Cross-domain intelligent fault classification of bearings based on tensor-aligned invariant subspace learning and two-dimensional convolutional neural networks

Abstract

Rolling element bearings faults are one of the main causes of breakdown of rotating machines. Aside from this, due to variation of operating condition, domain shift phenomenon results in important detection performance deterioration. Therefore, cross-domain intelligent fault detection and diagnosis of bearings is very critical for the reliable operation. In this paper, a new intelligent fault diagnosis approach based on tensor-aligned invariant subspace learning and two-dimensional convolutional neural networks (TAISL–2DCNN) is proposed for cross-domain intelligent fault diagnosis of bearings. The vibration signals of bearings fault are first formulated as a third-order tensor via trial, condition and channel. For adapting the source domain and the target domain tensor representations directly, without vectorization, the domain adaptation (DA) approach named TAISL is first proposed for tensor representation in bearing intelligent fault diagnosis field. Then the 2DCNN is utilized to recognize different faults. The performance of the presented algorithm has been thoroughly evaluated through extensive cross-domain fault diagnosis experiments. The verification results confirm that the developed approach is able to reliably and accurately identify different fault categories and severities of bearings when testing and training data are drawn from different distribution.