Abstract
•LLM-driven semantic feature selection for cloud-native 5G base-station KPIs.•A causal graph generation module based on score-based deep belief network for multidimensional KPIs.•State-space enhanced graph learner for long-range spatiotemporal fault diagnosis.•Superior diagnostic accuracy validated on China Mobile 5G base station dataset.
The cloudification of 5G base stations introduces a decoupled and multilayer architecture that significantly improves resource utilization, deployment flexibility, and operational efficiency, but also increases the complexity of fault diagnosis. To address this challenge, we propose CausaLM-Net, a unified diagnostic framework that integrates semantic reasoning with causal dependency modeling. First, a multi-large language models (LLMs) semantic feature selection module aligns key performance indicator (KPI) descriptions with operational logs to automatically identify fault-relevant indicators. Second, a score-based causal graph generation module learns the directional dependency structure among KPIs and constructs a sparse, interpretable causal graph. Finally, a selective state-space-enhanced graph learning module adaptively regulates node dynamics and edge dependencies, enabling the model to emphasize fault-critical causal pathways while suppressing noise. Experiments based on a real cloud native 5G base-station dataset from China Mobile demonstrate that the proposed method achieves strong robustness and clear performance advantages over mainstream fault diagnosis methods, offering a deployable, high-performing solution for intelligent fault diagnosis in cloud native 5G base stations.