Optimized Random Forest Model for Remaining Useful Life Prediction of Experimental Bearings
Muhammad Gibran Alfarizi, Bahareh Tajiani, Jørn Vatn, Shen Yin
Abstract
Bearings are essential to the reliable operation of rotating machinery in manufacturing processes. There is a rising demand for accurate bearing remaining useful life (RUL) predictions. The data-driven technique for predicting RUL of bearing has demonstrated promising prospects to facilitate intelligent prognostics. This article proposes a new data-driven prediction framework for bearing RUL utilizing an integration of empirical mode decomposition, random forest (RF), and Bayesian optimization. The proposed framework consists of two main phases: 1) feature extraction and 2) RUL prediction. The first phase of this framework focused on decomposing the empirical input signals using empirical mode decomposition into distinct frequency bands to filter out irrelevant frequencies and determine the fault characteristics of the signals. In the second phase, the RUL prediction is then carried out by an RFs-based model with its hyperparameters tuned by Bayesian optimization. The proposed approach is validated using datasets obtained from an actual run-to-failure experiment of roller bearings. The experiment results significantly improved compared to the standard data-driven and stochastic approaches.