Data center cooling system fault detection and diagnosis: A comprehensive review and outlook

Data center industry has become a rapidly growing electricity consumer, as cooling systems can account for 24-40% of total data center energy use. Improving cooling performance is therefore essential not only for thermal reliability and operational continuity, but also for reducing energy consumption and associated indirect carbon emissions, thereby supporting low-carbon operation and renewable-friendly demand response. Compared with conventional buildings, data center cooling systems exhibit stronger spatiotemporal coupling, multi-scale fault propagation, and stringent thermal constraints, which limits the direct transferability of traditional fault detection and diagnosis (FDD) methods. Despite the growing importance of these challenges, existing FDD reviews have mainly addressed conventional building HVAC systems or general diagnostic algorithms, leaving data center cooling FDD insufficiently synthesized as a distinct research problem. This review addresses this gap by synthesizing FDD methods, applications, and quantitative performance evidence in data center cooling systems. It compares knowledge-driven and data-driven approaches, including hybrid deep learning methods across five dimensions: modeling dependency, data requirement, interpretability, computational complexity, and energy/carbon impact measurability. Commonly used evaluation methods and metrics for fault detection and diagnosis are also summarized. Finally, five critical challenges are identified, and future directions are outlined toward knowledge-enhanced, multimodal, and self-evolving FDD systems that can support sustainable, reliable, and autonomous operation of next-generation intelligent data centers.