Anomaly Detection and Classification in Predictive Maintenance Tasks with Zero Initial Training

Abstract

The Fourth Industrial Revolution has led to the adoption of novel technologies and methodologies in factories, making these more efficient and productive. Among the new services which are changing industry, there are those based on machine learning algorithms, which enable machines to learn from their past observations and hence possibly forecast future states. Specifically, predictive maintenance represents the opportunity to understand in advance possible machine outages due to broken parts and schedule the necessary maintenance operations. However, in real scenarios predictive maintenance struggles to be adopted due to a multitude of variables and the heavy customization it requires. In this work, we propose a novel framework for predictive maintenance, which is trained online to recognize new issues reported by the operators. Our framework, tested on different scenarios and with a varying number and several kinds of sensors, shows recall levels above 0.85, demonstrating its effectiveness and adaptability.