Machine Learning Methods for Monitoring of Quasiperiodic Traffic in Massive IoT Networks

One of the central problems in massive Internet-of-Things (IoT) deployments is the monitoring of the status of a massive number of links. The problem is aggravated by the irregularity of the traffic transmitted over the link, as the traffic intermittency can be disguised as a link failure and vice versa. In this article, we present a traffic model for IoT devices running quasiperiodic applications and we present unsupervised, parametric machine learning methods for online monitoring of the network performance of individual devices in IoT deployments with quasiperiodic reporting, such as smart metering, environmental monitoring, and agricultural monitoring. Two clustering methods are based on the Lomb-Scargle periodogram, an approach developed by astronomers for estimating the spectral density of unevenly sampled time series. We present probabilistic performance results for each of the proposed methods based on simulated data and compare the performance to a naive network monitoring approach. The results show that the proposed methods are more reliable at detecting both hard and soft faults than the naive-approach, especially, when the network outage is high. Furthermore, we test the methods on real-world data from a smart metering deployment. The methods, in particular the clustering method, are shown to be applicable and useful in a real-world scenario.