Microseismic signal recognition based on power spectral density differences and fracture network evolution in coal seam hydraulic fracturing

To address the low accuracy of microseismic signal identification and the difficulty in reliably characterizing fracture network evolution under strong noise conditions, field microseismic monitoring data from hydraulic fracturing at the Yushutian coal mine were used to investigate signal identification methods and fracture network spatial evolution in coal seams. By revealing the intrinsic differences between microseismic events and typical noise signals in terms of dominant frequency energy concentration, spectral flatness, and spectral oscillation index, a dual-branch signal identification framework was developed by integrating multi-scale time–frequency deep convolutional features with power spectral density (PSD) statistical features. The results show that the dual-branch model outperforms the single PSD-based statistical feature model in terms of accuracy, recall, and F1-score, with the average accuracy increasing from 0.83 to 0.93, recall from 0.79 to 0.92, and F1-score from 0.80 to 0.92. After model validation, the framework was applied to the full-stage hydraulic fracturing microseismic dataset at the Yushutian site, and a total of 581 microseismic events were identified from continuous raw waveforms, providing a reliable dataset for subsequent characterization of fracture spatial structures. For hypocenter determination, the rapid earthquake association and location (REAL) method was employed, in which event associations were established using the number of P/S-wave picks and arrival-time residuals, and the results were further refined using a high-precision relative location method. The introduction of the REAL method improves the location efficiency of the microseismic sequence and constrains the three-dimensional distribution of fracturing events. The hypocenter locations and fracture network distribution indicate that the scale and geometric complexity of fracture networks vary among different fracturing stages; the fracture network is predominantly oriented in the NW–SE direction (approximately 320°–340°), with secondary branches in the NE–SW direction (approximately 40°–70°), reflecting the combined control of regional structural joints and the in-situ stress field. In terms of stimulated reservoir volume (SRV) quantification, the total SRV of all fracturing stages is 318.1×10⁴ m³. Among them, Stage 8 exhibits the largest SRV (45.2×10⁴ m³), followed by Stages 6 and 7, whereas Stages 1 and 9 show relatively low SRV values; the spatial density of microseismic events is generally consistent with the SRV distribution.