LLM-Driven Observability for Private 5G Networks: A Modular Platform for Industrial Environments
Resumo
Private 5G networks in industrial environments demand low-latency, trustworthy observability of standardized Key Performance Indicators (KPIs), such as those defined in ETSI TS 128 554. However, existing literature rarely addresses the unified integration of on-premises telemetry pipelines with natural-language analytics and the systematic validation of LLM-generated SQL against complex observability schemas. This paper proposes an on-premises observability architecture that orchestrates the ingestion of ETSI/3GPP-aligned KPIs from an Open5GS/OpenAirInterface testbed into a time-series database via a scalable Prometheus–Telegraf–Kafka pipeline. Furthermore, we introduce the Onion Validation framework, a multi-layer inference protocol that enforces intent classification, schema grounding, syntactic correctness, execution-plan conformance, and result-level verification with full auditability. An experimental evaluation was conducted using five 8-bit quantized Large Language Models (2B–8B parameters) processed against 130 domain-specific queries, comparing unified versus partitioned execution architectures. The empirical results reveal a maximum functional correctness of 27.7% in Answerable queries, highlighting the strictness of the validation protocol. The findings demonstrate that while current edge-deployed quantized models exhibit reasoning limitations, the proposed layered validation mechanism is essential to mitigate operational risks, thereby delineating the critical trade-off between on-premises data sovereignty and the semantic accuracy of automated analytics.
Referências
Arya, M. and Simmhan, Y. (2025). Understanding the performance and power of llm inferencing on edge accelerators. arXiv e-prints.
Azariah, W., Bimo, F. A., Lin, C.-W., Cheng, R.-G., Nikaein, N., and Jana, R. (2024). A survey on open radio access networks: challenges, research directions, and open source approaches. Sensors, 24(3):1038.
Dandoush, A., Kumarskandpriya, V., Uddin, M., and Khalil, U. (2024). Large language models meet network slicing management and orchestration. arXiv preprint arXiv:2403.13721.
Eswaran, S. and Honnavalli, P. (2022). Private 5G networks: A survey on enabling technologies, deployment models, use cases, and research directions. Telecommunication Systems, 82(1):3–26.
ETSI (2021). Ts 128 554 v16.7.0 - 5g; management and orchestration; 5g end to end key performance indicators (kpi). Technical report, ETSI. [Online].
Husom, E. J., Goknil, A., et al. (2025). Sustainable llm inference for edge ai: evaluating quantized llms for energy efficiency, output accuracy, and inference latency. arXiv preprint arXiv:2504.03360.
Kan, K. B., Mun, H., Cao, G., and Lee, Y. (2024). Mobile-llama: instruction fine-tuning open-source llm for network analysis in 5g networks. IEEE Network, 38(5):76–83.
Mani, S. K., Zhou, Y., Hsieh, K., Segarra, S., Eberl, T., Azulai, E., Frizler, I., Chandra, R., and Kandula, S. (2023). Enhancing network management using code generated by large language models. In Proceedings of the 22nd ACM Workshop on Hot Topics in Networks, pages 196–204.
Mcnamara, J., Camps-Mur, D., Goodarzi, M., Frank, H., Chinchilla-Romero, L., Cañellas, F., Fernández-Fernández, A., and Yan, S. (2023). Nlp powered intent based network management for private 5g networks. IEEE Access, 11:36642–36657.
Nascimento, E. R. S., Garcia, G., et al. (2025). Llm-based text-to-sql for real-world databases. SN Computer Science, 6(2):130.
Ollama Inc. (2025). Ollama: local large language model runtime. [link].
Open5GS (2025). Open5gs: open source 5g core network. [link].
OpenAirInterface Software Alliance (2025). Openairinterface 5g wireless implementation. [link].
Panek, M., Jabłoński, I., and Woźniak, M. (2025). Automatic performance assessment: Step toward autonomous mobile network management systems. IEEE Communications Magazine, 63(4):73–79.
Saha, N., Shahriar, N., Boutaba, R., and Saleh, A. (2023). Monarch: network slice monitoring architecture for cloud native 5g deployments. In NOMS 2023 - 2023 IEEE/IFIP Network Operations and Management Symposium, pages 1–7.
Saha, S. and Viswanathan, H. (2022). Demonstrating network slice kpi monitoring in a 5g testbed. In NOMS 2022-2022 IEEE/IFIP Network Operations and Management Symposium. IEEE. Demo Track.
SQLAlchemy Authors (2025). Sqlalchemy: the database toolkit for python. [link].
The PostgreSQL Global Development Group (2025). Postgresql: the world’s most advanced open source relational database. [link].
Timescale Inc. (2025). Timescaledb: an open-source time-series database powered by postgresql. [link].
Xu, X., Chen, H., Simsarian, J. E., Ryf, R., Mazur, M., Dallachiesa, L., Fontaine, N. K., and Neilson, D. T. (2023). Demonstration of voice user interface for intelligent network orchestration. In Optical Fiber Communication Conference (OFC). Paper M3Z.11, Optica Publishing Group.
Yaqub, M. Z. and Alsabban, A. (2023). Industry-4.0-enabled digital transformation: prospects, instruments, challenges, and implications for business strategies. Sustainability, 15(11):8553.
Zahir, J. and Qadi, A. E. (2016). A recommendation system for execution plans using machine learning. Mathematical and Computational Applications, 21(2):23.
Zhan, Y., Cui, L., Weng, H., Wang, G., Tian, Y., Liu, B., Yang, Y., Yin, X., Xie, J., and Sun, Y. (2025). Towards database-free text-to-SQL evaluation: A graph-based metric for functional correctness. In Proceedings of the 31st International Conference on Computational Linguistics (COLING), pages 4586–4610.
