Unsupervised Deep-Learning for Distributed Clock Synchronization in Wireless Networks

Abstract

One of the major factors which limits the throughput in wireless communications networks is the accuracy of time synchronization between the nodes in the network. Synchronization methods based on pulse-coupled oscillators (PCOs) have the advantage of simple implementation and achieve high accuracy when the nodes are closely located. However, such schemes tend to have poor synchronization performance for distant nodes, as well as in the presence of clock frequency offsets between the nodes. In this paper we present a novel PCO-based Deep neural network (DDN)-Aided Synchronization Algorithm coined DASA. We design DASA as a novel low-complexity and interpretable architecture by converting classic PCO-based synchronization into <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a trainable discriminative model</i> . To enable DASA to operate in dynamic settings, we propose a novel, unsupervised, distributed, fast <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">online</i> training scheme which is able to train DASA <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">within a few sampling instances, locally</i> , thereby avoiding the need for information exchange between the nodes or for a central node for coordination. DASDA is demonstrated to achieve an improvement by a factor greater than ten compared to the classic reference scheme. Lastly, we propose another novel, distributed <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">offline</i> training scheme for DASA, which is demonstrated to offer a tradeoff between performance and simplicity of deployment compared to the online training scheme, yet, at the same time, DASA with offline training still achieves superior performance compared to the classic reference scheme.