Abstract
Rockbursts are characterised by the sudden and violent release of stress energy within a rock mass, often leading to the explosive displacement of rock fragments. These phenomena are commonly associated with seismic activities and are particularly prevalent in deep underground mining and tunnel constructions. Due to their frequent occurrence in these environments, rockbursts pose substantial and potential risks, necessitating predictive research to enhance safety measures. This study primarily focused on predicting rockburst occurrences through the application of artificial neural networks (ANNs). To achieve this goal, two distinct approaches for implementing neural networks were utilised: Matlab pattern recognition and Matlab classification learner. Both approaches were employed to identify the most effective neural network model for accurate predictions of rockbursts. Subsequently, the developed neural networks underwent multiple training and testing sessions using a collected and prepared dataset comprising 185 rockburst incidents. Following an extensive experimental phase, the study's findings revealed a notable contrast in performance between the two distinct approaches. The first approach, Matlab pattern recognition, delivered a neural network model with outstanding performance, achieving a training accuracy of 99.1%, a validation accuracy of 80%, and a test accuracy of 76%. Cumulatively, across these phases, the overall accuracy reached an impressive 92.7%. In contrast, the second approach, Matlab classification learner, developed a neural model with acceptable yet comparatively lower performance, with a training accuracy of 70.34% and a test accuracy of 70.12%. The comparisons between the two approaches underscored the superiority of the Matlab pattern recognition approach in developing highly effective artificial neural networks compared to the Matlab classification learner approach. This study emphasises the critical importance of developing efficient models capable of delivering accurate predictions. It underscores that relying on a single approach for implementing machine learning models does not guarantee optimal results. Instead, as demonstrated in this study, the preference is to employ two or more implementation approaches, facilitating comprehensive model comparisons that assist in identifying the most optimal model for predictive tasks.
Keywords: Rockbursts, prediction, artificial neural network (ANNs), Matlab pattern recognition, Matlab classification learner.