Innovative and Effective Prognostics and Health Management in Data Center Cooling Systems

SUMMARY & CONCLUSIONSCooling systems—particularly liquid cooling—are a critical mission segment in data centers. Any malfunction or operational failure in these systems can lead to substantial Data centers operational disruptions. To mitigate such risks, proactive Prognostics and Health Management (PHM) offers a promising framework for early anomaly detection and estimation of the system’s Remaining Useful Life (RUL). Various methodologies exist for collecting and analyzing health data to support PHM, ranging from sensor-based monitoring to advanced data analytics and machine learning techniques. Foundation models—large-scale, pre-trained neural networks developed on vast datasets —offer transformative potential for enabling data-driven PHM applications. These models deliver scalable and adaptable solutions that can be fine-tuned for specific tasks, accelerating innovation and significantly reducing the time and resources required to develop AI-powered systems. The model also performs well when health data from the system is limited, making it suitable for data-constrained environments.This study presents a novel approach to PHM in liquid-cooled data center environments by leveraging foundation models for time-series analysis. Liquid cooling is increasingly adopted in high-performance computing due to its superior thermal efficiency over traditional air-based cooling systems. However, the complexity of liquid cooling infrastructure and the cost of PHM implementation—particularly sensor deployment, data availability, data processing, and model training—pose significant challenges to broader adoption.To address the challenges of implementing PHM in liquid-cooled data centers, it is crucial to optimize sensor placement, standardize data formats for efficient storage and transmission, enhance data processing capabilities, and ensure customer data privacy [1]. This study focuses on the data processing aspect by proposing the use of foundation models as a modular feature extraction backbone for PHM tasks. These models excel at capturing rich temporal patterns from sensor data and exhibit strong generalization performance, even when trained on limited labeled or unlabeled datasets. Specifically, we employ Chronos-Bolt, a transformer-based foundation model for time-series forecasting, which is adapted to process continuous signals using quantization and tokenization techniques. This approach significantly reduces the need for manual feature engineering and large-scale domain-specific training.The degradation process and abnormal operation due to random failures in the data center cooling system is analyzed through a simulated model in Simulink. The proposed methodology is also validated using an HVAC dataset from the Lawrence Berkeley National Laboratory (LBNL). The current research is focused on the chiller component, which is central to energy consumption and thermal regulation. In our case study, the model successfully detected faults and classified different severity levels using a set of selected relevant input features.The dataset encompasses various failure types, including sensor bias, leakage, and fouling of cooling system components. In particular, several measurement biases are introduced to the temperature sensors on the chiller, and the corresponding operating conditions are simulated to reflect these anomalies. Experimental results demonstrate that the proposed method effectively detects sensor bias faults and accurately estimates fault intensity, with especially strong performance in identifying positive biases. Additionally, a feature impact analysis was conducted to evaluate the influence of selected input variables on model performance. The observed dependencies provide valuable insights for optimizing feature selection across different operational scenarios.In conclusion, this work demonstrates the practicality and effectiveness of using foundation models for PHM in liquid cooling systems, even under data-scarce conditions. The modularity, low data requirements, and adaptability of this approach make it well-suited for real-world deployment in data centers. Future work will explore broader fault types, richer sensor sets, and integration with edge computing to enable real-time health monitoring and control.