Sustainable and Intelligent Public Facility Failure Management System Based on Large Language Models

Siguo Bi, Jilong Zhang, Wei NiarXiv 2025 · DOI · arXiv

Abstract

Modern public facilities increasingly rely on a diverse array of smart devices, creating significant management challenges due to data heterogeneity and the limitations of reactive maintenance. This paper proposes a novel framework utilizing Large Language Models (LLMs) to move toward proactive, semantic-aware facility management. By fusing multimodal data—such as error logs and surveillance imagery—with localized knowledge extracted from technical manuals, the system enables sophisticated root cause analysis and risk assessment while maintaining data privacy through local deployment.

Introduction

Public facilities like libraries, museums, and gymnasiums are now ecosystems of interconnected smart devices, ranging from self-service kiosks to complex climate control systems. Managing this "zoo" of infrastructure is difficult because devices generate disparate data types (structured logs, text, audio, and video) that traditional systems cannot easily correlate. Furthermore, existing maintenance strategies are largely reactive, addressing failures only after they occur.

Overview of the diverse smart systems and public facilities managed by the proposed framework.

Methodology

The proposed LLM-based Smart Device Failure Management framework is structured into three distinct layers designed to handle data complexity and ensure privacy.

The three-layer architecture of the LLM-based Smart Device Failure Management framework.
  1. Data Layer: This layer ingests heterogeneous data sources, including structured operational logs, raw text, and visual data from surveillance systems.
  2. Localized LLM Layer: Utilizing a locally deployed Llama 3.2-Vision model, this layer performs semantic understanding fusion. It converts multimodal inputs into a unified semantic representation (embeddings), allowing the model to reason across different data types.
  3. Knowledge Localization: To prevent "hallucinations" and provide expert-level guidance, the system extracts troubleshooting protocols directly from product manuals. This creates a local knowledge base that grounds the LLM's responses in manufacturer-specific facts.

Experimental Validation

The framework was validated through a "Scene Simulation" within a library environment. The study utilized:

  • Simulated Log Data: Operational logs for self-service machines containing timestamps and specific error codes (e.g., "Network exception").
  • Synthetic Visual Evidence: Multi-angle surveillance footage simulating various device states.
  • Technical Documentation: Product manuals used to populate the local knowledge database for precise troubleshooting.

Key Results and Insights

The prototype verification demonstrated several critical capabilities:

  • Semantic Fusion: The system successfully correlated timestamped log entries with visual cues. For instance, it could link a "Network exception" log to a visual identification of a disconnected Ethernet cable in a surveillance frame.
  • Root Cause Analysis: In test scenarios, the LLM combined disparate data points to conclude that failures were caused by specific physical issues, such as accidental disconnections during device transport.
  • Automated Reporting: The framework generated comprehensive "Device Anomaly Reports" in natural language, summarizing the fault, evidence, and potential remedies for facility managers.

Conclusion and Future Work

By deploying LLMs locally, this framework provides a privacy-preserving, intelligent solution for public infrastructure management. It effectively bridges the gap between raw device data and actionable maintenance insights. Future research will focus on scaling the system to noisier real-world environments and integrating IoT security features to protect against malicious cyber-physical attacks.

Summary updated

BibTeX
@misc{bi2025sustainable,
  author = {Siguo Bi and Jilong Zhang and Wei Ni},
  title = {{Sustainable and Intelligent Public Facility Failure Management System Based on Large Language Models}},
  year = {2025},
  eprint = {2501.06231},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.06231},
}
Figure