Reference-supervised vision-based vibration measurement and time–frequency enhancement for variable-speed bearings under coarse ROI localization

To address the issues of Region of Interest (ROI) sensitivity, background interference, and unclear time–frequency structures in visual vibration analysis of variable-speed rolling bearings, this article proposes a loosely localized ROI-based visual vibration recovery and multiscale time–frequency enhancement method under a “training-with-reference, vision-only inference” framework. During training, a contact reference signal supervises the mapping from video motion to mechanical vibration; during inference, only a loosely localized ROI from high-speed video is required. The method first recovers a one-dimensional vibration signal via dense optical flow, rigid-body motion suppression, block-level residual motion representation, and a dynamic fusion network. It then constructs a multiscale time–frequency representation and employs a pyramid fusion enhancement network to enhance frequency structures consistent with the mechanical response. Experimental results demonstrate stable recovery under ROI variations and shifts, high consistency with the contact reference, and clearer, more continuous time–frequency structures, outperforming traditional Lucas–Kanade optical flow and short-time Fourier transform. This provides an ROI-adaptive, physically interpretable noncontact visual analysis approach for fault feature extraction in variable-speed bearings.