Nonnegative matrix factorization-based blind source separation for full-field and high-resolution modal identification from video

Abstract

Traditional modal analysis requires physically attached sensors for data acquisition and vibration-based monitoring. Although traditional modal analysis presents well-established techniques for dynamics analysis, they can impose mass-loading effects on lightweight structures and increase budgetary demands on the maintenance of such data acquisition systems. Recently video-based techniques have become of increasing interest in the identification of the dynamic properties of infrastructures with arbitrary complexity. However, most applications rely on frame by frame tracking of fixed speckle targets to derive time-varying physical parameters. This imposes serious limitations for real-world applications, especially in scenarios where the structure is out of reach. Therefore, to address these issues, we propose a novel output-only operational modal analysis method based on vision-based blind source separation scheme. The proposed algorithm makes use of each pixel as a potential measurement point. This enables an increase in the spatial density of sensors conventionally used on a structure by orders of magnitude. This simultaneous processing of all pixel time-series derives full-field high-resolution mode shapes instead of low spatial resolution mode shapes achieved when measuring a limited number of discrete locations with typical sensors. Compared to other approaches, we propose a blind source separation scheme simpler than the ones based on phase extraction and complex steerable pyramids that still capable of disentangling local structural vibration from video measurement only. Moreover, a simple method to magnify and visualize independent vibration modes is introduced using the extracted modal information only. We validate our method by laboratory experiments on a bench-scale building structure and a cantilever beam. The results demonstrate that the proposed technique can decompose high-resolution modal parameters, visualize and reconstruct even those weakly-excited vibration modes.