New trend in enhancing bearing remaining useful life prediction

Abstract EN

Generally the two main strategies taken in data-driven remaining useful life (RUL) prediction of a component/system can be summarized in 1) identifying a health indicator and predicting its trend until a predefined threshold; 2) mapping directly the health indicator (HI) to RUL by regression.

Under the first category, traditional extracted features for RUL prediction usually show undesirable behaviors such as fluctuation, non-monotonicity and abrupt increase at the end which hampers the accuracy of the RUL prediction.

To enhance the prediction accuracy, this paper brings a new feature selection method, based on preprocessing further the extracted features in a way that the identified prognostic feature results in an obvious trend quality.

A set of established and proposed suitability metrics for the prognostic task are used to assess the identified features qualities.

The Particle Filtering technique is adopted as a projection tool as well for the prediction of the RUL due to its capability to carry nonlinear systems in presence of non-Gaussian process/observation noise.

Datasets from bearings run-to-failure experiments provided by FEMTO-ST Institute - IEEE PHM 2012 challenge- were used to validate our approach.

A mean percentage error of 12.18% was achieved indicating that the model worked accurately and reliably on every tested bearing..