Towards bearing failure prognostics: a practical comparison between data-driven methods for industrial applications

Research studies on data-driven approaches to rotating components and rolling element bearing (REB) prognostics have recently witnessed a rapid increase. These data-driven methods rely on sensor data for condition monitoring and degradation assessments; however, the problem of mining features from these sophisticated data using appropriate intelligent methods and choosing a practically reliable predictive model(s) has become a global concern. Vibration monitoring for REBs have over the years shown great effectiveness. Although monotonic statistical features serve as reliable health indicators (HIs), relying on a single feature for optimal bearing degradation assessment is inefficient. By fusing highly monotonic features using appropriate methods, a more reliable HI can be constructed and from this, various degradation states/stages and time to start prediction (TSP) can be identified by mapping known failure modes/degradation states to cluster points from clustering algorithms. Emphatically, the choice of regression algorithms for prognostics poses more concern as engineers and data scientists are faced with choosing between Bayessian machine learning (ML) and deep learning (DL) methods. This study presents a methodology for constructing a reliable HI for bearing prognostics, choosing a reliable TSP, and provides a comparison between ML and DL methods for bearing prognostics. As representatives of both domains, the Gaussian process regression (GPR) and the deep belief network (DBN) are introduced and compared. The results provide a reliable paradigm for prognosible feature representation for REBs and for choosing between both domains while considering their dependencies, efficiencies, and deficiencies.