A TAXONOMY OF UNIVARIATE ANOMALY DETECTION ALGORITHMS FOR PREDICTIVE MAINTENANCE

Abstract

Anomaly detection has a wide variety of applications, ranging from intrusion detection in cybersecurity to fraud detection in finance. Among the most prominent applications is predictive maintenance in manufacturing, which involves performing maintenance only when truly necessary, based on the condition of relevant equipment instead of following a fixed maintenance schedule. When implemented correctly, predictive maintenance can lead to more significant cost savings than other preventative maintenance approaches. Unfortunately, the unique challenges present in anomaly detection (including the very broad definition of an anomalous instance) make it particularly difficult to choose an appropriate algorithm, since each algorithm’s performance is so dependent on the use case. In this paper we present an up-to-date taxonomy of univariate anomaly detection approaches to predictive maintenance, which is aimed at aiding practitioners to design effective predictive maintenance models for their specific use cases, based on numerical benchmark tests.