This study addresses the core challenges of insufficient feature extraction capability and semantic disentanglement difficulties in traditional methods for unsupervised learning of vector graphics. We propose an innovative solution based on BetaVAE, systematically exploring the modeling capabilities of the variational autoencoder framework for latent features in vector data. Through the integration of theoretical analysis and experimental validation, the research reveals the unique advantages of BetaVAE in balancing reconstruction accuracy and feature disentanglement, while establishing a parameter optimization mechanism tailored to the topological characteristics of vector graphics. The results show that, in both feature representation capability and cross-domain adaptability, the proposed method substantially outperforms conventional unsupervised learning approaches in learning hierarchical semantic features of vector drawings. The main contribution of this work is the methodical application of BetaVAE to vector graphics unsupervised learning tasks, so extending the application limits of variational autoencoders and offers a fresh theoretical framework and technical path for processing challenging vector data. For promoting algorithmic innovation in intelligent design and allied disciplines, these developments have great worth.