[Paper Review] Demonstration of Spectrum Sensing with Blindly Learned Feature

Peng Zhang, Robert C. Qiu+1 more|arXiv (Cornell University)|Feb 24, 2011

Blind Source Separation Techniques6 references5 citations

TL;DR

This paper proposes a blind feature learning algorithm (FLA) and a feature template matching (FTM) detector that use the leading eigenvector of the signal covariance matrix as a learned, local feature for spectrum sensing in cognitive radio. Implemented on an SDR platform, FTM achieves approximately 3 dB gain in detection performance over the blind CAV detector, demonstrating that blindly learned features can significantly improve sensing accuracy without prior knowledge of signal parameters.

ABSTRACT

Spectrum sensing is essential in cognitive radio. By defining leading extit{eigenvector} as feature, we introduce a blind feature learning algorithm (FLA) and a feature template matching (FTM) algorithm using learned feature for spectrum sensing. We implement both algorithms on Lyrtech software defined radio platform. Hardware experiment is performed to verify that feature can be learned blindly. We compare FTM with a blind detector in hardware and the results show that the detection performance for FTM is about 3 dB better.

Motivation & Objective

• To address the challenge of spectrum sensing in cognitive radio under low SNR, noise uncertainty, and unknown channels.
• To develop a method for learning signal features blindly without prior knowledge of primary user signals.
• To improve detection performance by using the leading eigenvector of the signal covariance matrix as a robust, local feature.
• To implement and validate the proposed algorithms in a real-world SDR hardware platform.

Proposed method

• The leading eigenvector of the sample covariance matrix of received signals is defined as the signal feature, leveraging its stability and noise robustness in non-white WSS signals.
• A blind feature learning algorithm (FLA) is proposed to estimate the signal feature by measuring similarity between consecutive sensing segments using a threshold-based similarity metric.
• A feature template matching (FTM) detector is developed that uses the learned leading eigenvector as prior knowledge to improve detection performance.
• The FTM detector computes a test statistic based on the projection of received signal vectors onto the learned feature vector, enhancing detection under low SNR.
• The algorithms—FLA, FTM, and CAV—are implemented on a Lyrtech SDR platform using FPGA for covariance computation and DSP for eigenvector and test statistic calculation.
• Fast-PCA (FPCA) is employed to reduce computational complexity, enabling real-time implementation with N=32 and Ns=2^20.

Experimental results

Research questions

• RQ1 Can the leading eigenvector of the signal covariance matrix be learned blindly in a real-time, hardware-implemented cognitive radio system?
• RQ2 Does using a blindly learned signal feature as prior knowledge improve spectrum sensing performance compared to fully blind detectors?
• RQ3 How does the detection performance of the proposed FTM detector compare to the theoretical upper bound (EC) and lower bound (MME/CAV) in both simulation and hardware?
• RQ4 What is the achievable gain in SNR performance when using learned features versus blind detection in a non-line-of-sight indoor environment?
• RQ5 Can the learned feature remain stable over time and across varying propagation conditions in a real-world setting?

Key findings

• The leading eigenvector of the signal covariance matrix is a stable and robust feature that can be learned blindly in real time, even under low SNR and unknown channel conditions.
• In hardware experiments, the similarity between the first and last sensing segment features was 94.3%, confirming feature stability over 20 seconds in an NLOS indoor environment.
• The FTM detector achieved a 3 dB improvement in detection performance over the blind CAV detector, requiring only -119 dBm received power to reach 100% detection probability, compared to -116 dBm for CAV.
• Simulation results show that FTM performance lies between the theoretical upper bound (EC) and lower bound (MME/CAV), with FTM requiring -18 dB SNR to achieve 100% detection probability, compared to -16 dB for MME/CAV.
• The implementation using Fast-PCA achieved real-time performance with a computation time of less than 20 ms for eigenvector calculation, demonstrating feasibility for practical SDR deployment.