A Method to Handle Concept Drift in Predicting Remaining Useful Life

Predictive Maintenance (PdM), especially the prediction of Remaining Useful Life (RUL), has long been recognized as a pivotal capability in the progression towards autonomous manufacturing. However, a primary challenge impeding the scalability and generalization of RUL prediction is the concept drift associated with machines - specifically, the changes in machine data patterns over time due to aging and wear, or sudden changes in production recipes, etc. Such concept drifts compromise the predictive accuracy of static trained machine learning models, thereby posing challenges for RUL prediction. In this paper, we present a method to counteract these concept drifts in the RUL prediction task. We first outline a framework for RUL prediction and then devise a concept drift handler utilizing a gradient descent weighting method. This entire approach is subsequently tested on a datasets from a CNC spindle’s bearing at the Model Factory@SIMTech to assess its viability. The developed technology would later be used as an advance feature in the PdM app at our Model Factory for training and engagement with local industry.