Anomaly detection in heterogeneous media via saliency analysis of propagating wavefields

This work investigates the problem of anomaly detection by means of an agnostic inference strategy based on the concepts of spatial saliency and data sparsity. Specifically, it addresses the implementation and experimental validation aspects of a salient feature extraction methodology that was recently proposed for laser-based diagnostics and leverages the wavefield spatial reconstruction capability offered by scanning laser vibrometers. The methodology consists of two steps. The first is a spatiotemporal windowing strategy designed to partition the structural domain in small sub-domains and replicate impinging wave conditions at each location. The second is the construction of a low-rank-plus-outlier model of the regional data set using principal component analysis. Regions are labeled salient when their behavior does not belong to a common low-dimensional subspace that successfully describes the typical behavior of the anomaly-free portion of the surrounding medium. The most at tractive feature of this method is that it requires virtually no knowledge of the structural and material properties of the medium. This property makes it a powerful diagnostic tool for the inspection of media with pronounced heterogeneity or with unknown or unreliable material property distributions, e.g., as a result of severe material degradation over large portions of their domain.