توصيفگر ها :
آتشباري , رگرسيون , شبكه عصبي مصنوعي , خردايش , آناليز تصويري , SPSS
چكيده انگليسي :
Blasting operations are one of the most important operations in mining. The fragmentation
resulting from blasting directly affects mining costs, including loading, hauling, and
crushing. In addition to the economic costs, blasting has undesirable effects such as ground
vibration and air blast, which pose risks in mines. Therefore, mining experts have always
sought to provide a suitable function to predict the fragmentation resulting from blasting
accurately. In this study, 64 blasting operations in the Saryguni gold mine were investigated.
Initially, a database was formed, including burden, number of holes, explosive quantity, free
face area, RQD index, and point load pressure. Then, imaging of the fragmentation products
of each blast was performed in such a way that images were taken from each blast in different
directions and angles. After imaging, image analysis was performed using Split-Desktop
software, considering the effects of light and shadow by changing the direction and angle of
imaging, which ultimately revealed that imaging from a fixed angle does not accurately
represent the actual size of the fragmentation products and the presence of shadows can
affect the analysis values. After analyzing each blast, D20, D50, and D80 passing
percentages were considered as dependent variables, and other factors mentioned were
considered as independent variables. After determining the independent and dependent
variables, the data were normalized, and the correlation between them was determined using
Pearson correlation test. Then, three methods, including multiple linear regression, multiple
nonlinear regression, and artificial neural network with distribution function, were used to
predict the performance of each model, considering the coefficient of determination (r2) to
determine the degree of correlation in the relationships under study. Multiple linear
regression was performed using SPSS software, multiple nonlinear regression was
performed using Datafit software, and for the artificial neural network, deep learning
technique was used in two different cases, one hidden layer and two hidden layers. After
determining the presented relationships and the results obtained, the coefficients of
determination of all methods were compared, showing that linear and nonlinear regressions
provide similar values and the neural network with two hidden layers performs better. After
comparing the results of each method, the correlation between independent variables and
different passing percentages was also examined, showing that the highest correlation is for
the D80 value. Finally, for a more detailed examination of the prediction results of the last
13 blasts using the neural network with two hidden layers, the Kazi-Ram model and image
analysis were compared, revealing that the neural network can predict blasting results much
more accurately.
