The Valkenburg calcarenite mines in the Netherlands, dating back to the 15th century, represent an important historical monument. Initially, the mines were exploited using the room-and-pillar method, and nowadays serve as a tourist attraction. Throughout the years, these mines have suffered structural instabilities due to unregulated extraction practices and the inherent soft nature of the calcarenite (σci = 2.5 MPa). This study focuses on investigating the long-term creep behaviour of mine pillars using a dataset collected from an advanced fibre optic monitoring system, covering four years of continuous observations across 81 pillars. Traditional geospatial and statistical analysis results provided weak correlations and conflicting results between stability indicators (pillar height, effective width, and overburden thickness) and the observed deformation trends, necessitating a more advanced approach. To address this, Long Short-Term Memory (LSTM) neural networks were employed to predict pillar deformation and assess long-term stability risks. The LSTM models successfully captured the complex, non-linear deformation patterns of the mine pillars, accurately predicting deformation values up to approximately 150 days, after which deviations from the observed values occurred. The LSTM models also faced challenges in certain training sets where deformations were associated with pronounced seasonal patterns. The findings highlight the potential of continuous monitoring combined with predictive modelling to improve early warning and preserve the structural integrity of historical underground sites, while ensuring their safe utilisation for future generations.
4th International Symposium on the Preservation of Monuments and Historic Sites
Underground Monuments and Rock Engineering