Adaptive learning mode comparison for Predictive Maintenance Systems

The industry is increasingly using data-driven predictive maintenance with machine learning to prevent equipment failures. However, manufacturing systems are inherently complex and operate in dynamic, non-stationary environments, leading to the phenomenon of concept drift. Due to edge device limitations, traditional full retraining methods are often impractical, making incremental learning a potential alternative. This study evaluates both approaches through simulated real-world scenarios. Results show that full retraining generally performs better, especially in ideal and degraded conditions. While incremental learning shows acceptable performance with accurate labels and imbalanced datasets, it struggles when dealing with significant labelling inaccuracies.