Industrial application of Machine Learning: Predictive Maintenance for failure detection

Industry 4.0 is characterized by production systems that integrate multiple sensors to collect and transmit data. This huge amount of information can assure benefits if properly analyzed. Machine Learning provides a a set of techniques to extract correlations in such a complex scenario. This work focuses on the development of algorithms to perform predictive maintenance; the goal is to detect failures from the alarms collected by refrigeration systems. Multiple solutions are investigated both using boost decision trees and neural networks. In addition, Natural Language Processing studies are applied to automatically classify failures from the maintenance ticket requests sent from the customers. State-of-the-art algorithms are compared with the proposed SpectrumBoost method, which extract features from texts using a spectrum kernel and performs classification through an XGBoost model.