A Multimodal Knowledge Graph Completion Method for Human–Robot Collaborative Maintenance of Industrial Robots Toward Industry 5.0

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Industrial robot maintenance knowledge graphs often contain missing relationships, while semantic differences among maintenance text, graph structure, and operational condition signals complicate multimodal link prediction. Sparsely connected entities further limit the available structural evidence. To address these challenges, this study proposes a prototype-enhanced multimodal knowledge contrastive learning (Pro-MKCL) framework. The framework combines modality-specific encoders, multi-head neighborhood attention, and a cross-modal Transformer to learn integrated entity representations. Ontology-constrained multi-prototype contrastive learning further promotes within-category compactness and between-category separation. Experiments were conducted on an in-house industrial robot maintenance multimodal knowledge graph and two public benchmarks, MKG-W and DB15K. Across five independent runs, Pro-MKCL achieved a mean MRR of 0.9193 and Hits@1 of 0.8789 on the industrial robot dataset, compared with 0.9122 and 0.8709 for MOUNT, the strongest multimodal baseline. Its MRR and Hits@1 were 0.3529 and 0.2960 on MKG-W, and 0.3658 and 0.2777 on DB15K, respectively. On the industrial robot dataset, MRR remained at 0.8905 when 40% of the multimodal attributes were randomly masked. These findings support improved link prediction across the evaluated datasets and robustness under the tested masking condition, providing a knowledge foundation for fault association and collaborative maintenance decision support.