Dados do Trabalho

Título

NEURAL NETWORKS FOR PREDICTING SLOPE STABILITY: A CASE STUDY IN NATURAL SLOPES

Resumo

Ensuring the natural stability of slopes is crucial for averting environmental catastrophes triggered
by landslides. These incidents not only lead to material losses but also imperil the lives of residents in
vulnerable areas. Grasping the factors influencing slope stability is essential for effective risk management
and disaster prevention. Given the substantial uncertainties in geotechnical data and parameters, achieving
an analysis with reduced error involves utilizing a comprehensive database detailing the impact of material
characteristics and climate variations in a specific region. Neural networks emerge as a valuable tool for
handling vast datasets and conducting safety factor analyses for natural slopes. To construct an Artificial
Neural Network (ANN) model, an extensive dataset was compiled, encompassing soil parameters (cohesion
and friction angle), geometric factors (soil layer height, saturation layer width), slope inclination, and soil
type. Input parameters were generated within established intervals from literature sources—for instance,
cohesion ranged from 0 to 20 MPa, slope height from 1 to 5 meters, and saturation width from 10 to 80%
of the slope height. A total of 6,600 simulations were executed using HYRCAN to determine Factor of
Safety (FOS) values. Subsequently, the dataset was divided, with 80% allocated for training and optimizing
model hyperparameters using a Bayesian optimization approach. The remaining 10% served as a validation
set, while the rest was designated for testing. Post-optimization, the best hyperparameter set achieved Mean
Absolute Percentage Error (MAPE), R², and Root Mean Square Error (RMSE) values of 1.02%, 0.999, and
0.00078, respectively. The optimal model comprised three hidden layers employing the linear rectifier
activation function (ReLu), each with 100 neurons. The model was then applied to an actual topological
map of Rio de Janeiro, yielding results that were not only interesting but also physically congruent.
Additionally, a sensitivity study was conducted to explore the impact of each input in determining FOS.
The outcomes revealed that these parameters varied within their domains and were interdependent.
Consequently, an Analytical Hierarchical Process (AHP) was employed to quantify these sensitivities.

Palavras-chave

Slope stability, risk management, neural networks, AHP.

Arquivos