A feature extraction procedure based on trigonometric functions and cumulative descriptors to enhance prognostics modeling

Performances of data-driven approaches are closely related to the form and trend of extracted features (that can be seen as time series health indicators). (1) Even if much of data-driven approaches are suitable to catch non-linearity in signals, features with monotonic trends (which is not always the case!) are likely to lead to better estimates. (2) Also, some classical extracted features do not show variation until a few time before failure occurs, which prevents performing RUL predictions in a timely manner to plan maintenance task. The aim of this paper is to present a novel feature extraction procedure to face with these two problems. Two aspects are considered. Firstly, the paper focuses on feature extraction in a new manner by utilizing trigonometric functions to extract features (health indicators) rather than typical statistic measures like RMS, etc. The proposed approach is applied on time-frequency analysis with Discrete Wavelet Transform (DWT). Secondly, a simple way of building new features based on cumulative functions is also proposed in order to transform time series into descriptors that depict accumulated wear. This approach can be extended to other types of features. The main idea of both developments is to map raw data with monotonic features with early trends, i.e., with descriptors that can be easily predicted. This methodology can enhance prognostics modeling and RUL prediction. The whole proposition is illustrated and discussed thanks to tests performed on vibration datasets from PRONOSTIA, an experimental platform that enables accelerated degradation of bearings.