Unravelling Precursory Rockfall seismic signatures via multiscale clustering analysis

Seismic monitoring is an effective tool for studying rock mass stability, playing a crucial role in detecting the precursory assessment of damage and cracking processes preceding and accompanying macroscopic failures.

We first present the seismic monitoring results from the Lorgino Quarry in Crevoladossola (NW Italy) where a 6000 m³ rock-fall occurred on January 26, 2023 shortly after deploying a small-aperture array (about 100 meters) of three seismic stations, equipped with a tri-axial, velocimetric sensor and data-loggers sampling at 250 Hz. The rock fall took place about a month after the site-specific seismic array installation at the lithological contact between folded gneisses and a dolomitic limestones unit, mainly composed of dolomites and dolomitic saccharoid marbles. The rockfall seismic signature lasted 15s and the spectral analysis shows the occurrence of multiple sub-episodes of slip triggered by the initial rupture. 

As no obvious correlations between precursory activity and the rockfall occurrence were observed via traditional seismological approaches, we applied an unsupervised deep-learning method that combines a deep scattering network, for automatic feature extraction, with Gaussian mixture model clustering. This approach successfully identified low-amplitude signals occurring nearly one hour before the rockfall, nearly undetectable in raw seismic records and likely associated with a nucleation phase occurring well before the acceleration to failure.

In order to investigate the physical mechanisms driving the nucleation phase, we carried out rock deformation laboratory experiments, where marble cylindrical samples (100x40mm) from the quarry were triaxially loaded in compression to failure at constant effective pressure (20MPa) while an array of 16 Piezoelectric Transducers recorded the ongoing Acoustic Emissions (AE). The time and spatial distribution of AE reveal the nucleation and growth of patches led by limited occurrence of low energy AE events and the coalescence of microfractures into cm-scale macroscopic ruptures planes leading to AE clustering and stress drop and a peak in number of events and energy. Preliminary source mechanism analysis, carried out by developing an automated focal mechanism inversion workflow for AE based on P-wave first-motion, integrating polarity and amplitude measurements, suggests that the inverted focal mechanisms are stable and broadly consistent with the imposed stress conditions, highlighting the potential of the workflow to improve source mechanism quality by identifying and excluding unreliable solutions.