Abstract
Due to the harsh and complex operating conditions, rolling element bearings (REBs) are prone to failures, which can result in significant economic losses and catastrophic breakdowns. To efficiently extract weak fault features from raw signals, singular value decomposition (SVD)-based signal denoising methods have been widely adopted in the field of rolling bearing fault diagnosis. In traditional SVD-based methods, singular components (SCs) with significant singular values are selected to reconstruct the denoized signal. However, this approach often overlooks low-energy SCs that contain important fault information, leading to inaccurate diagnosis. To address this issue, we propose a new selection scheme based on frequency domain multipoint kurtosis (FDMK), along with a reweighting strategy based on FDMK to further emphasize weak fault features. In addition, the estimation process of fault characteristic frequency is introduced, allowing FDMK to be calculated without prior information. The proposed FDMK-SVD can adaptively extract periodic fault features and accurately identify the health condition of REBs. The effectiveness of FDMK-SVD is validated using both simulated and experimental data obtained from a locomotive bearing test rig. The results show that FDMK-SVD can effectively extract fault features from raw vibration signals, even in the presence of severe background noise and other types of interferences.