A Physics-Informed Modeling of Data Center Cooling and Self-Supervised Learning-Based Autoencoder for Fault Detection and Diagnosis

Effective Fault Detection and Diagnosis (FDD) is crucial for data center energy efficiency but often struggles with labeled data scarcity. We propose the Physics-Informed Spatio-Temporal Graph (PI-STG) framework, a two-stage self-supervised approach. Utilizing the AlphaDataCenterCooling virtual testbed, we generate high-fidelity data via signal-level fault injection. Stage 1 employs an auto-encoder for robust unsupervised anomaly detection. Stage 2 utilizes a graph neural network, explicitly modeled on the hydraulic cooling topology, to diagnose root causes by tracing fault propagation. Experiments demonstrate that PI-STG achieves an F1-score of 0.931, outperforming purely data-driven baselines like LSTM-AE and GAT, particularly in diagnosing complex actuator failures. Furthermore, the framework reduces energy waste from undetected faults by over 15%, bridging the gap between theoretical deep learning and practical facility management.