Multi-Class Vibration Event Classification in a Monitored Heritage Steel–Glass Hall

On 4 July 2025, a liquefied petroleum gas (LPG) explosion occurred at a petrol station located 6.4 km from the Marcus Aurelius Exedra Hall in Rome. The semi-elliptical steel–glass pavilion, equipped with a permanent accelerometric monitoring system, captured the resulting ground shock across multiple sensors. This study investigates the recorded response through a combined analysis of wave propagation, structural filtering, and multi-class event discrimination under real monitoring conditions. The measured signals show a short-duration impulsive transient followed by a rapidly decaying tail, clearly distinguishable from longer-duration environmental excitations such as thunderstorms. To support physical interpretation, a finite element model calibrated from ambient vibration data is used to compare measured and simulated responses. A feature-based calibration procedure aligns key statistical and spectral descriptors of synthetic excitation signals with measured response characteristics, including RMS amplitude, dominant frequency content, band energy distribution, and peak-related features. For event classification, two complementary machine learning branches are developed: a lightweight 1D-CNN trained on real monitored data, and a physics-informed XGBoost classifier trained exclusively on calibrated synthetic structural responses using response-derived statistical, spectral, and temporal ratio descriptors. A hybrid z-score anomaly detector bridges the synthetic-to-real domain gap, enabling deployment on unlabeled field recordings. The XGBoost classifier achieves 90% window-level accuracy on synthetic test validation, while the hybrid z-score–XGBoost framework successfully identifies a confirmed blast event and a thunderstorm episode in real data, demonstrating the practical transferability of the physics-informed training strategy.