?
Refrigerant Leak Detection in Data Centers Using Topologically Determined Graph Neural Networks
This paper investigates the problem of detecting slow refrigerant leaks in a data center cooling system using a graph neural network. The study addresses the challenge of early fault identification, proposing a method for constructing a topological graph based on the engineering diagram, the physical layout, and the cause-and-effect relationships in the cooling system. This graph structure effectively captures the spatial and functional dependencies between system components. Comparative testing of the GConvGRU model with topological, fully connected and correlation graphs, as well as the classic LSTM, was conducted on a real dataset from an industrial container-based data center. The experiments showed that the topological graph approach demonstrates superiority in all metrics: accuracy, F1, and detection time. Furthermore, the model proves effective even with limited labeled anomaly data, highlighting its robustness and practical applicability for real-world monitoring systems. The results confirm that incorporating domain knowledge of the system’s physics can significantly improve the quality of slow anomaly detection, reducing time to detection while minimizing false positives.
