Similarity-based anomaly score for fleet-based condition monitoring

Abstract

Monitoring machines and early fault detection reduces production downtime, repair costs, and human casualties. Traditionally, machine condition monitoring requires performing a time-consuming manual analysis to find indicators for each potential fault. Moreover, these approaches often only focus on a single machine, while many industrial applications involve monitoring a fleet of machines. In such applications, it is often safe to assume that the majority of machines are in a healthy state. In comparable operating states, the behavior of these healthy machines is similar and any deviating machine is thus likely to be faulty. Previously, we proposed a fleet-based anomaly framework that can assess the health status of each machine in the fleet by detecting these deviations. It groups together similarly behaving machines and assumes that the healthy ones form the largest group and assigns an anomaly score to each machine based on the size of the group it belongs to. In this work, we propose a similarity-based anomaly score that offers multiple benefits over the cluster-based anomaly score. First, this score better represents the severity of a machine fault. Second, it allows to assess the health status of individual machines instead of machine groups. Finally, using similarities provides more nuanced insights in a machine's health status, especially for gradual degrading machines. Experiments show that the similarity-based anomaly score is superior to the cluster-based approach.