Seismic-While-Tunneling imaging for coal mine roadways via a physics-based passive seismic imaging workflow assisted by deep learning

The disparity between rapid TBM excavation rates and static geological forecasting creates a significant safety gap in deep coal mine tunneling. This study presents a Seismic-While-Tunneling (SWT) framework that utilizes TBM cutterhead vibrations as a continuous passive source for look-ahead imaging during excavation. To address the low signal-to-noise ratio inherent in mechanical sources, we propose a physics-based passive seismic imaging workflow assisted by deep learning. First, a specialized Convolutional Neural Network (NoiseNet) is employed to segregate effective seismic body waves from non-stationary mechanical clutter based on signal morphology. The screened signals are then processed via seismic interferometry and 3D diffraction stack migration, ensuring that the core wavefield reconstruction remains governed by wave-propagation physics. Finally, a SOLOv2 model is applied to automate the instance segmentation of geological anomalies from the migration profiles. Field validation across eight monitoring cycles at the Shoushan No. 8 Mine demonstrates that the system effectively suppresses false triggers compared to traditional thresholding. The framework achieved a Mean Absolute Error (MAE) of 3.94 m in positioning geological hazards, demonstrating its potential as a continuous geological foresight workflow for engineering guidance.