Using machine learning approach for microseismic events recognition in underground excavations: Comparison of ten frequently-used models

Correctly distinguishing microseismic and blasting events in underground excavations is fundamental to subsequent geophysical analysis activities such as rock burst early warning and tunnel deformation monitoring. Conventional distinguishing approaches such as waveform analysis require operating personnel strong physical and signal processing knowledge for predetermining man-crafted identification criteria, which may not achieve decent recognition accuracy since the collected signals are always varied. In addition, increasingly growing monitoring data cause conventional methods to be time-consuming and inefficient. Based on these considerations, machine learning approach was introduced to efficiently and accurately distinguishing microseismic/blasting events without developing explicit identification instructions for signals. This study investigated the performances of ten frequently-used machine learning models for microseismic/blasting events recognition. A complete identification process encompassing feature selection, model training, hyperparameter tuning was demonstrated to achieve ten models' precedence measured by five selected performance indicators. Then, models' performances were comprehensively considered by a fuzzy evaluation model to generate final conclusions: The Logistic regression model yielded the best performance for microseismic/blasting events identification while the Gaussian process model did the worst. The results gained from this study should serve to encourage machine learning in microseismic/blasting events identification, as well as give geological engineers reliable prior knowledge in model selection.