Separability index-based feature selection and a two-tier classifier for improving diagnostic performance in bearings

Abstract

Faults in bearings have traditionally been diagnosed by identifying the characteristic defect frequencies in the envelope power spectrum (EPS). Although EPS-based methods are effective for constant operating conditions, it makes difficult to detect a fault when the machine operates at variable shaft speeds. To address this issue, we propose a method that performs as follows: (a) a feature extraction technique is used to extract as many information as possible from different sub-bands using a non-stationary signal analysis method (e.g., discrete wavelet pack transform (DWPT)), (b) a novel feature selection method is then applied to select the most informative features from the extracted high-dimensional feature pool, and (c) the selected features are further fed to a two tiers classifiers to identify the condition of a bearing. This two tiers classifier includes the combination of support vector machine groups in the first tier and the combination multilayer perceptron in the second tier. The proposed method is tested for a bearing fault diagnosis application using acoustic emission (AE) signal under various conditions.