Data-Driven Predictive Maintenance in Evolving Environments: A Comparison Between Machine Learning and Deep Learning for Novelty Detection

Predictive Maintenance (PdM) is the newest strategy for maintenance management in industrial contexts. It aims to predict the occurrence of a failure to minimize unexpected downtimes of equipment and maximize the useful life of the monitored components. In a data-driven approach, PdM makes use of Machine Learning (ML) algorithms to extract relevant features from historical signals, identify and classify possible faults (diagnostics), and predict the components' remaining useful life (RUL) (prognostics). The major challenge lies in the high complexity of industrial plants, where both operational and environmental conditions change over time and a large number of unknown a priori modes may occur. A solution to this problem is offered by novelty detection, where a representation of the normal operating state of the machinery is learned and compared with online measurements in order to identify new operating conditions. In this paper, a comparison between ML and Deep Learning (DL) methods for novelty detection is conducted, to evaluate their effectiveness and efficiency in different scenarios. To this purpose, a case study considering vibration data collected from an experimental platform is carried out. Results show the superiority of DL on traditional ML methods in all the evaluated scenarios.