Air Handling Units (AHUs) are critical for maintaining indoor air quality and energy efficiency in buildings. A common but often undetected fault in AHUs is valve leakage, which can lead to significant energy losses, reduced thermal comfort, and increased operational costs. This thesis investigates a datadriven approach to detecting valve leakage using machine learning, with the goal of developing models that can generalize from controlled experimental conditions to real-world installations.
The work involved designing and executing 80 controlled experiments on a laboratory AHU test rig, simulating both normal and leakage conditions using temperature, flow, and pressure sensors. The collected dataset was used to train and evaluate several machine learning algorithms, with a focus on XGBoost models using both manually engineered and AutoFeat-generated features. A dedicated preprocessing pipeline was created to adapt and align field AHU operational data with the lab dataset, enabling real-world validation despite limited metadata.
Results show that the XGBoost models could identify leakage patterns with high accuracy (97.41%) in lab conditions and retain detection capability when tested on real-life field data, even with only partial operational context. Cross-verification with coil temperature differentials supported the validity of the predictions. The findings demonstrate that machine learning can detect subtle operational anomalies indicative of leakage, offering a foundation for predictive maintenance in HVAC systems. However, larger volumes of domain-labeled operational data are needed before large-scale deployment.