Bearing Remaining Useful Life Prediction Combining Trajectory Similarity and Relevance Vector Machine

Bearing remaining useful life (RUL) prediction is critical for safe operation of rotating machinery. In this paper, we propose a combined RUL prediction approach that leverages both trajectory similarity and relevance vector machine (RVM). The similarity based prediction relies on historical degradation trajectories that are highly similar to the online data, hence would perform poorly if all historical trajectories have low similarity with the online data. The RVM based prediction relies solely on a regression model learned from the available online data, thus gives an inaccurate prediction when insufficient data are available in the early stage of degradation. A weighted sum of these above two predictions is proposed to address the limitation of each single prediction method, whose weights are determined by solving a non-negative least squares fitting problem. To further improve RUL prediction accuracy, we distinguish between fast and slow degradation modes, so that each mode uses a different set of historical degradation trajectories and kernel functions. By doing so, we predict RUL under the identified mode. The case study using the PHM2012 dataset demonstrates the effectiveness of the proposed RUL prediction approach.