Enhancement of Automatic Signal Classification by Data Augmentation from Collapsed Mine

Abstract

Summary Microseismic monitoring is a promising technique for detecting signs of mine collapse. Recently, rapid progress in machine learning is accelerating the ability of microseismic monitoring for safety monitoring of mines. However, most mines in operation are structurally stable and therefore risky microseismic signals are seldom recorded unlike other mining-related signals. This makes it difficult to apply data-oriented machine learning techniques to the safety monitoring of mines. In this study, we try to solve this problem by augmenting collapse-related data obtained from a different mine where the collapse has occurred. We also introduce a new attribute named pseudo frequency to properly characterize microseismic signals and minimize processing time. The pseudo frequency attribute can provide frequency information without the Fourier transform. Automatic classification is performed by using a Random Forest model and the optimized model is designed through hyper-parameter tuning. By putting more weight on the risky microseismic signals, we achieved a high recall score, which is the most important goal for safety monitoring. From the application of our algorithm to the simulation of actual safety monitoring, we can confirm that the developed classification technique will be very suitable for the real-time safety monitoring in mines.