Semi-random subspace with Bi-GRU: Fusing statistical and deep representation features for bearing fault diagnosis

Statistical features and deep representation features have been widely used in bearing fault diagnosis. These two kinds of features have their superiorities, however, few studies have explored combining them and considering their heterogeneousness. Therefore, a Semi-Random Subspace method with Bidirectional Gate Recurrent Unit (Bi-GRU), i.e., SRS-BG, is proposed in this paper, to take full advantage of fusion features for bearing fault diagnosis. Firstly, the statistical features are obtained by multiple signal processing methods, and the deep representation features are obtained by Bi-GRU. Secondly, the heterogeneousness among these features are considered by proposing a novel structure sparsity learning model, which is further utilized to produce based classifiers in the proposed Semi-Random Subspace method. Finally, experiments on bearing vibration datasets derived from Case Western Reserve University validate that the proposed feature fusion strategy greatly enhances diagnostic performance, and outperforms other existing ensemble learning methods.

Automated summary

The paper proposes a Semi-Random Subspace method with Bidirectional Gate Recurrent Unit (Bi-GRU) to fully utilize fusion features for bearing fault diagnosis. Statistical features obtained by multiple signal processing methods and deep representation features obtained by Bi-GRU are combined and the heterogeneity among them is considered through a novel structure sparsity learning model. Experimental results on bearing vibration datasets validate that the proposed feature fusion strategy greatly enhances diagnostic performance, outperforming other existing ensemble learning methods.