Harnessing machine learning for seismic event discrimination in deep underground mining: a case study from Western Australia

Abstract

This paper presents a comprehensive study on applying machine learning (ML) techniques to discriminate seismic events in deep underground mining from blast and noise records using data collected from the Vivien gold mine in Western Australia. The dataset/catalogue comprises parameters derived from signals recorded by an ESG microseismic monitoring system, encompassing various parameters such as magnitude, seismic moment, total radiated energy and more, totalling 33,298 records. A rigorous statistical analysis was conducted to address potential multicollinearity issues and identify key input variables. Additionally, the local outlier factor (LOF) method was utilised to remove anomalies, ensuring homogeneity in the dataset for further analysis. The synthetic minority oversampling technique (SMOTE) was then applied to address imbalanced datasets, particularly in classifying seismic record types as seismic events, blasts or noise attributed to rockfall. Eight robust ML algorithms were employed to develop classifiers for predicting record class types. The performance of each model was evaluated using statistical indices, ultimately identifying random forest (RF) as the most accurate method for distinguishing between different record types. Furthermore, a user-friendly graphical user interface (GUI) was also developed to facilitate data analysis based on the proposed RF model, enhancing the interpretation of microseismic monitoring results in practical applications. This study underscores the efficacy of ML approaches in seismic event discrimination to ensure the seismic dataset is clean and reliable for use in geotechnical assessments of seismic hazards and seismic characteristics at the mine.

Keywords: seismic event, microseismic monitoring, deep underground mining, machine learning, random forest