A Novel Clustering Algorithm Based on Information Geometry for Cooperative Spectrum Sensing

Spectrum sensing is used to detect whether primary user are using authorized spectrums, which can be regarded as a key and core issue for opportunistic spectrum access in cognitive radio networks. In traditional information theory and clustering algorithm-based spectrum sensing methods, they need to evaluate noise environment for constructing a reference point. However, the reference point is fixed, which is unreasonable in dynamic cognitive radio environment. Moreover, these methods convert signal feature from manifold onto Euclidean space, which will cause to overall performance degradation, since sensing information losses. To address this problem, an information geometry (IG)-based K-means clustering algorithm, namely IGK, is developed, it clusters samples on manifold instead of Euclidean space with an unsupervised way to train a classifier for spectrum sensing. Specifically, secondary users observe and collect data from a selected authorized spectrum, which needs to be detected, and send these sensing data to a fusion center (FC). Then, the FC transforms these data into samples on the manifold to obtained a classifier by using the proposed IGK algorithm. According to the trained classifier, we can get related result of the authorized spectrum. Finally, in simulation section, the effectiveness of the proposed scheme is verified under different conditions.

I. Introduction

The development of wireless devices and communications, along with spectrum scarcity during past decades, has triggered the development of cognitive radio (CR) technique [1]–[5]. This technique includes spectrum sensing, spectrum decision, spectrum sharing, and spectrum mobility, which is foreseen as technique capable of improving spectrum utilization [6], [7]. As the first step of CR technique, the work of spectrum sensing is to accurately judge states of primary users (PUs) and find spectrum holes for subsequent operations. Over the past few years, many research works have been reported to spectrum sensing, such as energy detection, cyclostationary detection, and matched filtering detection. The energy detection algorithm is simple but its computational complexity is low. However, it is greatly affected by noise uncertainty [8]. The cyclostationary detection has better performance than energy detection, however, it has higher computational complexity [9]. Matched filtering has the best performance when priori information of PU signal is known [10]. However, the priori information is difficult to acquire in practice. It should be noted that the aforementioned research works used single spectrum sensing techniques, which have poor performance in channel fading and shadow environment [11].