GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.2 GNGTS 2024 Exploring rockfall precursors through unsupervised deep-learning clustering analysis G.M. Adinolf 1 , C. Comina 1 , S.C. Vinciguerra 1 1 Department of Earth Sciences, University of Turin, Turin, Italy Studying rock mass stability is greatly enhanced by the powerful tool of seismic monitoring. This technique allows for contnuous recording, making it easier to analyse spatotemporal actvity related to gravitatonal instabilites. Seismic monitoring is essental for detectng and assessing damage and cracking processes as precursors to macroscopic failures. In this research, we present inital fndings from seismic monitoring conducted at a test site where a rockfall occurred shortly afer deploying a small-aperture array of three seismic statons, covering approximately 100 meters. These statons were equipped with a tri-axial velocimetric sensor and data-loggers sampling at 250 Hz. The specifc focus of the seismic array was to survey potental instabilites originatng from structural weaknesses, including deformaton bands, joints, and lithological contacts. The installaton of the site-specifc seismic array preceded a rockfall event by about a month. The rockfall occurred at the lithological contact between folded gneisses and a unit of dolomitc limestones, predominantly composed of dolomites and dolomitc saccharoid marbles. The seismic signature of the rockfall persisted for approximately 10 seconds, and spectral analysis revealed the occurrence of multple sub-episodes of slip triggered by the inital rupture. No apparent correlatons between precursory actvity and the rockfall occurrence were identfed through traditonal seismological approaches. In response to this challenge, we applied a recently developed unsupervised deep-learning method for clustering signals in contnuous multchannel seismic tme series. This method combines a deep scatering network for automatc feature extracton with a Gaussian mixture model for clustering. Our successful applicaton of this approach led to the identfcaton of seismic signals associated with the rockfall, encompassing various slip episodes, along with their initaton and propagaton. Additonally, we detected precursory actvity occurring approximately 2 hours before the rockfall, consistently identfed through the clustering analysis (Fig. 1).