Event-Based Predictive Maintenance on Top of Sensor Data in a Real Industry 4.0 Case Study

Abstract

We deal with a real predictive maintenance case study encountered in modern Industry 4.0 settings: based on logs of past failures, we train a model to predict critical failures of equipment without any other domain expert knowledge well in advance. We propose a novel methodology that combines and extends the state-of-the-art in event-based predictions with advanced time-series analytics. This renders our technique applicable directly onto the sensor data, as it is produced in a modern factory setting. Further, we show that employing unsupervised learning techniques, such as continuous outlier monitoring, is a competitive approach. Although our techniques are developed and tested in a specific case study, they can be transferred to arbitrary settings.