Fault Diagnosis of Air Handling Units in an Auditorium Using Real Operational Labeled Data across Different Operation Modes

Abstract

Fault detection and diagnosis (FDD) in air handling units (AHUs) is essential for ensuring indoor air quality and prolonging system life span. However, the use of real operational data for FDD has been limited in existing research, largely due to challenges in data availability and the complexity of fault annotation. This study specifically focuses on AHUs with constant air volume (CAV) systems, a critical component of HVAC operations within an auditorium. Data from 15 different sensors installed across 13 AHUs in the auditorium were collected and analyzed. Seven operational conditions were identified - including normal operation and six fault types - and were annotated accordingly. Nine supervised machine learning methods were applied to FDD tasks involving these operational conditions. Three types of models were developed: an integrated model trained on both cooling and heating data sets, a cooling model trained exclusively on cooling data, and a heating model trained exclusively on heating data. Each method underwent hyperparameter tuning across 10 configurations, resulting in a total of 270 models (9 methods × 10 configurations × 3 model types). These models were validated to identify the best-performing model in each category based on F1-scores. The integrated model achieved average F1-scores of 95.43% for the cooling season and 95.65% for the heating season. The cooling model achieved F1-score of 97.82%, and the heating model reached 96.87% in their respective seasons. Based on these findings, the potential applications of these models for real operational conditions are discussed in this research.