Détection de fuites de liquide frigorigène dans les installations frigorifiques industrielles à compression de vapeur

Efficient detection of refrigerant leakage is crucial for industrial refrigeration systems, as leaks can significantly impact both system performance and environmental sustainability.Existing fault detection and diagnosis (FDD) methods rely primarily on experimental or laboratory data.However, in the industrial use case, achieving accurate early detection poses significant challenges.Refrigerant leakage can be identified by tracking unexpected drops in the liquid level within the system receiver.This PhD thesis presents a novel data-driven and knowledge-guided approach for refrigerant leak detection in industrial vapor compression refrigeration systems.Our method predicts the liquid level in the receiver under fault-free operating conditions and identifies leaks by comparing the actual and predicted values, incorporating both operational and external factors.The experiments were conducted using real-world data from industrial refrigeration systems. We evaluated five machine learning models for their ability to predict fault-free liquid levels and explored various fault detection techniques for leak identification, along with preliminary research on leak diagnosis.To ensure that the model remains robust over time despite the evolving nature of refrigeration systems, we developed an automated approach for concept drift detection and adaptation.Domain knowledge played a key role in guiding every stage of our approach. By integrating real-world data with knowledge-driven enhancements, our method shows strong potential for a reliable and scalable system.Ultimately, we aim to develop a fully automated leak detection system that minimizes human intervention, adapts to system changes, and scales across different installations.