila spent May-June at Aalto Also many professor-level research visits, e.g. Prof. Visa Koivunen at Princeton University, 5 months The work will also be complemented by the collaboration with Drexel university in READS/RADSA project VTT and Univ. Oulu)

10 Some Selected Snapshots of Research Outcomes

11 Collaborative Sensing and Spectrum Access In collaboration with Princeton Research Snapshot 1 Co-operation among multiple spectrum sensors in order to jointly optimize identifying and exploiting idle spectrum Local sensors build on cyclostationary features of communications waveforms, fusion center accumulates local measurements towards spectrum access grants Local sensors can report either hard or soft sensing data Impact of reporting errors have also been analysed Global optimization in spectrum access can be emphasizing, e.g., the total sumrate of the cognitive network or the fairness between opportunistic devices, and also imposing interference control towards primary radio system

12 Collaborative Sensing and Spectrum Access (cont d) In collaboration with Princeton Research Snapshot 1 Jointly optimized sensing and access policies using restless multiarm bandit problem (RMAB) formulation. Logarithmic regret of the developed policy has been proofed. In general, in many cases, very coarse reporting is actually sufficient with sophisticated fusion and learning rules in the fusion center Modern machine learning methods are applied to efficiently identify and access time-frequency-location varying spectral resources.

13 Example: Logarithmic regret of jointly optimized sensing and access policy

14 Bayesian Detection based Spectrum Sensing Research Snapshot 2 Classical (frequentist) detectors do not deploy/assume any prior information about the target(s) Such information can, however, be available through spectral measurement campaigns; e.g. probabilistic models for spectrum utilization Bayesian detectors can utilize such information and thus improve spectrum sensing performance against classical solutions (like plain local energy detection) This could be seen as one tool trying to bridge the gap between local sensing based and database based approaches General Bayesian detector formulated for the spectrum sensing task, taking also noise and/or SNR uncertainty into account

15 Research Snapshot 2 Bayesian Detection based Spectrum Sensing (cont d) As an example, compared to classical energy detection, this allows reducing the sensing time especially in good SNR conditions Stemming from proper uterative use of Bayes Odds Ratio, accumulated over time when new samples arrive

16 PSD Beyond-FFT methods for efficient wideband radio scene analysis and spectrum sensing Developing frequency domain optimized filter bank solutions providing high-resolution, high dynamic range spectrum analysis tool Can be utilized for very flexible, multiband spectrum sensing for different PU waveforms Can be used as a highresolution spectrum analyzer tool, independently of waveform PU1 SH PU2 Research Snapshot 3 PU3 OFDM FBMC Normalized frequency [ / ] SH

17 Analog FFT (A-FFT) based spectrum sensor for 500 MHz instantaneous BW In collaboration with UCLA Research Snapshot 4 Targeting to extremely wide instantaneous bandwidths in the order of 500 MHz Classical solutions unfeasible Idea: use analog FFT (e.g. 8 bins ) for coarse channelization, followed by digitization Further sensing processing then inside each bin Wideband RF Frontend S/P Ideal FFT Err. 1: Twiddle Factor Analog FFT Err. 2: Charge Sharing A/D A/D A/D A/D X 1 ' X 8 ' Digital Correction H T x 1 x 8 Massive effort overall; our work is related e.g. to digital calibration and correction of analog FFT imperfections

18 Research Snapshot 4 Analog FFT (A-FFT) based spectrum sensor for 500 MHz instantaneous BW In collaboration with UCLA Current work at RTL level, target for tape out is Q4/2012-Q12013 Wideband RF Frontend Analog FFT A/D A/D A/D A/D X 1 ' X 8 ' Digital Correction x 1 x 8 X 1 ' X 2 ' X 8 ' Coef D Coef D Coef D D D D Output Register X 1 X 2 X 8 40nm Area: 0.1 mm 2, Power: 10.1 mw Bandwidth: 500 MHz H T 8th row of H T Comparison: A-FFT approach vs. only ADC FFT ADC Cal. Total N A-FFT [bin] MS/s Qty bits mw mw mw N/A

19 RF originated distortion and interference problems HW originated interference and distortion cause false detections on unoccupied frequency bands. Sources of HW originated interferers: Nonlinearity, especially 2nd and 3 rd order IQ imbalance Clock feed-through and phase noise etc Due to high sensitivity requirement, nonidealities of the RF front-end can be a remarkable source of false detections, especially under presence of strong out-ofband interferers. Research Snapshot 5

