Integrating surface and borehole sensing for accurate microseismic localization: A velocity-model-free deep learning approach

With the increase in production depth and pressure, microseismic monitoring and location technology has become particularly important for ensuring the safety of oil and gas production. Therefore, we established a comprehensive well and ground microseismic location network specifically designed to accurately locate microseismic events induced by shale gas extraction. The proposed model, JGF-LocNet, integrates convolutional neural networks (CNNs), transformers, long short-term memory (LSTM) networks, and graph convolutional networks (GCNs) to achieve high-precision event localization with real-time computational efficiency, which closely reproduces the QuakeMigrate reference catalog. The method is velocity-model-free at its inference, but inherits velocity-model assumptions through training labels generated from the catalog. Unlike traditional methods that rely on velocity models, our approach requires only the input waveform, enabling JGF-LocNet to complete seismic event predictions within 0.3 s, with a positioning error of less than 40 m. Blind tests on two unseen arrays confirmed a <10% increase in AE-P90 while maintaining real-time latency. With an adequate number of training samples, JGF-LocNet is expected to deliver rapid, high-precision positioning with minimal errors. We conducted comparative analyses of the computational efficiency and positioning accuracy of our method against those of traditional approaches and other deep learning models. Our results demonstrate that the proposed method most closely reproduces the QuakeMigrate reference catalog in real time while retaining subsecond latency. Given the increased frequency and risk of dynamic disasters in high-pressure, deep, and complex geological structures, the real-time monitoring and precise localization of induced microseismic events provided by our proposed method are critical for ensuring industrial safety.