Real-time microseismic source localization in deep tunnel using physics-informed neural networks

Microseismic (MS) monitoring has become an essential technique for preventing rockburst hazards in deep-buried tunnel engineering. However, accurate and real-time localization of MS sources remains challenging due to the presence of void structures and dynamic changes in sensor layout. Traditional travel-time difference methods struggle to balance travel-time modeling accuracy and computational efficiency, while existing deep learning approaches often rely on large volumes of labeled data and are unsuitable for dynamic monitoring environments. To address these challenges, this study proposes a novel travel-time modeling framework for real-time MS source localization based on physics-informed neural networks, referred to as MSLoc-PINN. By incorporating physical constraints governed by the eikonal equation into the loss function of a neural network, the proposed method achieves high-accuracy travel-time modeling. The introduction of residual connections and an adaptive weighting mechanism further enhances the model’s stability and convergence performance. Moreover, by integrating the transfer learning strategy, the model can rapidly adapt to variations in sensor layouts and velocity structures that occur during tunnel excavation, further improving training efficiency. Numerical experiments on gradient, layered, and void-containing velocity models demonstrate that MSLoc-PINN achieves lower travel-time prediction errors than conventional methods, including the fast marching method and uniform velocity model. Practical engineering applications further validate the potential and practical feasibility of MSLoc-PINN in dynamic field conditions. The results indicate that MSLoc-PINN enables high-precision, real-time MS source localization, providing valuable technical support for safety monitoring and hazard prevention in complex underground engineering environments.