Data-Driven Detection of Leaking Valves in Air Handling Units

Maheen Abdul Ghani

Keywords: Valve leakage, HVAC systems, Fault detection, Machine learning, XGBoost, Predictive maintenance

Abstract:

Valve leakage in Air Handling Units (AHUs) silently degrades building energy efficiency and thermal comfort, yet it is hard to detect because compensating controls mask the fault. We study a data-driven approach that uses standard AHU telemetry to detect leakage robustly and in (near) real time.

We designed 80 controlled experiments on a full-scale laboratory AHU rig to simulate normal vs. leakage conditions under varying leakage rates, fan speeds, and water temperature profiles (1 Hz sampling; 20 min per run), yielding a labelled dataset of ~96k rows. Sensors included water/air temperatures, flow, and differential pressure (later excluded for deployability). We constructed two pipelines around gradient-boosted trees (XGBoost): (i) physically informed manual features (coil ΔT, rolling statistics, short lags), and (ii) automated feature generation (AutoFeat) to capture non-linear interactions. We also built a preprocessing bridge to align real operational logs to the lab schema to enable field validation despite limited metadata.

On unseen lab runs, the manual-feature XGBoost reached 97.41% accuracy (ROC-AUC 0.994); the AutoFeat variant reached 99.0% accuracy (ROC-AUC 0.998). To address the data drift between controlled lab data and real-world building data (domain shift), we injected 1,000 high-confidence field samples into training and retrained; the manual-feature model then achieved ~99% accuracy with ROC-AUC 0.9987 and tight cross-validation stability. Qualitative checks against coil temperature differentials (traditional leakage estimation method) corroborated predicted leakage patterns in the absence of ground-truth labels.

We implemented a lightweight inference stack—LabVIEW logging → Python watchdog → shared preprocessing/features → XGBoost inference → Streamlit dashboard—that produces operator-visible alerts with <2 send-to-end latency on standard hardware.

The study shows that: (1) a small, interpretable feature set tied to heat-exchange physics is sufficient for reliable leakage detection; (2) a lightweight adaptation step can bridge lab-trained models to real-world building data without privileged control signals; and (3) a deployable, vendor-agnostic pipeline can be realized with commodity sensors and open-source tooling. Unlike previous HVAC fault detection studies, this work specifically targets valve leakage detection, a fault type that has received little to no prior attention, providing a practical path toward predictive maintenance with measurable energy and sustainability impact, while highlighting the need for larger, domain-labelled field datasets to enable large-scale deployment.

Submission Number: 253