20 RF originated distortion and interference problems Example: Digital Intermodulation Suppression of DVB-T / WLAN Spectrum Sensing Receiver Cognitive radio spectrum sensing platform designed by Nokia 3nd order nonlinearity and spectral regrowth identified as a problem (Aalto) Interference on baseband has also cyclostationary features causes false alarms Research Snapshot 5 Digital compensation with reference nonlinearity (third power) and LMS weighted interference subtraction Spreading because of nonlinearity

21 RF originated distortion and interference problems Research Snapshot 5 Example: Digital Intermodulation Suppression of DVB-T / WLAN Spectrum Sensing Receiver With the developed digital cancellation method, the interference is efficiently removed from the vacant WLAN channel (20 MHz bandwidth)

22 Research Snapshot 6 Power efficient Adaptive Coarse-Fine (hybrid) Spectrum sensing Energy detector is assisted by cyclostationary feature detector Energy (coarse) detector: Fast and efficient Vulnerable to noise level uncertainty Unable to differentiate between systems May detect also hardware originated interferers Cyclostationary (fine) feature detectors Increased complexity Insensitive to noise level uncertainty and some hardware originated interferers. Known detection SNR threshold SNR can be evaluated from calculated test statistics.

23 Research Snapshot 6 Detector development FPGA implementation properties Power diss. [mw] Logic eleme nts Regist ers Memory Energy detector FFT-based cyclostationary feature detector Time domain cyclostationary detector Spatial sign Cyclic Correlator Spatial sign Cyclic Correlator with angular domain computation Evaluated with FPGA implementation

24 Research Snapshot 6 Benefits of using multiple detectors Sensing without energy detector, WLAN, DVB-T and free channel Black=occupied White=free Grey=no decision Sensing with energy detector, WLAN, DVB-T and free channel Fast energy detector decreases the detection time on occupied channels Detection efforts can be concentrated on likely free channels Detection time on likely free channels is only dependent on a number of possible primary systems on channel (and implementation of the detectors)

25 Research Snapshot 6 Adaptive energy detection In real application, noise and interferers are not stationary. (Gain, noise figure, temperature, adjacent channel leakage, nonlinearity, feed-through and coupling etc.) Changes in noise and interferer power degrades the reliability of energy detector. Number of parameters affecting reliability is large. Single-run calibration not sufficient for compensation due to variability of parameters. Problem can be alleviated with adaptive decision threshold of the energy detector. Adaptive parallel sensing, WLAN, DVB-T and free channel

26 Research Snapshot 6 Adaptive energy detection Problem: Energy detector can not differentiate if the increase of energy is due to signal or due to change in noise power. Solution: Decision threshold of the energy detector tracks the changes in noise power through the SNR estimate. Outcome: Energy detector is adaptively performing detection on desired SNR level, rather than on desired noise power estimate level. SNR estimate can be obtained from the test statistics of the cyclostationary feature detector. Open question: Required rate of adaption? Adaptive parallel sensing, WLAN, DVB-T and free channel Desired detection SNR

27 Research Snapshot 6 Adaptive energy detection Case DVB-T Detector Power [mw] Samples Rate [MS/s] Time/detectio n [ms] Energy detector Energy/detection [uj] Spatial sign Cyclic Correlator with angular domain computation Adaptive parallel sensing, WLAN, DVB-T and free channel In stable state, the detection of signals is performed mainly with the energy detector. In DVB-case, this results in reduction of energy per detection by factor of 400 or increase of detection rate by factor of 256.

28 DSP-enhanced A/D interfaces for Cognitive Radio receivers Research Snapshot 7 Here we study A/D interface signal conditioning problems and nonlinearities in cognitive radio receiver context, assuming Wideband radio front-end with only coarse selectivity Desired signal spectral chunks can be scattered over wide frequency range in a non-contiguous manner No control over signal energy in-between interesting spectral slices Still trying to cope with one radio chain and one A/D interface (within reasonable total bandwidth, though) I AGC LPF A / D BPF LNA LO frf DSP Q AGC LPF A / D

29 Digital Output Codeword DSP-enhanced A/D interfaces for Cognitive Radio receivers (cont d) Research Snapshot 7 A/D interface nonlinearities is one element in defining the overall receiver spurious-free dynamic range Examples are, e.g., integral and differential nonlinearity and even clipping Ideal Levels Nonindeal Levels INL = -1.0 LSB INL = -0.5 LSB 0.5 LSB, DNL = -0.5 LSB 1 LSB, DNL = 0 LSB 0 LSB, DNL = -1.0 LSB (missing code) 1.5 LSB, DNL = 0.5 LSB 0 1/8 2/8 3/8 4/8 5/8 6/8 7/8 FS Clipping Analog Input Voltage Clipping

