Application of a Data Augmentation Technique on Blast-Induced Fly-Rock Distance Prediction

Abstract

AbstractFly-rock induced by blasting is an inevitable phenomenon in quarry mining, which can give rise to severe hazards, for example, causing damage to buildings and human life. Thus, successfully estimating fly-rock distance is crucial. Many researchers attempt to develop empirical, statistical, or machine learning models to accurately predict fly-rock distance. However, for most previous research, a worrying drawback is that the amount of data related to fly-rock distance prediction is insufficient. This is because the measurement work of fly-rock distance is costly for manpower and material resources. To deal with the problem of data shortage, we first separated the original data set that was collected from four granite quarry sites in Malaysia into two parts, i.e., the training and testing sets, and then adopted a data augmentation technique termed tabular variational autoencoder (TVAE) to augment the amount of the training (true) data, so as to generate a fresh synthetic data set. Subsequently, we utilized several statistical visualization methods, such as the boxplot, kernel density estimation, cumulative distribution function, and heatmap, to testify to the effectiveness of the synthetic data generated by the TVAE model. Lastly, several commonly used machine learning models were developed to verify whether the mixed data set—which is obtained by merging the training and synthetic data sets—can benefit from the addition of the synthetic data. The verification work is implemented on the testing data set. The results demonstrate that the size of the training data set has increased from the initial 131 to 1000 to obtain a synthetic data set, and the statistical methods proved that the synthetic data set not only preserves the inner characteristics of the training data set but also generalizes more diversities compared with the training data set. Further, by comparing the performance of five machine learning models on three data sets (i.e., the training, synthetic, and mixed data sets), it can be concluded that the overall performance of all machine learning models on the mixed data set outperforms that on the training and synthetic data sets. Consequently, it can be asserted that the application of the data augmentation technique on the fly-rock distance issue is fruitful in the present study and has profound engineering application value.KeywordsFly-rock distance predictionData augmentationTabular variational autoencoderMachine learning prediction models