Predicting Machine Errors based on Adaptive Sensor Data Drifts in a Real World Industrial Setup

We present a dynamic error prediction system for industrial production machines. We implemented a flexible data collection tool to create error warnings for a production line, which aims to improve the already existing static alarm models. For industrial machines, there are threshold-based alarm models set by prior experiences and observations of the operator. For machines without standardized interfaces and communication protocols, which are not Industry 4.0 compatible, it represents a challenge to implement and add a dynamic and opportunistic system behavior. Machines need to learn from past errors autonomously and adapt the production properties dynamically. We implemented a framework that makes production machines conform to the Internet of Things (IoT) concepts, by making previously non-IoT enabled resources available to get new insights into the production processes.The system component recognition and the database setup is done fully automatically by our developed system.We designed and applied a feature-based data drift model in a real-world industrial setting to determine data deviation between normal and erroneous work-pieces in real-time to predict upcoming erroneous behavior. The drift analysis flagged and predicted work-pieces as erroneous several minutes before the pre-defined machine alarms would have been raised. The resulting flagged sensors and values can be compared to the system determined errors to get new insights into the abnormal machine behavior. For the reduction of downtime, the most valuable immediate result of the system is the ability to notify the operator earlier and reduce overall downtime.