Abstract
Fault detection in Air Handling Units (AHUs) is critical for improving energy efficiency and maintaining indoor environmental quality in modern buildings. However, the practical deployment of data-driven fault detection remains challenging due to the scarcity of labeled fault data, the need for expert annotation, and the presence of noisy, incomplete, and low-resolution building management system (BMS) data. This study aims to evaluate the effectiveness of multiple unsupervised learning methods for fault detection in fully unlabeled industrial AHU data. To address these challenges, this study investigates robust unsupervised methods for fault detection using a real industrial AHU dataset that contains missing values, measurement noise, outliers, and low-resolution sampling. A comparative framework is developed to evaluate multiple unsupervised learning approaches, including Principal Component Analysis (PCA), autoencoders, K-Means clustering, and statistical control charts, without relying on labeled data. The results show that autoencoder-based models achieve the best overall performance by effectively capturing complex nonlinear relationships, while PCA and K-Means identify more pronounced patterns, and control charts provide simple but limited univariate insights. These findings highlight the importance of robust data preprocessing and demonstrate that unsupervised methods can provide practical and scalable solutions for fault detection in real-world Heating, Ventilation, and Air Conditioning (HVAC) systems under realistic data constraints.