چكيده انگليسي :
This research aims to present an empirical model for predicting the penetration rate of jumbo drills in underground mines
under conditions of geomechanical uncertainty. Initially, by reviewing the existing literature and models for predicting
penetration rates, the factors affecting the penetration rate were identified, including rock mass properties, drilling operation
parameters, and other ancillary parameters. Due to the multitude of rock mass characteristics and the high cost of including
all these parameters, the Rock mass Drillability Index (RDi) was used as a representative of the rock mass properties in the
analyses. Field data were collected from the two mining complexes of Bama Lead and Zinc and Sormeh, resulting in a
database of 737 drill holes with two diameters of 52 and 64 millimeters. In this process, prior to drilling, information related
to discontinuities in the face was collected, and samples were taken for laboratory studies. During drilling, operational
parameters of the jumbo drills, including hydraulic oil pressure for feed, rpm, percusion, and drilling duration for each
borehole were recorded. After forming the database, the relationship between penetration rate and the Rock mass
Drillability Index was examined using classification and regression tree algorithms (CART). The results showed a
significant correlation between these two parameters, with a linear relationship yielding a high coefficient of determination
(R²) above 0.9. To develop the empirical model for predicting penetration rates, the database was divided into two parts:
80% of the data was used for model development and 20% for validation. Several statistical parameters, such as the
coefficient of determination, root mean square error (RMSE), mean absolute error (MAE), percentage of mean absolute
error (MAPE), and variance accounted for (VAF), were used to assess the accuracy of the models. Initially, multiple linear
and nonlinear regression models were proposed for predicting penetration rates. The linear and nonlinear regression models
were developed, with the nonlinear model performing better, yielding an R² of 0.88 and lower relative error. Subsequently,
models were developed using machine learning algorithms, including multilayer perceptron artificial neural networks
(MLP), support vector regression (SVR), and random forests (RF). The results indicated that the neural network model
outperformed regression models, with an R² of 0.92 and a root mean square error of 0.16 for training data and 0.09 for
validation data. Additionally, the support vector regression model showed a higher accuracy with an R² of 0.94 compared
to regression models. Alongside machine learning methods, deep learning techniques, including deep random forests
(DRF), deep support vector regression (DSVR), and fully connected deep artificial neural networks (FCDNN) combined
with genetic optimization algorithms, were also utilized. The results demonstrated that the accuracy of deep learning
methods surpassed that of other models, with the deep neural network yielding the best results. The deep neural network
achieved an R² of 0.94 for both training and testing data, with root mean square errors of 0.16 and 0.2, respectively. To
assess the uncertainty of the models, Monte Carlo simulation was employed. The probability distribution function of the
output parameters was specified, and the simulation space was prepared by generating 2000 random samples. At a 95%
confidence level, the uncertainty range of the actual average values was compared with the values predicted by the models....
