Automated mode tracking via supervised classification and adaptive parameter calibration for seismic monitoring with sparse sensors

One of the most important issues to address in the practical implementation of permanent dynamic Structural Health Monitoring (SHM) systems is undoubtedly that of Mode Tracking (MT). Indeed, the influence of environmental and random fluctuations, as well as the uncertainty inherent in the identification algorithms themselves, especially the spill-over effects linked to unmodeled dynamics, can make it difficult to disentangle the various modal behaviours. This separation process, i.e the MT procedure, involves comparing vibration mode estimates with a reference set of modal properties. Although this operation can be straightforward for simple structures, in many practical applications of structural engineering, when there is strong modal concentration (e.g. lattice structures) or high geometric and mechanical complexity (e.g. monumental buildings) greater challenges arise, which grow in the presence of sparse sensor setups (civil structures in general), the superposition of exogenous frequency components (industrial structures, bell towers etc.) and environmental fluctuations. This study presents an innovative MT methodology that combines supervised classification, using advanced machine learning algorithms, with adaptive multi-threshold calibration to overcome the limitations of current MT techniques. The approach incorporates clustering analysis to characterize vibration modes by their natural frequencies and mode shapes, ensuring accurate identification and rejection of spurious data. The method was validated with a simplified numerical model and then demonstrated on a baroque monumental structure equipped with a long-term monitoring system. In addition to being efficient and robust compared to traditional techniques, the proposed procedure is effective for automating the monitoring of modal parameters in SHM systems, even in scenarios with limited sensor deployments.