Causal reasoning subgraph retrieval for coal mine safety risk assessment

Abstract

Coal mine safety risk assessment can reduce accidents, but traditional methods rely on subjective expert judgment or isolated sensor monitoring, which cannot assess complex and interrelated diverse security risks. This paper proposes a retrieval-augmented generation (RAG) framework based on causal reasoning knowledge subgraphs. We first construct a coal mine safety knowledge graph that integrates risk factors, solutions, and legal bases by employing a Bidirectional Encoder Representations from Transformers with whole word masking combined with Bi-directional Long Short-Term Memory and Conditional Random Field (BERT‑wwm‑BiLSTM‑CRF) for entity recognition and BERT-wwm-BiLSTM-Attention for relation extraction. Then we retrieve a causal reasoning subgraph from this graph through relation expansion, node screening, and semantic pruning. This subgraph is provided to the large language model (LLM) to make risk assessment. Experimental results show an accuracy of 95.9%.