Fisher embedding shift-invariant dictionary learning for weak feature recognition in bearing health monitoring

Abstract

Dictionary learning has gained widespread attention in weak feature recognition for bearing health monitoring due to its advantages of possessing fewer parameters, simple and efficient algorithms, and strong interpretability. However, conventional methods often fail to fully utilize the interclass and intraclass characteristics of dictionary atoms and do not take into account the shift-invariance of the dictionary, which in turn limits their performance. To address these challenges, a new method named Fisher embedding shift-invariant dictionary learning (FESIDL) is proposed in article. Specifically, Fisher constraints are imposed on both atoms and coding coefficients to promote their discriminability, and the dictionary optimization is designed based on singular value decomposition to make the learned dictionary shift-invariant and more discriminative. Finally, FESIDL simultaneously utilizes the reconstruction error and coding coefficient vector as classification criteria to ensure a high classification accuracy. Two experiments are conducted, and the results demonstrate the effectiveness and stability of the proposed method for weak feature recognition in bearing health monitoring when compared to several advanced methods in the field.