Improved Variational Mode Decomposition Based on Scale Space Representation for Fault Diagnosis of Rolling Bearings.

Accurate extraction of weak fault information from non-stationary vibration signals collected by vibration sensors is challenging due to severe noise and interference. While variational mode decomposition (VMD) has been effective in fault diagnosis, its reliance on predefined parameters, such as center frequencies and mode number, limits its adaptability and performance across different signal characteristics. To address these limitations, this paper proposes an improved variational mode decomposition (IVMD) method that enhances diagnostic performance by adaptively determining key parameters based on scale space representation. In concrete, the approach constructs a scale space by computing the inner product between the signal's Fourier spectrum and a Gaussian function, and then identifies both the mode number and initial center frequencies through peak detection, ensuring more accurate and stable decomposition. Moreover, a multipoint kurtosis (MKurt) criterion is further employed to identify fault-relevant components, which are then merged to suppress redundancy and enhance diagnostic clarity. Experimental validation on locomotive bearings with inner race faults and compound faults demonstrates that IVMD outperforms conventional VMD by effectively extracting fault features obscured by noise. The results confirm the robustness and adaptability of IVMD, making it a promising tool for fault diagnosis in complex industrial environments.