COMPLEX RADIO SCENE ANALYSIS: SUPERVISED AND UNSUPERVISED MACHINE LEARNING APPROACHES

Abstract

The analysis of radio frequency (RF) scenes is a critical component of RF situational awareness and dynamic spectrum management. This work addresses the challenges of RF signal recognition and transmission pattern analysis problems forcomplex RF scenes using machine learning techniques. First, a feature learning technique is proposed, where the classiÞers trained on non-mixture single-label RF transmissions can be used for classifying mixed multi-label RF signals. Such an approach can signiÞcantly reduce the training burden.SpeciÞcally, novel supervised dictionary learning algorithms are developed with various feature-shaping constraints. The proposed algorithms are tested using real wide-band RF measurement data and robust performance is obtained even when a mixture of heterogeneous signal classes with widely di?erent powers is observed, and the number of component signals is not known a priori. Then, innovative unsupervised methods are proposed for complex radio scene analysis, where the patterns of individual transmissions as well as their channelcharacteristics are estimated using multiple antennas and deep neural networks(DNNs). Two types of algorithms are developed. First, a method that combines established signal processing algorithms with neural networks is derived. Then, a method based entirely on neural networks is developed. The methods utilize temporal/spatial/spectral Þltering to separate individual RF signals arriving from di?erent directions with the help of time-frequency (TF) patterns captured by the neural networks. The DNNs can be trained in an unsupervised and end-to-end fashion, rendering the training and operation e?cient and practical. Experimental results on synthetic and real-world datasets demonstrate the e?ectiveness of the proposed methods and the complementary merits of the two types of algorithms.