Bearing Fault Diagnosis Based on Discriminant Analysis Using Multi-View Learning

Abstract

Bearing fault diagnosis has been a challenge in rotating machinery and has gained considerable attention. In order to correctly classify faults, the conventional fault diagnosis methods are mostly based on vibration signals. However, features extracted from a single view of vibration signals may leave out useful information, which can cause the incompleteness of intrinsic information and increase the risk of the performance degradation of fault classifications. In this paper, a novel bearing fault diagnosis method, discriminant analysis using multi-view learning (DAML), is proposed to tackle this issue. Multi-view datasets referring to vibration and acoustic signals are obtained by carrying out a fast Fourier transform (FFT). Then, multi-view feature (MVF) representation, including view-invariant and category discriminative information in a common subspace, is achieved based on canonical correlation analysis (CCA) and uncorrelated linear discriminant analysis (ULDA). Ultimately, with the help of the K-nearest neighbor (KNN) classifier built on the multi-view features, bearing faults are identified. The extensive experimental results show that DAML can identify the bearing fault accurately and outperforms other competitive approaches.