An approach for feature extraction and selection from non-trending data for machinery prognosis

With the paradigm shift towards prognostic and health management (PHM) of machinery, there is need for reliable PHM methodologies with narrow error bounds to allow maintenance engineers take decisive maintenance actions based on the prognostic results. Prognostics is mainly concerned with the estimation of the remaining useful life (RUL) or time to failure (TTF). The accuracy of PHM methods is usually a function of the features extracted from the raw data obtained from sensors. In cases where the extracted features do not display clear degradation trends, for instance highly loaded bearings, the accuracy of the state of the art PHM methods is significantly affected. The data which lacks clear degradation trend is referred to as non-trending data. This study presents a method for extracting degradation trends from non-trending condition monitoring data for RUL estimation. The raw signals are first filtered using a discrete wavelet transform (DWT) denoising filter to remove noise from the acquired signals. Time domain, frequency domain and time-frequency domain features are then extracted from the filtered signals. An autoregressive model is then applied to the extracted features to identify the degradation trends. Features representing the maximum health information are then selected based on a performance evaluation criteria using extreme learning machine (ELM) algorithm. The selected features can then be used as inputs in a prognostic algorithm. The feasibility of the method is demonstrated using experimental bearing vibration data. The performance of the method is evaluated on the accuracy of RUL estimation and the results show that the method can be used to accurately estimate RUL with a maximum error of 10\%.