Detecting Anomalies: A Case Study in Predictive Maintenance of Telescopes

A predictive maintenance (PdM) approach is considered among the best solution for the maintenance, as it introduces a lot of advantages among which there is the increase of reliability performances, the saving in maintenance and inventory holding costs. Although several techniques are available for predictive maintenance, those based on data science feature a widespread use; the reasons are mainly due to the availability of huge sets of telemetric data from the system to which maintenance must be applied, coming from Internet of Things – based sensors. Supervised learning – based approaches are the most used in these cases. Sometimes, data available features the lack of information allowing to discern a state of malfunction of the system from those relevant to the faulty-free functioning. In this case, suitable unsupervised learning – based approaches may be used. The paper aims at presenting an approach to the realization of a predictive maintenance system even in the absence of data labels related to failure information. The approach is described and applied to the real case of study of an astronomical telescope.