Smart contract security serves as a critical safeguard mechanism for ensuring trusted transactional interactions within blockchain ecosystems, emphasizing the imperative for systematic vulnerability detection processes. Deep learning-based techniques have emerged as a promising solution in blockchain security, through their capacity to automate the extraction of high-level semantic features. However, two challenges persist in current methodologies for smart contract vulnerability detection: (1) Feature representation mechanisms exhibit inherent limitations in comprehensively and precisely capturing the intrinsic characteristics of vulnerabilities; (2) Existing frameworks have not explored methods for feature enhancement during training. To overcome the aforementioned challenges, we propose a novel framework integrating feature enhancement techniques with self-supervised training for robust smart contract vulnerability detection. (1) Contract feature vectors are extracted using pre-trained models and convolutional neural networks, followed by retrieval of semantically similar contracts from existing repositories and generation of chain-of-thought vectors through large language models; (2) Contractual features and reasoning patterns are fused to create enhanced representations; (3) Knowledge distillation is employed to optimize feature learning during model training. Empirical findings from three standardized data corroborate that the developed method exhibits enhanced predictive accuracy relative to conventional benchmarks in the field.