Autoencoder-based modal clustering and structural states classification using variational approximation and triplet loss

With the advancement of sensor technology, vibration-based Structural Health Monitoring (SHM) has demonstrated its exclusive effectiveness due to its role in maintaining structural integrity during the expected lifecycle. Among them, the modal analysis provides crucial information about the changes in dynamic characteristics and is utilized for detection, identification, diagnosis, and prognosis in field applications. However, the changes in modal parameters can sometimes be insignificant and necessitate further examination of mode shapes and/or their derivatives. The latest developments in machine learning (ML), particularly clustering techniques, offer promising solutions for distinguishing between the subtle changes. This paper aims to develop an autoencoder-based modal clustering to discriminate structural damages from uncertainties and environmental influence. Furthermore, variational autoencoder (VAE) expands a continuous and smooth space by using variational approximation. To further leverage the dimensionality in latent space, deep metric learning is introduced by the triplet loss, embedding a more intuitive and interpretable space. In addition, both numerical simulation and experimental verification have been conducted to showcase the effectiveness of clustering changes between various damage scenarios. The results demonstrate the capability of the proposed approach for structural inspection during long-term monitoring.