Discovery Of Anomalous Patterns Within Multidimensional, Asynchronous Time-series With An Emphasis On The "internet Of Things"

Abstract

In this dissertation we examine ``Internet-scale'' systems that present us with multidimensional time-series data characterized by many sources sending symbols at irregular intervals over a common channel. We explore a unique method for the discovery of hidden populations of similar sources and their previously-unknown behavioral patterns, and using these discoveries, reveal anomalous sources and/or time-frames based on their statistical properties. To do so, we employ several well-studied mechanisms, such as k-means and Principle Component Analysis (PCA), and bring to bear analysis tools from other disciplines, such as the use of n-grams and "motifs," that have not previously been considered in these contexts. While applicable to the study of any system whose attributes can map to the generalized model we present, the approach is of particular interest when dealing with large numbers of remote devices that make up the "Internet of Things" (IoT). The method is applied to several discrete layers of cellular wireless communications infrastructure for a diverse set of commercial Machine-to-Machine (M2M) applications where the behavior of the devices was examined as they interacted with carrier network elements. While it is common to study these systems in the context of their application-level duties, the devices' behavior at lower levels of the solution stack has received less attention, and is often poorly understood by either the application engineers or network operators.