The rapid development of blockchain technology has laid the foundation for the widespread application of smart contracts. However, the immutability and code complexity of smart contracts also pose significant security challenges. Traditional vulnerability detection methods struggle to address emerging attacks due to limitations such as high computational overhead, poor generalization, and insufficient dynamic behavior analysis. To tackle these challenges, we propose a tree-based convolutional neural network framework (TVD-NN). This method compiles smart contracts into abstract syntax trees (ASTs), extracts structural and semantic features through tree-structured intermediate representations and keyword embedding techniques, and employs an improved tree-based convolutional neural network for vulnerability classification. Experimental results indicate TVD-NN outperforms traditional symbolic execution tools, providing scalable theoretical frameworks and practical tools for smart contract security.