The Construction of Knowledge Graphs in the Assembly Domain Based on Deep Learning

The operation scenario of human-robot collaboration assembly involves multiple channels to acquire assembly domain knowledge data. In order to equip collaborative robots with certain general knowledge of assembly work and logical reasoning ability, this paper proposes a Chinese-character-based names entity recognition model Bert-BiLSTM-CRF and a relational extraction model Bert-BiGRU-ATT, based on the multiple heterogeneous data in the assembly domain, to construct an assembly domain knowledge graph centered on <operator> and <assembly object> for human-robot collaboration. Among them, the Chinese-character-based names entity recognition model avoids the error bias caused by Chinese word segmentation in traditional methods, while the relation extraction model fully extracts the association information between entities and relations. The proposed entity recognition model and relation extraction model are validated in a real scenario of mechanical product assembly. The experimental results show that the proposed models exhibit excellent performance in the human-machine collaborative assembly domain, with overall average F1-score of 84.02% and 94.92% for entities and relations, respectively. The constructed knowledge graph of the assembly domain contains 2724 triples, and the knowledge graph is presented through a visual interactive interface.