Multi-stage faults are developed in the Shanxi Formation-Taiyuan Formation, the main strata in the northwestern margin of the South North China Basin. In this area, the heterogeneity is strong, the fracture formation mechanism is complex, and the small-scale hidden faults are relatively developed. The small-scale fracture boundary characteristics obtained by conventional prediction methods are fuzzy and the accuracy is low, which seriously restricts the development process of deep coal measure gas. Therefore, it is urgent to find a concealed structure prediction method suitable for the study area. Taking JF1 well area as an example, this paper proposes a method of hidden fracture identification based on texture analysis of gray level co-occurrence matrix and Mini Batch K-means deep clustering. Firstly, this paper uses time-varying frequency division deconvolution technology to carry out frequency expansion processing, and obtains broadband post-stack seismic data. Then, by optimizing the scale and gray level of the three-dimensional sliding window, the gray level co-occurrence matrix is generated in the sliding window according to the specified true dip angle and azimuth angle, and the texture features of entropy, difference, uniformity and energy are extracted respectively. The texture attributes based on true dip angle and azimuth angle constraints are used as sample data. The initial clustering center of Mini Batch K-means is set, and the small batch data subset is optimized to establish a Mini Batch K-means deep learning model suitable for JF1 well area. Finally, based on intelligent ant colony algorithm optimization.