Experimental Investigation of Failure Thresholds of Rolling Element Bearings

The rolling element bearing (REB) is the most common mechanical element used in industrial machinery. Usually, in data-driven prognostic models for predicting the remaining useful life (RUL) of REBs, the failure threshold (FT) is assumed as a constant value of a statistical feature, such as root mean square (RMS), peak or kurtosis, extracted from vibration signals collected by sensors. The procedure to define the FT value is conservative and expert-based. In addition, FT is typically defined only with respect to one of the features. In this paper, to analyze the effect of the definition of FT by only one feature, the correlation of statistical features is studied on accelerated-life datasets of REBs. First, by reviewing the literature in the field of PHM and FT, features including peak, kurtosis, RMS, and level crossing rate are selected to be analyzed. Then, the correlation of these features is investigated on the PRONOSTIA and the Sharif University of Technology accelerated-life datasets. The results show that the correlation of features varies in the different tests of these datasets. As a result, in this paper, it is proposed that the FT is to be defined as a multi-feature fusion, which arises from the relationship between features. This proposed method helps resolve the ambiguity in defining the FT as a single fixed value.