Low Data Prognostic Model for Rolling Element Bearing Remaining Useful Life

Abstract Rolling element bearings play a pivotal role in rotating machines and are key components of small gas turbine engines. Due to the limited lifespan of these components, they are critical under the maintenance point of view. Various techniques, relying on different algorithms, exist for damage identification. However, the prognostics and residual useful life (RUL) estimation for rolling element bearings present a more challenging task mainly due to the complexity of these components and of their damaging mechanisms. To characterize the bearing failure, extensive and costly testing of several bearings until failure is required. In this paper, a data-driven prognostic model is trained using data obtained from a single run-to-failure test. The bearing is already damaged at the beginning of the test. The bearing acceleration data is processed, and kurtosis is the selected feature to describe the damage evolution of the bearing. The Dynamic Mode Decomposition (DMD) algorithm is employed as surrogate model. To further enhance model accuracy, a time delay or Hankel matrix is used to facilitate state-space modeling and the identification of dynamic behavior within the time series. Subsequently, a particle filter is developed with the trained surrogate models to estimate the RUL of the bearing. The developed model is tested on run-to-failure data from the Center for Intelligent Maintenance Systems (IMS) of the University of Cincinnati. Remarkably, a relatively small volume of data is sufficient to obtain accurate predictions, especially close to the bearing end-of-life, for different bearing geometries and operating conditions.