Abstract: With the new requirements of Industry 5.0 for knowledge-driven intelligent manufacturing,the field of mechanical assembly is faced with challenges such as sparse multi-modal data,entity semantic boundary blurring and long tail effect of data distribution.Therefore,a BERT-Tiny Transformer-CRF model with polynomial loss function was proposed to improve the efficiency of domain knowledge extraction in low resource scenarios.The domain prior knowledge was injected through knowledge distillation and semantic enhancement technology.A dimension adaptive feature compression module was designed to realize cross-modal feature fusion.Finally,a dynamic edge-aware decoding mechanism was constructed to achieve accurate positioning of entity boundaries.A comparative study with different entity recognition models was carried out on the self-built automated assembly data set.The results demonstrated that the proposed model had good adaptability in the entity recognition task of the automatic assembly field.The model was superior to other comparison models with 86.62% accuracy,85.27% accuracy,85.67% recall and 85.46% F1 value,which provided an effective technical method for the construction of knowledge graph in the field of mechanical automated assembly under Industrial 5.0.

Key words: BERT-Tiny Transformer-CRF model, data enhancement, PolyLoss, automated assembly

摘要： 随着工业5.0对知识驱动的智能制造提出新要求,机械装配领域面临多模态数据稀疏、实体语义边界模糊、数据分布呈现长尾效应的挑战。为此,提出一种融合多项式损失函数的BERT-Tiny Transformer-CRF模型,旨在提升低资源场景下的领域知识抽取效率。首先,通过知识蒸馏与语义增强技术注入领域先验知识,其次设计维度自适应特征压缩模块实现跨模态特征融合,最后构建动态边缘感知解码机制实现实体边界的精准定位。利用自主构建的自动化装配数据集,将所提方法与不同实体识别模型进行对比,实验结果表明,所提模型具有良好的泛化识别能力,以86.62%的准确率、85.27%的精确率、85.67%的召回率和85.46%的F1值优于其他模型,为工业5.0下机械自动化装配领域知识图谱的构建提供了一种有效的技术方法。

关键词: BERT-Tiny Transformer-CRF模型, 数据增强, PolyLoss, 自动化装配