Differentiating between microseismic and blasting events is a classical task in underground excavation monitoring. Most of the machine learning methods utilized for this task are supervised and require labeled data. To this aspect, this study proposes a fast and automatic unsupervised machine learning method based on sparse graph spectral clustering theory. The proposed method segments the microseismic–blasting dataset using the eigenvectors computed from the sparse graph Laplacian matrix of the database. Unlike conventional spectral clustering, which operates on a dense similarity matrix, the proposed method leverages a sparse similarity matrix, leading to better results. Using a well-documented six-dimensional microseismic–blasting monitoring dataset from a coal mine in China, the effectiveness of the method is demonstrated. An overall segmentation accuracy of 97%, and a precision of 96% and 98% are achieved for blasting and microseismic events, respectively, with only 0.003 s run time; hence, it is fast. The method does not require any labeled data or hyperparameters to be manually tuned; hence, it is automatic. The results are benchmarked against 10 supervised machine learning algorithms. It is found that the method performs similar or better than supervised machine learning techniques in terms of accuracy while being approximately 100–1,000 times faster. In terms of accuracy versus run-time trade-off, the method achieves a Pareto front. Overall, this study provides a useful strategy for the segmentation of microseismic and blasting events in future underground monitoring projects without requiring any labeled data.
Sharma et al. (Tue,) studied this question.