矿山智能化建设大背景下，微震监测作为冲击地压等复杂动力灾害智能预警体系建设的基础保障技术现已得到广泛应用，如何基于微震监测数据实现矿山动力事件的高效精确捕捉与定位是当前研究的重点与难点。为解决矿山低信噪比震动波到时自动拾取不准、定位结果受人为因素干扰等问题，结合大数据与深度学习相关理论与方法，初步尝试建立基于深度迁移学习的矿山微震实时自动定位方法。设计了基于百万条地震波数据的矿山微震到时自动拾取初始模型，为进一步使该模型适用于矿山微震定位与信息解析，结合所建立的一万余条矿山微震到时拾取数据集，构建了矿山震动波到时自动拾取模型，实现矿山震动波P波到时自动精准拾取。在此基础上，设计了微震定位台站自动优选方法，提出定位台站波速自动微调策略，实现了矿山微震事件自动精准定位。以内蒙古某煤矿强开采扰动工作面的顶板爆破数据为验证对象，结果证明自动定位算法在水平空间平均定位误差为27.88 m，三维空间平均定位误差为28.40 m，满足矿山冲击地压等动力灾害的微震定位精度需求，有效降低台站波速标定精度不足对事件定位误差的影响，鲁棒性较强，并将耗时数分钟的人工定位缩短至200 ms内完成，初步实现了矿山微震事件的自动实时定位。研究结果可为矿山动力灾害微震信息准确解析挖掘与智能预警提供可靠技术支撑。

Under the background of mining intelligentization, the seismic monitoring technique has been widely used as the basic support for constructing the intelligent early warning systems for dynamic disasters such as rock bursts. The key issue for future research on seismic monitoring is to achieve accurate and efficient seismic signal capturing and events locating. To solve the inaccurate automatic picking up of low SNR seismic waves and avoid man-made interference during the events locating process, the authors proposed a transfer learning-based real-time automatic seismic locating method in mines using big data and deep learning. First, based on about one million seismic wave data of earthquakes, an initial model was designed to automatically pick up the arrival time of seismic waves in mines. Then, by using a set of more than 10,000 arrival times of seismic waves detected during mining, the initial model was further updated to fit the seismic location and information interpretation in mines, which achieves an automatic and accurate P wave pick-up for mining seismic waves. Finally, an automatic optimization method of locating sensors combination and a fine-tuning strategy for locating wave velocity was proposed, which can locate the seismic events automatically and accurately in mines. The proposed method was tested by using the blasting data from a longwall panel in an Inner Mongolia underground coal mine. The results show that the average planar and volumetric locating errors are 27.88 m and 28.40 m, respectively, which meets the accuracy requirements of seismic monitoring in mines. The method significantly reduces the impact of insufficient wave velocity accuracy of sensors on the locating events, and the results present strong robustness. Also, the seismic locating calculation time is shortened from minutes by manual work to within 200 milliseconds. The outcome can provide a reliable technical basis for real-time accurate seismic locating and intelligent hazard early warning in mines.