Local lowest-rank dynamic mode decomposition for transient feature extraction of rolling bearings

The fault diagnosis mainly relies on the detection of periodic pulse components caused by the local damage. However, the impulsive component in bearing vibration signals is usually disturbed by noise and some harmonics due to the harsh working environment, which makes great challenge to fault diagnosis. To enhance the performance in fault diagnosis of rolling bearing, a local lowest-rank dynamic mode decomposition (LLRDMD) is proposed in this paper. Firstly, an improved projection operator is proposed to reduce the noise of augmented snapshot matrix and thus eliminating the bias of high-dimension dynamic system matrix. Secondly, by solving a local lowest-rank optimization problem, the impulsive component and interference components in the high-dimension dynamic system matrix are separated. Compared with DMD, the proposed method achieves better performance in the extraction of transient pulse component while avoiding the problems of parameter selection and mode selection. Lastly, the proposed method is applied to three sets of real datasets. Compared with existing methods, LLRDMD achieves a higher accuracy in fault diagnosis.