Scholarlink

<p>To address the issues where existing machining feature recognition methods overlook engineering geometric constraints and rely heavily on large-scale labeled data, this paper proposes a geometry-knowledge enhanced multi-relation graph contrastive learning framework tailored for data-scarce scenarios. First, we construct a Multi-Relation Attributed Graph (MRAG). Beyond traditional topological adjacency, this graph explicitly introduces three geometric relations: normal parallelism with area similarity, axis-plane parallelism, and orthogonality, to encode engineering constraints in a heterogeneous graph structure. Second, we design a Semantic Fusion Heterogeneous Graph Neural Network (SF-HGNN), which integrates local topological information with global geometric semantics through relation-differentiated message passing. Finally, the model undergoes pre-training on open-source unlabeled datasets to learn robust geometric representations from unlabeled data, followed by training on labeled datasets. Experimental results demonstrate that the proposed method outperforms existing models in machining feature recognition tasks on small-sample datasets, exhibiting significant robustness under few-shot conditions and reducing dependence on data annotation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MRAG: multi-relation attributed graph contrastive learning for few-shot machining feature recognition

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Ziyan Du,
  • Minrui Wang,
  • Ruizhe Wang,
  • Yibing Peng

摘要

To address the issues where existing machining feature recognition methods overlook engineering geometric constraints and rely heavily on large-scale labeled data, this paper proposes a geometry-knowledge enhanced multi-relation graph contrastive learning framework tailored for data-scarce scenarios. First, we construct a Multi-Relation Attributed Graph (MRAG). Beyond traditional topological adjacency, this graph explicitly introduces three geometric relations: normal parallelism with area similarity, axis-plane parallelism, and orthogonality, to encode engineering constraints in a heterogeneous graph structure. Second, we design a Semantic Fusion Heterogeneous Graph Neural Network (SF-HGNN), which integrates local topological information with global geometric semantics through relation-differentiated message passing. Finally, the model undergoes pre-training on open-source unlabeled datasets to learn robust geometric representations from unlabeled data, followed by training on labeled datasets. Experimental results demonstrate that the proposed method outperforms existing models in machining feature recognition tasks on small-sample datasets, exhibiting significant robustness under few-shot conditions and reducing dependence on data annotation.