Data-driven long-term condition-based maintenance using anomaly detection under concept drift: A case study of an ultrasonic sieve machine
Feng Zhua, Nicolas Jourdan, Beatriz Bretones Cassoli, Joachim Metternich · Procedia CIRP · 2024
Class imbalance and concept drift are two critical barriers to applying AI-based condition monitoring models in industrial manufacturing. In practical industrial production, there are usually more normal samples than abnormal samples, and the data distribution varies over time, limiting the training and the practical deployment of AI models. This paper proposes a solution based on anomaly detection and continual learning to address these two problems. Experiments on an ultrasonic vibrating sieve verify the solution's effectiveness. Learning on the normal samples can effectively identify sieve breakage, and continual learning can prevent model performance degradation.