“Small Data” Anomaly Detection for Unmanned Systems

Abstract

This paper presents an approach for a low cost, platform-agnostic, unmanned systems anomaly detection system that learns normal operating conditions from limited data and computing resources to track deviations from those conditions over time. Machine learning and automatic anomaly detection use has exploded in the Big Data arena with the availability of large volumes of historical data and extensive computing resources. However, in the case of unmanned systems, there is typically limited historical data available and computational resources are often restricted to embedded devices similar to cell phones. We discuss the application of two algorithms for anomaly detection in this “small data” context: 1) sparse modeling and 2) T-Digest. These algorithms are also designed and chosen to perform generically across a number of target application domains with a standalone health monitoring sensor box coupled with noninvasive sensors. Acoustic and inertial sensors have been initially selected to illustrate and validate the system capability and performance.