We deal with a real predictive maintenance (PdM) scenario in an Industry 4.0 setting. With a help of the Sibyl platform, we can monitor live data from key components of a Philips factory equipment; in this work, we focus on a cold-forming press. Due to the dynamic environment of the operation of this press, unsupervised anomaly detection techniques are used to timely detect the wear, where early anomalies are interpreted as warning signs of a forthcoming failure. Typically such techniques assign an anomaly score, and the problem we face is how to appropriately set a threshold for this score. We introduce and compare four generally applicable thresholding techniques, two of which are dynamic, i.e., they continuously refine the threshold during the episode lifetime. We discuss the properties of these techniques and quantitatively evaluate their behavior in our case study.

No abstract

No abstract

No abstract

No abstract