Physics-guided, noise-resilient and interpretable machine learning framework for fault detection and diagnosis in variable refrigerant flow systems

Fault detection and diagnosis (FDD) in variable refrigerant flow (VRF) systems is typically driven by improving classification accuracy using complex machine learning (ML) models trained on clean datasets, with limited consideration of robustness to sensor noise, generalization under distributional shift, and interpretability. Moreover, existing studies predominantly focus on refrigerant charge and fouling faults, while critical component-level hard faults remain underexplored. To address these gaps, this study proposes a Gaussian noise-resilient, interpretable, and physics-guided ML framework for component-level VRF FDD using only built-in sensor telemetry. The framework introduces noise-resilient physics-informed (NRPI) variants of three conventional classifiers (RF, SVC, and XGBoost), integrating curriculum-based Gaussian noise augmentation to explicitly account for noise and a post-hoc physics-guided probabilistic rule to refine class-specific predictions in a modular manner. A comprehensive three-stage evaluation protocol including hold-out testing, independent offline validation, and controlled noise-based stress testing was employed to assess accuracy, offline generalization, and noise robustness. NRPI variants consistently outperform their baseline counterparts with NRPI-RF and NRPI-SVC achieving the highest mean macro-F1 of 99.10% and 99.13% across different level of sensor noise. Averaged across the different noise tests, NRPI‑XGBoost achieved the largest macro‑F1 gain (38.15%) over its baseline variant, while NRPI‑RF and NRPI‑SVC improved their macro-F1 scores by 11.98% and 4.38%, respectively. Ablation and interpretability analyses reveal that noise-aware training enhances global robustness under additive Gaussian noise, while the physics-guided rule improves class-specific reliability. Overall, the proposed framework provides an accurate, robust and interpretable pathway for component-level FDD in VRF systems.