Abstract
Microseismic (MS) source localization is essential for mitigating rockburst hazards and optimizing the reconstruction of energy reservoir. However, conventional convolutional neural networks (CNNs) remain highly sensitive to noise, and LSTM-FCNN models fail to effectively capture the spatial–temporal correlations among geophone arrays, resulting in degraded performance under low signal-to-noise ratio (SNR) conditions, typically below 0 dB. To overcome these limitations, This paper presents a multi-scale convolutional localization model with coordinate attention (MSCAR-Net). The model exploits the heterogeneous responses of noise and the correlated features of effective signals through multiscale convolutions, while a feature fusion strategy balances noise suppression and signal enhancement. In addition, the coordinate attention (CA) mechanism establishes spatial-temporal correlations between geophone positions and signal sequences, significantly improving localization robustness under low-SNR conditions. Experimental results demonstrate that MSCAR-Net achieves a stable localization error of approximately 7.9 m at SNR ≥ −5 dB, representing an 87% improvement over the baseline CNN and a 44% enhancement compared with the LSTM-FCNN. Field validation using known perforation locations in the S-H2 hydraulic fracturing project confirms an average error of 8.2 m, while real-time localization of 65 events further validates the method’s applicability as a measurement solution for low-SNR MS monitoring systems.
Funding Information
  • Jilin Province Science and Technology Development Plan Item, China (20250201058GX)

This publication has 37 references indexed in Scilit: