In order to solve the problem that the traditional causal inference method is difficult to effectively model the nonlinear dependence between multi-dimensional indicators in root cause analysis, and the interpretability is low, a fusion of graph neural network (GNN) and large language model (LLM) causal traceability framework are proposed. Complex equipment and systems are modeled as peer-to-peer abstract networks, and a multi-level explanation system from system level to intervention level is constructed through the causal analysis strategy of single node and full graph level, combined with the structural modeling ability of GNN and the knowledge reasoning ability of LLM. The results of simulation network datasets show that the proposed method achieves competitive prediction performance and stronger explanatory power without prior knowledge, in which the full graph causal traceability provides a global relationship perspective while maintaining a high accuracy (MSE = 0.042 8), and the model combined with the true prior is optimal in the error index. The results show that this method not only expands the application boundary of large model in the root cause analysis of complex systems, but also provides an extensible and interpretable technical path for intelligent operation and maintenance and anomaly diagnosis.