An overbreak during the construction of underground mining tunnels is a common geotechnical and operational problem, which is caused by a combination of geological, geotechnical, structural and operational factors, in which partial or reduced information is available, thus conditioning tunnel stability and consequently the safety of personnel during construction. Additionally, studying an overbreak during early stages allows to validate assumptions applied during engineering stages.
Throughout history, different methods have been proposed for the overbreak estimation, these ranging from an empirical, analytical (including numerical modelling), observational or even through the application of machine learning.
This work proposes a different approach to most of the studies carried out, which usually consider an average or expected value of overbreak. On this occasion, Deep Learning architectures are used to characterize and predict the complete geometry of the tunnel based off of a training carried out using point clouds of the sectors already excavated.
The results obtained show that it is possible to use autoencoder-type architectures to carry out the characterization and prediction of the tunnel’s geometries from point clouds of previously excavated sectors, which has a relevant value for back analysis and potentially predictive analysis, which would in turn impact tunnel stability and/or safety in the different operation cycles during the construction of underground mining galleries and/or tunnels and civil works’ projects and operations.
