关键词: U-Net网络, 断层识别, 高角度走滑断裂, 伴生小断距次级断裂, 正态标准化, 聚焦均方损失函数, 白碱滩组, 三叠系, 玛湖凹陷

Abstract: The small faults of Triassic Baijiantan Formation in Mazhong area of Mahu Sag in Junggar Basin were identified by using the improved U-Net network small fault identification technology. The results show that: （1）Construct-guided filtering preprocessing can effectively improve the quality of seismic data and increase the accuracy of fault identification. The U-Net network, which incorporatesskip connections, relay supervision, normal standardization, and focused mean square loss function, has improved its fine identification ability of small faults（. 2）Using 200 training sample sets and 20 validation sample sets, the seismic data of the model were generated by convolution of reflection coefficients and Ricker wavelet, and faults were manually labeled. Optimal network model parameters were selected to test on the synthesized noisy seismic data by using coherent attributes, conventional U-Net network methods, and the improved U-Net network method. Construct-guided filtering effectively highlighted the boundaries of faults and enhanced the lateral continuity of the same phase axis. The improved U-Net network method can effectively identify faults with a fault distance of more than 7 meters.（3）The improved U-Net network method has significantly higher accuracy than coherent attributes and conventional UNet network methods in identifying high-angle strike-slip faults and associated secondary faults with small fault throw of Triassic Baijiantan Formation in Mazhong area of Mahu Sag. The No. 3 and No. 4 sand bodies in the northern Dazhuluogou fault in the study area are favorable areas for efficient exploration of Triassic Baijiantan Formation in MZ4 well area.

Key words: U-Net network, fault identification, high-angle strike-slip fault, associated secondary faults with small fault throw, normal standardization, focused mean square loss function, Baijiantan Formation, Triassic, Mahu Sag