Open Cut » Geology
Rockfalls pose significant risks to human lives, machinery and portal structures located at the toe of highwalls in open cut mines, with the potential for considerable financial losses. An efficient mitigation can be provided when the energy at first impact at the bottom of the affected highwalls and the block run-out distance on the mine floor are appropriately predicted. A common approach is to employ computer based simulations of rockfall trajectories and associated stochastic analyses to evaluate these risks.
This project investigated the feasibility for the implementation of Machine Learning (ML) models for fast and reliable predictions of rockfall hazard. Probabilistic rockfall simulations are performed to generate trajectories using high resolution 3D photogrammetric models of fifteen sub-vertical rock faces (highwalls) obtained from seven mine sites.
Natural slopes geometries are considered in this project, in contrast to the common approach in which the highwalls are represented by line segments with artificially introduced roughness values. An automated software solution is developed to extract 2D profiles as well as meaningful geometrical features from arbitrary 3D photogrammetry data of real highwalls. Consequently, rockfall risks can be predicted along a full strike length of highwalls. In total, 1,669 2D profiles capturing a wide range of slope geometries are extracted from these fifteen highwalls. A total of 4,550,500 rockfall trajectories are generated by conducting the associated kinematic simulations with respect to the natural slope profiles, before performing an automatised stochastic evaluation.
A meaningful split is performed on the stochastically evaluated simulation data to develop ML models for the target parameter (risk variables). Four representative highwalls are selected for model calibration based on an extensive data analysis. These highwalls have distinct characteristics and, overall, capture all possible ranges of wall geometric features. The remaining highwalls are used for model validation.
The block release height (falling height) as well as the slope local roughness and average angle are chosen as the principal features, i.e., ML input parameters. Kinematic energy at the first impact, the first impact position and the final run-out at the bottom of rock faces are considered as the target parameters. The application of several ML models is investigated for each individual target parameter and their performances compared. A multi-linear regression model shows the best predictive behaviour for the kinematic energy at the first impact and a multi-non-linear regression model with a quadratic term in the block release height shows the best predictions for the first impact position and the final run-out.
The results show a reasonable applicability of the approach for a fast prediction of rockfall hazard for arbitrary highwalls based on the fully automatised extraction of geometrical features of 3D photogrammetric data. The proposed framework allows improvement of the model by including real data for calibration and validation.