Abstract: To address the issues of untimely or excessive maintenance caused by the reliance on empirical judgment for punch wear in stamping production, this paper proposes an intelligent punch maintenance method based on punching defect features and graph convolutional networks (GCN). Firstly, punching images were acquired via a vision inspection system and underwent image preprocessing. Subsequently, techniques such as least squares circle fitting and convex hull fitting were employed to quantitatively extract multi-dimensional geometric defect features of the punched holes. To effectively utilize the complex correlations among features, a feature graph structure was constructed based on the Pearson correlation coefficient, and a GCN model was utilized to learn the deep non-linear mapping relationships between defect features. Experimental results showed that the proposed GCN model achieved a diagnostic accuracy of 0.991 and an F₁-score of 0.987 on the test set, outperforming baseline models such as MLP, SVM, and 1D-CNN. Finally, combining actual production line data, a predictive maintenance strategy based on the proportion of punching anomalies was formulated. Field verification showed that the predictive maintenance strategy based on the proportion of abnormal punched holes could reduce service-life waste caused by premature replacement and lower the batch quality risk associated with continued production using severely worn punches, providing a feasible technical approach for intelligent maintenance of stamping production lines.