[3D3] Combining data-driven features and physics-based models for the prediction of remaining useful life of bearings
F Hosseinpour, M Behzad and E Zio
In practice, the prediction of the remaining useful life (RUL) of rolling bearings faces challenges such as complex degradation processes, exposure to varying working conditions and insufficient run-to-failure data. This work addresses two challenges: the scarcity of data for training data-driven models and the related difficulty of generalisation; and the difficulty involved in developing physics-based models of the degradation mechanisms. An innovative framework for RUL prediction is developed by combining data-driven features and physics-based models. The proposed framework consists of the following: an abnormality indicator is derived from time-series data utilising a convolutional neural network autoencoder (CNN-AE) and the Manhattan distance; a dynamic degradation model is established based on the Paris crack growth model; and a particle filter (PF) method that integrates Monte Carlo simulation with the degradation model is developed to estimate the RUL and the associated uncertainty. The proposed framework has been applied to the benchmark PRONOSTIA dataset; the obtained results outperform those from statistical feature-based methods.
