Microseismic monitoring in mining areas is one of the key technical approaches to enhance mining safety and issue timely hazard early warnings. With the rapid growth of the seismic monitoring data from mining areas, the traditional manual seismic phase identification and catalog compilation methods can barely meet the demands of time-sensitive, high-precision and continuous monitoring.Our study area is the Daqiang Coal Mine, which sits in Zhangqiang Town, Kangping County, Shenyang, Liaoning. Since 2022, Liaoning Earthquake Agency and the mine’s company have been working together on building a dedicated seismic network, tweaking the station layout, and doing detailed analyses of microseismic activity. To speed up catalog construction and get more reliable locations, we used continuous recordings from 64 sets of short-period dense array deployed at the Daqiang Mining Area in Liaoning Province as the research dataset, covering the period from July 10 to August 11, 2023. We ran these data through an automated workflow that uses AI for phase picking and location, and then we looked at the locations, waveform features, and station residuals together. For phase picking, we used the DiTingPicker deep learning model to pick P-wave and S-wave seismic phases automatically, and we combined REAL for association, HypoInverse for absolute locations, and HypoDD for double-difference relative relocation to build our catalog. For most stations, the P-wave arrivals on the vertical component are fairly clear, and the P-wave picks are generally consistent, which suggests that the deep learning model has good detection capability for this type of microseismic event. The S-wave amplitudes, however, are weak and the onsets are not obvious, so the S-wave picks come with higher uncertainty than the P picks.

The results show that the workflow picked out and located 255 microseismic events. The processing time dropped from about a month with manual work to roughly two days, which is about 15 times more efficient. The catalog’s completeness magnitude is around M_L1.0, and the b-value is about 0.90. In terms of space, the events are mostly clustered around the mining face and near known faults, which matches pretty well with the shallow disturbance from mining. More detailed examination reveals that both S-wave picking quality and timing errors from the short-period instruments can affect the locations of shallow events, sometimes shifting the depths systematically. A few stations showed systematic mismatches between the theoretical arrival times and the real waveform features, suggesting possible overall timing offsets. But these anomalies are not necessarily caused by the phase-picking model itself; they could also be related to station timing errors, channel delays, time calibration issues in the data headers, or the operating status of the instruments. In this study, we tested how well deep learning works for detecting microseismic events and picking phases in a mining setting, built a catalog for the Daqiang mining area, and looked at its spatial temporal distribution and magnitude-frequency patterns. We also discuss how S-wave arrival quality, phase weighting choices, and timing errors affect the location results for shallow events. Our findings should be useful for automated monitoring, catalog quality control, and identifying dynamic hazard risks in mines. Looking at the locations overall, the events tend to cluster on the map mainly within and around the mining area, showing pretty clear spatial clustering that seems to relate to both mining disturbances and local structures. Events are mostly located near the mine boundaries and within the array coverage, suggesting that the dense array has good monitoring capability for microseismicity in this area. In terms of magnitude distribution, small events dominate, while relatively few larger ones occur – this matches the general pattern of mining-induced microseismicity. AI-based methods have significantly improved the efficiency of event detection and phase picking, providing important technical support for rapid catalog construction. However, in the complex observational environment of a mine, automated results still need to be combined with waveform verification, station residual analysis, and location sensitivity testing. Especially for shallow microseismic events, weak S-wave development and station timing errors can have a notable effect on the estimated focal depths.

In summary, AI-based methods can significantly enhance the automation and efficiency of microseismic catalog construction in mining areas. In practice, however, particular attention should be paid to the timing accuracy of short-period instruments. DiTingPicker demonstrates good applicability to the Daqiang dataset, performing well on the dense array data. The observed spatial clustering, together with the magnitude-frequency characteristics （M_L 1.0, b≈0.90）, suggests that microseismicity is associated with both mining disturbances and local structural conditions. Shallow events tend to be scattered, likely related to near-surface activities, while deeper events exhibit linear trends that may reflect buried faults or stress concentration boundaries. Timing errors and unstable S-wave picking are identified as two key factors affecting the reliability of shallow event locations in this study.