This study proposed a fracture prediction method based on unsupervised clustering of fracture sensitivity attributes to accurately characterize the distribution of fractures with different attitudes in tight sandstones of the Upper Triassic Xujiahe formation in the Yuanba area, northeastern Sichuan. First, the sensitivity of fractures was extracted and selected based on the optimized post-stack seismic data. Then, the convolutional neural network, a deep learning algorithm, was used to learn global massive fault and fracture databases of various types, obtaining the intensities, dip angles, and azimuths of fractures. In combination with high-precision-guided curvature attributes, an unsupervised clustering algorithm based on the Bayesian probability model was used to predict the intensities of fractures with different dip angles through dimensionality reduction using principal component analysis (PCA). The prediction results are highly consistent with both the fracture interpretation results from imaging logs and the geological results. The results of this study show that the third member of the Xujiahe Formation has more developed fractures than the second member. Fractures in the third member include both the fault-induced fractures distributed near the faults in the southeast flank of the Jiulongshan anticline and the fold-induced fractures in the areas with large formation flexures. By contrast, only fault-induced fractures near the faults occur in the second member.