Drift-Aware Edge Intelligence for Remaining Useful Life Prediction in Industrial Internet of Things

Advances in the Industrial Internet of Things (IIoT) enable complex machine data analysis and information extraction to improve manufacturing throughput. However, existing approaches are fixated on improving prediction accuracy and overlook the implications of model drift on edge-based devices. For this reason, we present a real-time system for monitoring RUL model performance drift via incremental learning. Besides, edge devices are compute and memory bounded with diverse configurations. To address this problem, we propose to re-train the RUL prediction model on an edge server and conduct extensive experiments to determine the optimal quantity of batchsized data that exhibits the lowest error deviation compared to tabula rasa model training. Our findings demonstrate that incremental training yields an average training time savings of 60% for an open-source industrial dataset while incurring a mean difference of -1.1% and -4.9% for prognostics performance metrics root-mean-square-error and score, respectively.