Classification, Denoising, and Deinterleaving of Pulse Streams With Recurrent Neural Networks

Abstract

Pulse streams of many emitters have flexible features and complicated patterns. They can hardly be identified or further processed from a statistical perspective. In this paper, we introduce recurrent neural networks (RNNs) to mine and exploit long-term temporal patterns in streams and solve problems of sequential pattern classification, denoising, and deinterleaving of pulse streams. RNNs mine temporal patterns from previously collected streams of certain classes via supervised learning. The learned patterns are stored in the trained RNNs, which can then be used to recognize patterns-of-interest in testing streams and categorize them to different classes, and also predict features of upcoming pulses based on features of preceding ones. As predicted features contain sufficient information for distinguishing between pulses-of-interest and noises or interfering pulses, they are then used to solve problems of denoising and deinterleaving of noise-contaminated and aliasing streams. Detailed introductions of the methods, together with explanative simulation results, are presented to describe the procedures and behaviors of the RNNs in solving the aimed problems. Statistical results are provided to show satisfying performances of the proposed methods.