SAGE - scalable multi-agent architectures for facility management and energy efficiency

Starting point / motivation

Fault detection and diagnosis (FDD) for heating, ventilation, and air conditioning (HVAC) systems in buildings currently requires considerable human effort. Only after user complaints, sample analyses, or unspecific alarms from the building management system have occurred is data manually evaluated or inefficient corrective action is taken (e.g., increasing setpoints). 

Studies show that 15-30% of a building's energy consumption can be attributed to HVAC system faults or inefficient operation, affecting both existing and new buildings. Software solutions available on the market for optimizing operations often require highly qualified specialists, are very complex to set up, or require the existing measurement technology in the HVAC systems to be upgraded first. Fault diagnosis remains a labor-intensive, manual task for building operators.

Content and goals

In SAGE, we are working on an AI-based, self-learning, and self-diagnosing fault detection and diagnostics (FDD) solution for complex building technology systems. Through the use of Large Language Models (LLMs) and agent workflows, a structure is created that gives buildings an "awareness" of their internal and external status. This should make it possible to recognize operational anomalies, react dynamically to changing environmental conditions, and work in a more resource-saving and energy-efficient manner. 

A central aspect of the project is the development of a human-in-the-loop approach (HITL), which focuses on the interaction between man and machine. The methods developed are intended to support the work of building operators by providing intuitive communication interfaces such as chatbots and voice control. 

The LLM agents will act as intelligent assistants that present decision-making processes transparently, provide recommendations for action, and respond to requests in natural language (including translations in up to 20 industry-relevant languages).

Methods

In order to achieve the objectives of the project, SAGE is researching the solution of the following objectives using appropriately adapted methods: 

1. Develop a scalable solution for FDD in complex building energy systems.
   This solution is based on multi-agent architectures that can continuously collect and analyze building data and make adaptive decisions. In SAGE, we will implement and evaluate different agent workflow architectures to find the best one for the diagnostic domain. 
2. Develop robust machine learning methods for fault detection and diagnostics on both the generation and consumption side of a building.
   With the help of technologies, new reasoning structures can be developed that enable the system to continuously learn from new error cases and to adapt independently. 
3. Enable a comprehensive analysis of building data.
   To enable human-in-the-loop, SAGE uses multimodal models and the evaluation of different multimodal data sources (text, video, audio, ...) integrated in multi-agent architectures to enable a comprehensive analysis of building data. LLM agents are to be enabled by specific tools to better understand building data and use it more effectively.

Expected results

Based on the stated objectives and the corresponding methods, we expect the following results: 

• A scalable architecture for FDD in complex building energy systems that are based on open-source technologies and can be used freely as an open-source solution. In particular, this should also include an explanatory component that makes it possible to justify FDD results. 
• Proof of the usability of the human-in-the-loop approach through empirical validation.

Project management

Graz University of Technology, Institute of Software Engineering and Artificial Intelligence

Project or cooperation partners

• DILT Analytics FlexCo
• Vienna University of Technology, Institut für Hoch- und Industriebau

Prof. Dr. Franz Wotawa
Inffeldgasse 16b/2
A-8010 Graz

Tel.: +43 (316) 873 5724
E-mail: wotawa@tugraz.at
Web: www.tugraz.at/institute/sai/home