Machine Learning Based Spectrum Sensing In Cognitive Radio

The useable electromagnetic radio spectrum, a scarce resource in itself, has beenexclusively allocated to different wireless services by government and regulatory bodies,though the licensed part of the radio spectrum is poorly utilized. However, there are moreand more wireless devices and applications throughout the world. Therefore, the dilemmaof the scarcity of spectrum resources and the expanding demand for them has become aserious concern.Cognitive radio (CR) is a novel paradigm of wireless communications and is re-garded as a key technology for solving low efficiency in spectrum utilization. Accord-ing to the definition of CR by the Federal Communications Commission (FCC), CR is awireless communication system with intelligence that senses its outside electromagneticenvironment and learns from the surroundings, then adapts its internal states by changingcertain operating parameters such as transmit-power, carrier frequency, and modulationstrategy in order to adapt to the changes of its environment.Spectrum sensing plays a primary and important role in CR. The aim of spectrumsensing is to detect a spectrum hole precisely in order to share the spectrum without harm-ful interference to other users. The term"spectrum hole"refers to a band of frequencieswhich are not being occupied at a particular point of time and specific geographic locationby a licensed user (also known as the primary user). In order to guarantee the unlicenseduser (also known as secondary user) access the spectrum hole reliably and to utilize theband effectively without interference to the primary user, various detection methods havebeen discussed. The spectrum sensing techniques for CR are classified into three cate-gories: transmitter detection, cooperative detection, and interference based detection. Inthis paper, we focus on transmitter detection, which mainly includes energy detection,cyclostationary feature detection and matched filter detection.In this thesis, spectrum sensing is studied from the machine learning perspectiveand an artificial neural network based spectrum sensing method is proposed. Artificialneural network is introduced into spectrum detection in our approach, with which theenergy detection and cyclostationary feature detection are combined. Previous to neuralnetwork training, the signal information of samples from the primary user is preprocessed, while four feature values are extracted simultaneously. Through neural network training,learning and accumulation of signal information from the primary user is implemented.Finally, spectrum sensing of the received signal is processed to detect whether there is aprimary user. Simulations based on AM and OFDM signals are presented to demonstratethe superiority of the proposed scheme over conventional spectrum sensing methods.