Development of an FDD model for an existing building using transfer learning

• Published in: HVAC&R research Vol. 30; no. 10; pp. 1183 - 1195
• Main Authors: Chu, Han-Gyeong, Cho, Seongkwon, Park, Cheol-Soo
• Format: Journal Article
• Language: English
• Published: Philadelphia Taylor & Francis 25-11-2024; Taylor & Francis Ltd
• Subjects: Air conditioning; Fault detection; Fault diagnosis; First principles; Learning; Synthetic data; Thermal comfort
• ISSN: 2374-4731; 2374-474X
• DOI: 10.1080/23744731.2024.2411160

Building heating, ventilation, and air conditioning (HVAC) systems are affected by several errors that can cause thermal discomfort to occupants and waste energy in buildings. Therefore, early and accurate Fault Detection and Diagnosis (FDD) is vital for enhancing the system's efficiency. Data-driven FDD is promising because it is convenient compared to the first-principles-based rule set that demands in-depth expertise. However, to realize data-driven FDD for real-life cases, the data-imbalance problem in FDD must be solved. In this study, the authors suggest a novel approach that generates synthetic data from a building system simulation tool, HVACsim+, using them as a source model to apply transfer learning to a target air handling unit system. Even though from the existing target system only the normal operational data were available, the transfer learning approach was satisfactory, confirming that the proposed method effectively mitigates data imbalance in developing data-driven FDD.