Multi-objective variational autoencoder: an application for smart infrastructure maintenance

Abstract Multi-way data analysis has become an essential tool for capturing underlying structures in higher-order sets where standard two-way techniques often fail to discover the hidden correlations between variables multi-way data. We propose a multi-objective variational autoencoder (MO-VAE) method smart infrastructure damage detection and diagnosis sensing based on reconstruction probability of deep neural network (ADNN). Our fuses from multiple sensors one ADNN at which informative features are being extracted utilized identification. It generates probabilistic anomaly scores detect damage, asses its severity further localize it via new localization layer introduced ADNN. evaluated our laboratory-based real-life structural datasets area health monitoring purposes. The was collected deployed acquisition system cable-stayed bridge Western Sydney, reinforced concrete cantilever beam replicates major components Sydney Harbour Bridge laboratory building structure obtained Los Alamos National Laboratory (LANL). Experimental results show that proposed can accurately damage. also able estimate different levels severity, capture locations unsupervised aspect. Compared state-of-the-art approaches, shows better performance terms localization.