A Comparison of dimensionality reduction and blind source separation techniques for video-based modal identification

Abstract

Understading the dynamic properties of a structural system is indispensable for a reliable study of the structural behavior. Video-based structural dynamics identification has been effectively used as a key method for modal analysis in recent years.With several different approaches, the ones based on the blind source separation strategy have received increased attention for identifying structural characteristics. Blind source separation addresses the problem of separating or extracting the original source waveforms from a sensor array. Although the literature addresses several techniques to perform the source separation, only one of them (named complexity pursuit) is often employed for video-based solutions. This work aims to explore other blind source separation algorithms to perform video-based modal analysis. In order to perform the modal decomposition, a set of blind source separation methods is combined with different dimensionality reduction techniques for full-field high-resolution structural dynamics from video. Specifically, Principal Component Analysis (PCA) and NonnegativeMatrix Factorization (NNMF), two dimensionality reduction techniques, are used for video compression along with six source separation algorithms, resulting in twelve different frameworks tested over a laboratory cantilever beam structure and a bench-scale model of a three-story building structure. The blind source separation techniques used are: Complexity Pursuit (CP), Idependent Component Analysis (ICA), Second Order Blind Identification (SOBI), Second Order Blind Identification with Robust Orthogonalization (SOBIRO), Equivariant Robust Idependent Component Analysis (ERICA) and Algorithm for Multiple Unknown Signals Extraction (AMUSE). The twelve techniques based on dimensionality reduction and blind source separation algorithms are evaluated here using as the criteria comparison their mode shape, modal coordinates and MAC values. The main goal is to provide a range of alternatives for the vide-based structural dynamics evaluation. For specific algorithms, the results indicate that both dimensionality reduction techniques and the blind source separation methods play a major role in the mode estimation performance. In the experiment utilizing the cantilever beam structure, all expected modals were successfully identified using the algorithms based on PCA-CP, PCA-ICA, PCA-SOBIRO, PCA-ERICA, NNMF-CP, NNMFSOBI, NNMF-SOBIRO and NNMF-ERICA. For the second scenario, using the model of a three-story building structure, the methods that correctly perfomed modal analysis are based on PCA-CP, PCA-SOBI, PCA-SOBIRO, PCA-ERICA, NNMF-CP and NNMF-SOBIRO. It is suggested that The effectiveness of combining NNMF and blind source separation methods for modal analysis may be contingent upon the complexity of the system under investigation.