Unsupervised Radio Scene Analysis Using Neural Expectation Maximization

Abstract

An unsupervised learning-based blind RF scene analysis method is proposed. The method can analyze a complex radio scene containing a mixture of different transmission types and estimate the constituent signals with associated channel vectors from multi-antenna measurements. A deep neural network is trained to learn the unique time-frequency patterns of various signal types. The channels, noise powers, and encodings input to the neural network are estimated in a maximum likelihood framework via an expectation-maximization algorithm. Numerical tests using scenes constructed from real RF measurements verify the effectiveness of the proposed method.