To address the limitations of conventional post-stack seismic attribute methods (such as coherence and curvature), including insufficient use of single attributes, reliance on manual interpretation, and difficulty in effectively integrating OBN multicomponent data, this study proposes an unsupervised fracture prediction method based on a multi-branch Variational Autoencoder (VAE) for OBN seismic data. The method fully leverages the multivariate attributes acquired by Ocean-Bottom Nodes (OBN), including PP-wave and PS-wave data as well as their corresponding coherence and curvature attributes, and constructs a multi-branch deep learning architecture for intelligent fracture system identification.A variational autoencoder framework is adopted, in which a four-branch encoder is designed to extract deep features from PP waves, PS waves, coherence, and curvature attributes, while multi-scale convolutional blocks and the CBAM attention mechanism are incorporated to enhance feature representation. Considering the geological continuity inherent in seismic data, a geology-aware smoothing loss is introduced and combined with the reconstruction loss and KL divergence, enabling data-driven unsupervised learning. A 128-dimensional latent space is used in the variational bottleneck, and fracture-related anomalies are identified through reconstruction-error analysis. Experimental results demonstrate that the proposed method effectively integrates multicomponent seismic data and seismic attribute information, achieving fracture prediction without the need for manual labels. Compared with traditional single-attribute approaches, the method successfully suppresses ring-shaped artifacts commonly observed in curvature attributes, reduces the excessive boundary responses in coherence attributes, and produces results with improved spatial continuity and geological plausibility, providing a new technical solution for fracture prediction under complex geological conditions.