A practical and synchronized data acquisition network architecture for industrial robot predictive maintenance in manufacturing assembly lines

Abstract

This manuscript presents a methodology and a practical implementation of a network architecture for industrial robot data acquisition and predictive maintenance . We propose a non-intrusive and scalable robot signal extraction architecture, easily applicable in real manufacturing assembly lines. The novelty of the paper lies in the fact that it is the first proposal of a network architecture which is specially designed to address the predictive maintenance of industrial robots in real production environments. All the infrastructure needed for the implementation of the architecture is comprised of traditional well-known industrial assets. We synchronize the data acquisition with the execution of robot routines using common Programmable Logic Controllers (PLC) to obtain comparable data batches. A network architecture that acquires comparable and structured data over time, is a crucial step to advance towards an effective predictive maintenance of these complex systems, in terms of effectively detecting time dependent degradation. We implement the architecture in a real automotive manufacturing assembly line and show the potential of the solution to detect robot joint failures in real world scenarios. The architecture is therefore specially interesting for industrial practitioners and maintenance personnel. Finally, we test the feasibility of using one-class novelty detection models for robot health status degradation assessment using data of a real robot failure. To the best of our knowledge, this is the first contribution that uses robot torque signals of a real production line failure to train one-class models. • A non-intrusive and scalable robot data acquisition architecture is proposed for real manufacturing assembly lines. • The network acquires synchronized comparable robot operational data over time. • The architecture is able to detect a significant deviation in a faulty ABB IRB 6400r robot data. • One-class novelty detection models detect deviations in real operational conditions.