Kernel-Based Dynamic Ensemble Technique for Remaining Useful Life Prediction

Abstract

Prognostic and health management (PHM) has been widely used in manufacturing system, particularly, for predictive maintenance (PdM). The purpose of PdM is to predict whether equipment or parts is in health. Typically, the statistical exponential models with the health index were often applied for the remaining useful life (RUL) estimation. However, due to the diverse equipment characteristics and rapid environmental change, no single prediction model can predict well for RUL based on its assumptions and limitations. This study proposes a kernel-based dynamic ensemble technique (KDET) embedded with Inference Confidence Index (ICI) to build the weight adjustment of each model and model retraining mechanism. The ICI is built to measure the belief of the prediction by evaluating the similarity of multiple prediction models, and thus guide the concept drift to update the models immediately for the incoming streamline data. Two datasets are applied to validate the proposed KDET, and the result shows that the KDET can dynamically and effectively integrate multiple models for robust RUL prediction over time and thus improve the PdM system.