Prediction of boulder generation in opencast bench blasting using artificial neural network and their limitation

• Main Authors: Murldhar, Bhatawdekar Ramesh, Mohmad, Edy Tonnizam, Kumar, Abhijeet, Kumar, Sinha Rabindra, Armaghani, Danial Jahed, Pathak, Pranjal, Dan, Firdaus Md
• Format: Article
• Published: Books and Journals Private Ltd. 2021
• Subjects: TA Engineering (General). Civil engineering (General)
• Online Access: http://eprints.um.edu.my/36093/; https://www.scopus.com/inward/record.uri?eid=2-s2.0-85123994496&partnerID=40&md5=cbc8a052f94bd946d7689462e2c50aec
• Institution: Universiti Malaya

This paper discusses the various artificial neural techniques used to analyze 285 blasting data set from limestone quarry in Thailand consisting of blast design data and percentage of boulders as blast performance criteria. In the beginning, the data sets have been divided into train and test sets using genetic algorithm to maintain their statistical properties. Five-fold cross validation technique has been used for the selection of the network configurations and the regularization constant. Step by step analysis of data has been carried out. Four types of models are used for analysis namely - neural networks with whole set of features, neural networks with feature transformation using principal component analysis, neural networks with feature selection using information gain by decision trees and neural networks with feature selection using forward search. Neural network with feature selection using forward search, produced the best results among the four models. However, the model has not been able to produce any significant improvement in the results. The analysis shows that there exists an insignificant correlation and mean square error values with the collected data samples from the blast results of the quarry. The methods to forcibly produce significant mean square error and correlation values, that show apparently good results, have been shown. However, such models are not fit for generalizing the results. These models will not be able to predict the results for new and unnoticed inputs. © 2021, Books and Journals Private Ltd.. All rights reserved.