Enhancing Fault Detection in CO2 Refrigeration Systems: Optimal Sensor Selection and Robustness Analysis Using Tree-Based Machine Learning

Abstract

This study investigates the reliability and robustness of data-driven Fault Detection and Diagnosis (FDD) models for CO2 refrigeration systems (CO2-RS) in supermarkets, focusing on optimal sensor selection and resilience against sensor noise. Using tree-based machine learning algorithms - Random Forest (RF), XGBoost, CatBoost, and LightGBM - we developed FDD models to classify six common faults in a laboratory-scale CO2-RS. The Recursive Feature Addition (RFA) approach identified optimal sensor sets, achieving a 99% F1-score with minimal sensors: four for RF, seven for XGBoost, five for CatBoost, and five for LightGBM. Condenser-side sensors consistently ranked as critical for fault detection. Robustness was assessed by injecting Additive White Gaussian Noise (AWGN) at a signal-to-noise ratio (SNR) of 3 dB into the most important sensor, with XGBoost showing superior resilience at 85.24%, followed by CatBoost (57.07%), LightGBM (49.1%), and RF (49.46%). Sensitivity analysis across high-SNR (10 dB), low-SNR (0 dB), and sensor failure scenarios revealed XGBoost's robustness peaking at 90.23% and retaining 79% under failure, contrasting with sharper declines in other models. These findings highlight a trade-off between sensor count, cost, and reliability, with larger ensembles enhancing noise resilience. This work bridges gaps in FDD literature by integrating sensor optimization with comprehensive robustness analysis, offering a practical framework for improving energy efficiency and fault management in CO2-RS. Future efforts could explore adaptive SNR thresholds and redundant sensor configurations to enhance real-world applicability.