Multi-Objective Optimized Machine Learning for Early Fault Detection in Water-Cooled Chillers
In large office buildings and commercial complexes, HVAC systems account for nearly two-thirds of total electricity consumption. However, early fault detection and diagnosis (FDD) in water-cooled chillers remains challenging because faults usually develop slowly, produce weak initial signatures, and exhibit strongly nonlinear thermodynamic behavior. Moreover, overlapping operational characteristics between faults make accurate diagnosis under mild and moderate conditions difficult. Previous studies using the ASHRAE RP-1043 dataset reported limited diagnostic performance for incipient faults, particularly refrigerant leakage and condenser fouling. To address these gaps, this study proposes a hybrid optimization-based FDD framework for early fault diagnosis in water-cooled chillers, integrating the Non-Dominated Sorting Genetic Algorithm III with Local Search (NSGA-III-LS) and the M5 Prime regression model for hyperparameter tuning and nonlinear operational modeling. The proposed framework offers a balanced trade-off between fault sensitivity, residual stability, and diagnostic accuracy. Validation results demonstrate high detection rates of 62.5-95.83% for mild faults and nearly 100% for severe faults, outperforming conventional methods, especially during incipient fault stages. The proposed method also supports earlier detection of abnormal thermal behavior, contributing to energy savings of approximately 15-30%, extended equipment lifespan, and more effective predictive maintenance planning.