NeuralCDE-Based Bench Stability Mapping for Open-Pit Mines Using Photogrammetric and Blast Vibration Data
Authors
Files
Abstract
Stability of an open-pit bench is influenced by surface deformation, discontinuity-controlled kinematics, and short-duration blast vibration, but these observations are acquired at different times and spatial resolutions. This study applies a neural controlled differential equation model to integrate irregularly sampled photogrammetric and vibration descriptors into cell- based screening maps. Photogrammetric inputs describe elevation residuals, horizontal displacement, roughness, and discontinuity traces; vibration inputs describe peak particle velocity, dominant frequency, duration, and charge-delay characteristics. The reported monitoring dataset covers one limestone bench over 14 weeks and includes 2,184 cells and 36 blast events. The reported macro-F1 was 0.887 for the proposed model, compared with 0.831 for a gated recurrent unit and 0.806 for a temporal convolutional network. High-risk omission decreased from 14.6% to 6.8%, and the proposed model had a mean absolute index error of 0.056. Following blast events B18 to B20, the mapped high-risk area increased from 9.4% to 18.7%, with a concentrated warning belt near the central berm. The output is an instability-oriented screening index rather than a mechanical factor of safety or a validated failure probability. These observations suggest potential for combining asynchronous monitoring streams in inspection planning. Their interpretation remains limited by a single-bench study, incomplete temporal-split documentation, and locally selected warning thresholds. Independently documented monitoring records and prospective evaluation are required before general operational performance can be established.
