A Combination of Quickest Detection with Oracle Approximating Shrinkage Estimation and Its Application to Spectrum Sensing in Cognitive Radio Feng Lin, Zhen Hu, Robert C. Qiu Department of Electrical and Computer Engineering, Center for Manufacturing Research, Tennessee Tech University, Cookeville, TN, USA Email: fenglin@ieee.org, zhu@tntech.edu, rqiu@tntech.edu Michael C. Wicks Sensor Systems Division, University of Dayton Research Institute, Dayton, OH, USA Email: Michael.Wicks@udri.udayton.edu Abstract—Spectrum sensing is a fundamental problem in cog- nitive radio. How to sense the presence of primary user promptly in order to avoid the unexpected interference is a key issue to the system. The motivation of our work is to detect the primary user signal using small size data in short time. In this paper, a quickest detection based approach is proposed for spectrum sensing. This approach employs covariance matrix estimation instead of sample covariance matrix as the first step, then the core idea of sequential detection or quickest detection is borrowed and utilized here to improve the performance of traditional eigenvalue based MME and AGM detectors. The main advantage of the proposed approach is that it requires short data to detect quickly and it works at lower SNR environments than some traditional methods. A performance comparison between the proposed approach and other traditional methods is provided, by the simulation on captured digital TV (DTV) signal. The simulation results show this proposed approach exhibits performance improvement while the threshold keeps robust. I. I NTRODUCTION Spectrum sensing is the key component for cognitive radio which is the next generation wireless communication system. Spectrum sensing tries to find the spectrum hole available for the unlicensed cognitive radio. Based on the results of spec- trum sensing, cognitive radio can perform dynamic spectrum access to make the spectrum utilization efficient. For traditional sample covariance matrix based spec- trum sensing methods, e.g., maximum-minimum eigenvalue (MME) [1], arithmetic-to-geometric mean (AGM) [2], func- tion of matrix detection (FMD) [3], [4], and so on, sample covariance matrix based on large-size sensed data is used to derive detection statistics. However, due to the short time dura- tion of spectrum hole, the accurate and quick spectrum sensing is the prerequisite of high-performance cognitive radio. If we can shorten the time needed for spectrum sensing, more time can be exploited and allocated for the reaction of cognitive engine and dynamic spectrum access. The straightforward challenge for quick spectrum sensing is how to guarantee the detection performance of spectrum sensing using small-size sensed data. This paper tries to address this issue and give the potential solution to the quickest spectrum sensing in the low SNR situation. From the statistical point of view, the sample covariance matrix is the maximum likelihood estimation of true covari- ance matrix when sample size goes to infinity. However, if sample size is small or the dimension of samples is in the same order of sample size, the sample covariance matrix can behave very badly, which cannot describe the accurate statistical relationship within each sample. Quickest detection [5] tries to detect the change of two different random processes with the shortest delay. If the change happens at the beginning of spectrum sensing, the goal of quickest detection is similar to that of sequential detection. The advantage of this paper is threefold: • The proposed method can detect the PU signal quickly without knowing the statistical distributions of PU signal or noise. • Given the number of data samples, the proposed method works at lower SNR environment compared with some other traditional approaches. • The detection threshold of the proposed quickest spec- trum sensing is invariant to the SNR and sample size. The rest of the paper is organized as follows. Section II gives the background knowledge of spectrum sensing together with the well-studied spectrum sensing methods. In section III, the quickest spectrum sensing with small-size sensed data are presented. Simulation results and performance evaluations are given in section IV followed by some remarks in section V. II. SYSTEM MODEL AND PRIOR SOLUTIONS A. Binary Hypothesis In a secondary network, we consider each secondary user (SU) with one receive antenna to detect one primary user (PU) signal based on its own observation. Let x(t) be the continuous-time received signal after unknown channel. Let T s be the sampling period, the received signal sample is x [n] = x (nT s ). There are two hypotheses to detect PU signal’s existence, H 0 , only noise (no PU signal) exists; and H 1 , both PU signal and noise exist. The received signal samples under the two hypotheses are given respectively as 978-1-4673-3/12/$31.00 ©2013 IEEE 978-1-4673-3/12/$31.00 ©2013 IEEE