Labeled Datasets for Air Handling Units Operating in Faulted and Fault-free States

Data-driven fault detection and diagnosis (FDD) for buildings' heating, ventilating, and air conditioning (HVAC) systems has gained popularity in recent years. However, the scarcity of well-labeled data that represents true fault symptoms presents a challenge for developing new FDD methods. Furthermore, there is growing interest in applying transfer learning (TL) for building applications, where well-labeled data from one building is used to diagnose faults in a related but different building. Successful evaluation of TL algorithms requires at least two datasets that share similarities yet exhibit differences in some operational conditions. Unfortunately, the lack of comparative studies to identify suitable dataset pairs has slowed the progress of TL or other inter-dataset studies. To address these challenges, this paper focuses on the air handling unit (AHU), a key HVAC subsystem, and 1) presents the publication of eight new datasets, operating under fault-free and various faulty conditions; and 2) conducts a comprehensive study on AHU fault datasets to identify dataset pairs and their associated faults that are most suitable for evaluating TL algorithms.