Video vibration measurement and composite fault feature extraction of rolling bearings based on multi-scale spatial filtering

Abstract Accurate and robust non-contact fault diagnosis of rolling bearings remains a challenging task in rotating machinery, particularly under complex operating conditions and varying light intensities. To address this, this paper proposes a video-based framework for vibration measurement and composite fault feature extraction based on multi-scale spatial filtering. High-speed industrial cameras are employed to capture vibration image sequences, while multi-scale spatial filtering is implemented using a multi-scale, parameter-adaptive, two-dimensional Gabor filter (MPA 2D Gabor). By adaptively adjusting wavelength and standard deviation according to local texture, density, and contrast features, the MPA 2D Gabor enhances phase responses and improves the acquisition of weak vibration signals. A multi-scale response fusion strategy further integrates phase information through confidence-weighted averaging, suppressing background noise and emphasizing fault-related spectral components. For composite fault feature extraction, adaptive MOMEDA is introduced, leveraging multipoint kurtosis spectra to determine the period parameter T and optimize the filter length L, enabling accurate separation of multiple fault frequency components. Extensive comparative experiments on outer race faults under different light intensities (100%, 75%, 50%, 25%) demonstrate that the proposed method outperforms fixed parameter 2D Gabor filtering and conventional accelerometer-based techniques in fault frequency prominence, noise suppression, and light intensity adaptability, maintaining an outlier proportion below 0.02. Additional composite fault tests confirm that the framework can distinctly separate and identify individual fault components. On the same videos and region of interest (ROI), under a CPU-only setup with 10,000 frames, the proposed method is approximately 3.58 times faster than the optical-flow baseline and provides more stable resolution of the first–third inner-race harmonics; comparisons with IMCKD, ACYCBD, and FDS+MOMEDA further confirm more reliable component separation, offering robust non-contact diagnostic support for the predictive maintenance of rotating machinery.