Multi-granularity Hierarchical RAG for Welding Parameter Recommendation

With Industry 4.0 driving intelligent transformation in manufacturing, optimizing welding parameters for industrial robots has become critical to achieving high-quality, cost-effective production. Traditional methods relying on expert experience and trial-and-error approaches suffer from inefficiency and poor knowledge reuse. While Retrieval-Augmented Generation (RAG) systems enhanced by Knowledge Graphs (KGs) and Large Language Models (LLMs) offer promise, they struggle with two key challenges: neglecting the hierarchical structure of welding knowledge and failing to capture complex interdependencies among parameters. To address these limitations, this paper proposes a multi-granularity hierarchical RAG framework for welding parameter recommendation. By decomposing the task into three reasoning layers, the framework integrates a progressive knowledge flow mechanism that combines relational vector matching, case knowledge graph querying and process manual subgraph retrieval. Additionally, we construct Wecommend, a dataset encompassing 1009 sets of on-site welding production records and 177 standardized parameters extracted from welding process manuals. Experimental results demonstrate that our multi-granularity hierarchical RAG approach effectively captures the hierarchical dependencies among welding parameters through multi-granularity knowledge enhancement, achieving a recommendation accuracy of 89.86% and outperforming baseline methods.