Physics-informed hybrid anomaly detection and fault diagnosis for an R744 booster refrigeration system with heat recovery and oil management

Abstract

Purpose. To develop and evaluate a physics-informed hybrid anomaly detection and fault diagnosis framework for an R744 booster refrigeration system with heat recovery and oil management by combining thermodynamic modelling and machine learning methods. Methodology. The proposed approach integrates thermodynamic feature engineering with two unsupervised machine learning techniques: K-means clustering and an autoencoder neural network. K-means clustering is used to identify global operating-mode deviations, while the autoencoder detects local nonlinear anomalies through reconstruction-error analysis. A hybrid decision logic based on union and intersection rules is applied to improve diagnostic reliability. The framework was validated using a synthetic physics-informed dataset representing seven days of operation with a 10-second sampling interval and including normal operation as well as four fault classes: flash-gas instability, heat-recovery disturbance, oil-return disturbance, and sensor drift. Findings. The developed hybrid framework demonstrated high anomaly-detection and fault-diagnosis performance. The Stage-1 anomaly detector achieved Precision = 1.000, Recall = 0.903, and F1-score = 0.949. The Stage-2 diagnostic classifier reached an overall accuracy of 0.992. The results confirmed the capability of the proposed approach to distinguish between different fault mechanisms and associate detected anomalies with physically meaningful thermodynamic processes occurring within the refrigeration system. Originality. A physics-informed hybrid fault detection and diagnosis framework for transcritical R744 booster refrigeration systems with heat recovery and oil management is proposed. Unlike conventional purely statistical approaches, the developed methodology combines thermodynamic interpretation with machine learning techniques, enabling both accurate anomaly detection and physically explainable diagnostics. Practical value. The proposed framework can be integrated into SCADA and building management systems as a supervisory diagnostic layer for commercial refrigeration applications. The approach supports predictive maintenance, improves system reliability, reduces energy consumption, and facilitates the implementation of intelligent monitoring strategies for low-GWP refrigeration systems.