TL-BiRNN-Pick: An End-to-End Automated Microseismic Monitoring Framework for Underground Mines via Transfer Learning

Abstract Microseismic monitoring is essential for ensuring mine safety, yet the accuracy of P-wave phase picking is often hindered by low signal-to-noise ratios and complex waveforms in mining environments. To address the scarcity of labeled mining data, this study develops a transfer learning–based bidirectional recurrent neural network phase picker (TL-BiRNN-Pick) for end-to-end automated monitoring. The model employs an encoder–decoder architecture integrated with bidirectional recurrent neural networks. It leverages a pretrained model derived from the large-scale DiTing 2.0 natural earthquake data set and is fine-tuned for specific mining seismic signals. By integrating the DBSCAN association algorithm, enhanced Geiger localization, and double-difference relocation, the framework enables a seamless waveform-to-catalog automated workflow. Performance evaluation on 996 mining seismic waveforms yields a precision of 0.947 and an F1 score of 0.933, with a root-mean-square error of 8.23 ms and a standard deviation of 15.24 ms. In a 16-day real-world application at the Dongtan Coal Mine, the framework identified 1271 microseismic events—a three-to-fourfold increase over traditional methods—and improved the magnitude of completeness to −1.64, representing a 0.24 magnitude unit enhancement. This research demonstrates the robust capability of the TL-BiRNN-Pick framework in enhancing the automated monitoring and safety assessment of underground mines.