Under the background of the deep integration of artificial intelligence into the protection, dissemination and reuse of cultural heritage, it is urgent to move from empirical inheritance to data-based, intelligent and interactive expression of traditional carving skills. This paper constructs a multi-modal data system that integrates images, videos, 3D point clouds, text and oral data. Combined with deep visual coding, cross-modal knowledge representation, knowledge graph constraints, conditional generative pattern reconstruction, style transfer and intelligent interactive feedback mechanism, this paper proposes a technical path for traditional sculpture maintenance and transformation. The experimental results show that the accuracy of the proposed method in carving category recognition and knife discrimination reaches 90.6% and 88.9%, respectively. In the generation task, SSIM, PSNR, style consistency and semantic integrity reach 0.893, 30.4 dB, 0.901 and 0.918, respectively. The system also improves the scores of pattern recognition, knife understanding and style judgment by 12.8, 13.2 and 12.9 points respectively, and maintains the recognition accuracy of 80.9% under the condition of 40% noise or occlusion, which reflects good cross-scene adaptation ability and robustness. It provides more operable technical support for the digital activation, accurate inheritance and innovative transformation of traditional carving skills
