The Problem with Real-World Novelty Detection - Issues in Multivariate Probabilistic Models

Abstract

Novelty and anomaly detection in real-world data streams are becoming more and more important for IoT, industry 4.0 and digital-twin applications. However, most of these algorithms are designed in-vitro and usually not very resilient against the failure behaviour of real-world systems, that is, minor system faults (e.g. a failing sensor, small damage, or firmware updates). In most scenarios, such a minor fault leads to a total failure of the detection engine, resulting either in the constant reporting of an anomaly or a total inability for further detection. In this article we investigate this problem in more detail and present simple approaches to circumvent them.