Performances of data-driven prognostics approaches are closely dependent on form, and trend of extracted features. Indeed, features that clearly reflect the machine degradation, should lead to accurate prognostics, which is the global objective of the paper. This paper contributes a new approach for features extraction / selection: the extraction is based on trigonometric functions and cumulative transformation, and the selection is performed by evaluating feature fitness using monotonicity and trendability characteristics. The proposition is applied to time-frequency analysis of non-stationary signals using Discrete Wavelet Transform. The main idea is to map raw vibration data into monotonic features with early trends, which can be easily predicted. To show that, selected features are used to build a model with a data-driven approach namely, the Summation Wavelet-Extreme Learning Machine, that enables a good balance between model accuracy and complexity. For validation and generalization purpose, vibration data from two real applications of Prognostics and Health Management challenges are used: 1) cutting tools from Computer Numerical Control (CNC) machine (2010), and 2) bearings from platform PRONOSTIA (2012). Performances of the proposed approach are thoroughly compared with the classical approach by performing: feature fitness analysis, cutting tool wear "estimation" and bearings "long-term predictions" tasks, which validates the proposition.

No abstract