Advancing Building Autonomy with LLM-Based Fault Detection and Preventive Maintenance

Modern building infrastructures are increasingly complex and heavily instrumented with sensors, yet fault detection and preventive maintenance remain challenging. Existing approaches primarily rely on expert-crafted rules or specialized machine learning models, both of which require extensive labeled data and often fail to generalize across diverse building configurations and equipment types. In this paper, we propose a novel framework that leverages Large Language Models (LLMs) for fault detection and predictive maintenance in building systems. Our method employs in-context learning to enable LLMs to synthesize heterogeneous data sources, including sensor logs, maintenance histories, and operational metadata, thereby generating contextualized diagnostics and actionable maintenance recommendations. We evaluate the framework on a simulated multi-system dataset encompassing HVAC, electrical, and elevator subsystems. Experimental results demonstrate that our LLM-based approach achieves superior fault classification accuracy compared to conventional anomaly detection baselines, while also producing human-interpretable diagnostic reports. These findings highlight the potential of LLM-powered frameworks to transform building management by providing a scalable, data-driven, and adaptive consultation tool for infrastructure maintenance.