LAUSR

Abstract Prognostic Health Management (PHM) is a maintenance policy aimed at predicting the occurrence of a failure in components and consequently minimizing unexpected downtimes of complex systems. Recent developments in condition monitoring (CM) techniques and Artificial Intelligence (AI) tools enabled the collection of a huge amount of data in real-time and its transformation into meaningful information that will support the maintenance decision-making process. The emerging Cyber-Physical Systems (CPS) technologies connect distributed physical systems with their virtual representations in the cyber computational world. The PHM assumes a key role in the implementation of CPS in manufacturing contexts, since it allows to keep CPS and its machines in proper conditions. On the other hand, CPS-based PHM provide an efficient solution to maximize availability of machines and production systems. In this paper, evolving and unsupervised approaches for the implementation of PHM at a component level are described, which are able to process streaming data in real-time and with almost-zero prior knowledge about the monitored component. A case study from a real industrial context is presented. Different unsupervised and online anomaly detection methods are combined with evolving clustering models in order to detect anomalous behaviours in streaming vibration data and integrate the so-generated knowledge into supervised and adaptive models; then, the degradation model for each identified fault is built and the resulting RUL prediction model integrated into the online analysis. Supervised methods are applied to the same dataset, in batch mode, to validate the proposed procedure.