Structural vibration measurement based on improved phase-based motion magnification and deep learning

Complex and changeable environments during structural vibration processes poses great difficulties to traditional vision-based vibration measurement methods, especially for small structural vibrations. This paper proposes an improved phase-motion magnification (IPMM) and deep learning-based Recurrent All-Pairs Field Transforms (RAFT) method to identify the structural fullfield displacement and modal parameters. The IPMM algorithm has the features of effectively identifying frequency bandwidths, as well as effectively suppressing background noise and illumination changes. First, the video image of structural vibrations was captured using a camera and preprocessed. The IPMM algorithm was used to amplify motion in videos by selecting the frequency bandwidths of interest and the magnification factor. Subsequently, the RAFT was employed to calculate the full-field optical flow and structural displacement time history of the region of interest from the magnified video. Finally, post-processing tasks, e.g., motion normalization of displacement time history and identification of modal parameters, were carried out. The proposed method was verified by laboratory experiments. The results indicate that the IPMM not only effectively magnifies the small vibrations but also exhibits the advantages of suppressing background noise in non-motion regions and resisting changes in illumination conditions; the improved IPMM combined with the RAFT significantly improves the identification performance of structural full-field displacement and modal parameters.