This study designs and optimizes a deep learning-based algorithm for 3D model defect detection tasks to improve detection accuracy and efficiency. Considering the full mining of features of spatial information and topological structure inherent in 3D data, a hybrid model is established by combining the respective characteristics of CNNs and GCNs. In this respect, algorithm optimization includes a weighted cross-entropy loss function to handle problems brought about by class imbalance, a cosine annealing learning rate scheduling strategy to speed up model convergence, and the introduction of Dropout and L2 regularization to avoid overfitting. Experimental results demonstrate that the optimized model outperforms the baseline model and the enhanced model in terms of accuracy, recall rate, and F1-score, reaching 90.2%, 88.0%, and 89.0%, respectively, reflecting the effectiveness and potential of the proposed algorithm in practical industrial applications. The stability and generalization ability of the algorithm are guaranteed through systematic model training and parameter tuning, and its superior performance in the detection of complex defect categories is verified.