Preventive Model for Slope Failure Forecasting at the Quellaveco Mine Early Detection Approach Using MLP-PyTorch, Random Forest, and Logistic Regression with SMOTE

Noriega-Vidal E.M.; Jimenez-Calvanapon A.S.; Aranda-Vega C.M.; Castillo-Orbegoso M.O. · International Journal of Safety and Security Engineering (2026) · DOI: 10.18280/ijsse.160716

Resumen

Neural Network implemented with PyTorch (21–20–10–5–1 architecture, Batch Normalization, 20% Dropout), Random Forest (200 trees), and Logistic Regression, on 1,500 historical records (2022–2024) with 21 geotechnical, environmental, and operational variables. Environmental variables were calibrated with data from four SENAMHI stations (Calacoa, Moquegua, Quinistaquillas, and Ubinas) using inverse distance weighting (IDW) interpolation. Facing class imbalance (91.33% stable / 8.67% failure), SMOTE was applied exclusively to the training set, generating 2,192 balanced samples. On the held-out test set (n = 300, original distribution) the Multi-Layer Perceptron (MLP) reached Accuracy = 0.870, Recall = 0.500, and AUC-ROC = 0.750; Random Forest obtained Accuracy = 0.867 but Recall = 0.423; Logistic Regression achieved the highest Recall (0.731) and AUC-ROC (0.785). These performance values are consistent with the variable-importance results presented below. The Gini importance ranking is led by the Factor of Safety (0.1513), followed by GSI (0.1184), RQD (0.0731), and Internal Friction Angle (0.0729). Because FS < 1.3 is also one of the two criteria used to label the target variable, the leading position of the Factor of Safety is reported as an internal consistency check of the model and not as an independent predictive finding; the ranking of the remaining variables, which do not intervene in the labelling rule, is unaffected by this constraint.

Instituciones: Universidad Privada del Norte