30 DSP-enhanced A/D interfaces for Cognitive Radio receivers (cont d) Research Snapshot 7 To suppress the intermodulation distortion, especially due to strong neighboring carriers, several methods have been developed joint iterative detection & interference suppression based methods Adaptive nonlinear interference cancellation based methods Parametric behavioral modeling and parameter identification based methods Details skipped, just one spectral illustration given below for one of the techniques 3rd order distortion suppressed by 17 db

31 DSP-enhanced A/D interfaces for Cognitive Radio receivers (cont d) Research Snapshot 7 Another performance example with parametric behavioral modeling based approach where parallel Hammerstein is used for dynamic nonlinearity modeling and correction in subsampling ADC s In collaboration with TU Karlsruhe Measured results with Linear Technologies LTC2208, 16-bit, 130 Msps; at 3 rd and 5 th Nyquist zone In collaboration with TU Karlsruhe

32 Reconfigurable multiband quadrature sigma-delta converters for CR Targeting to simultaneous multiband reception with frequency-agile reconfigurability in terms of bandwidths and center (IF) frequencies Research Snapshot 8

33 Reconfigurable multiband quadrature sigma-delta converters for CR (cont d) Tool is wideband RF front-end followed by higher-order quadrature (I/Q) sigma delta converter Single-stage vs. multi-stage sigma-delta structures Parameterization for reconfigurability, optimization and response analysis Clever designs against implementation impairments Can also build part of receiver selectivity inside the sigma delta loops Research Snapshot 8 Sigma-delta signal and noise responses can be tuned according to the signal conditions at hand Principal total NTF Principal total STF

34 Reconfigurable multiband quadrature sigma-delta converters for CR (cont d) Second-order sigma-delta block as the main building block, overall structure is a multistage connection of these building blocks Analysing the signal and noise responses in closed form Converter coefficient design and optimization to facilitate (i) noncontiguous spectrum, (ii) reconfiguration on the fly, (iii) robustness against implementation imperfections Research Snapshot 8

35 Reconfigurable multiband quadrature sigma-delta converters for CR (cont d) Research Snapshot 8 An example with 128 MHz sample rate and two non-contiguous frequencies

36 Example Journal Publications J. Oksanen, J. Lunden, V. Koivunen, Reinforcement learning based sensing policy optimization for energy efficient cognitive radio networks, Neurocomputing, Vol. 80, pp , S. Chaudhari, J. Lundén, V. Koivunen, and H. V. Poor, Cooperative sensing with imperfect reporting channels: Hard decisions or soft decisions?, IEEE Trans. Signal Processing, Vol. 60, No. 1, pp Lunden J, Koivunen, V., Poor, H.V. Spectrum Exploration and Exploitation Chapter 5 in Principles of Cognitive Radio, editors E. Biglieri, A. Goldsmith, L. Greenstein, N. Mandaym and H.V. Poor, 72 pages, Cambridge University Press J. Marttila, M. Allen, and M. Valkama, Multi-stage quadrature sigma-delta modulators for reconfigurable multi-band analog-digital interface in cognitive radio devices, EURASIP Journal Wireless Communications and Networking (Special Issue on Ten Years of Cognitive Radio), DOI: / M. Allen, J. Marttila and M. Valkama, Iterative signal processing for mitigation of analog-to-digital converter clipping distortion in multiband OFDMA receivers, Journal Electr. Comp. Eng., Volume 2012, Article ID , 16 pages, doi: /2012/ J. Marttila, M. Allen, and M. Valkama, Frequency-agile multi-band quadrature sigma-delta modulator for cognitive radio: Analysis, design and digital postprocessing, IEEE Journal on Selected Areas in Communications, accepted after minor revision, to appear, 2012.

37 Field test example All the reserved channels (44, 46, 53) were detected in every location with 100% probability False detections due to nonlinearity on adjacent channels Visible mostly in open areas Reduced by reducing RF front-end gain or bandwidth of the detector (as predicted)

38 Field measurement database Database for RSSI and test statistics with time and location information for cooperative sensing purposes Support for wireless microphone systems can be added Current status: Openly accessible db under developement.

39 ENCOR-Clarifying The Big Picture Probability of detection I AGC LPF A / D BPF LNA LO frf DSP Q AGC LPF A / D Algorithm development and research Harware nonideality compensation with DSP Algorithm implementations Field tests/ applications

40 Contact Information and Project Website Aalto: Prof. Visa Koivunen, Dept. Signal Processing and Acoustics, TUT: Prof. Mikko Valkama, Dept. Communications Engineering, Project website: