Early Detection and Severity Estimation of Non-Condensable Gas Faults in Water-Cooled Chiller Using Physics-Informed Machine Learning Model

الملخص

Non-condensable gases Contamination (NC) in water-cooled chillers reduces thermodynamic efficiency, leads to higher power consumption for the compressor, and lowers system life span; however, such problems are challenging to identify in low levels of severity via traditional rule-based approaches. In this paper, an FDD algorithm of three-layer architecture integrating Gordon-Ng Thermodynamic Residuals, COP calculation, and RBF-SVM classification for identifying different NC concentrations at varying operating conditions. The method is validated on two independent datasets: a 58‑ton ASHRAE RP‑1043 laboratory chiller and a 338‑ton OpenStudio EnergyPlus simulation with synthetic NC faults injected via a custom OpenStudio measure. On the RP‑1043 data, the framework achieves a detection rate of 85.2% at 1% NC, rising to 100% at 5% NC, with a false alarm rate of 12.3% and SVM accuracy of 87.7%. On the 338‑ton simulation, detection reaches 99.6% at 2% NC and 100% at ≥3% NC, with 0.0% false alarms and 98.2% accuracy. The Gordon–Ng power residual on the larger chiller is nearly perfectly linear (~10 kW per 1% NC; R² ≈ 1.00), enabling continuous severity estimation from compressor power alone. Results show that physics‑informed residuals provide highly discriminative features for robust NC detection even at clinically low contamination levels, and that signal‑to‑noise ratio scaling with chiller capacity is a key calibration factor for field deployment. Future work will extend the pipeline to additional RP‑1043 faults, investigate multi‑fault concurrent detection, and pursue live BACnet/MQTT integration with commercial building management systems.