Using simulation to evaluate a concept drift detector for condition based maintenance

Industry 4.0. has empowered the integration of real-time sensor readings and process knowledge for early identification of faults. In this paper, simulation was used to evaluate the process of classifying data streams as drifting towards defective behavior. Focusing on the challenges imposed by real-word, unlabeled and non-stationary data, the proposed method was assessed by conducting a case-study in the plastic extrusion domain. The knowledge gained from initial tests on generated synthetic data sets was used to isolate symptoms of a clogged screen pack from the concept drifts detected by the Adaptive Windowing algorithm. Experimental results for 40 days of running time of an extruder show that the proposed method helped in identifying a defective screen pack state upon the appearance of defects in the extrudate.