Predictive Maintenance of Hydro Turbine-Generator Units: A Review

Abstract

Reliable operation of hydro turbine-generator units (HTGUs) is central to safe, flexible, and efficient hydropower generation. State-of-the-art approaches to condition monitoring, fault diagnosis, and early fault detection increasingly rely on artificial intelligence (AI)-driven methods. However, labeled fault data remain scarce and industrial validation of these AI methods remains limited. This paper presents a systematic, algorithm-focused review of predictive maintenance (PdM) for HTGUs. Using a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-guided literature search, we map data-driven and AI-enabled methods from classical machine learning through deep Autoencoders and hybrids to Transformers and state-space models (SSMs), and we compare them with respect to their interpretability, data requirements, and industrial deployability. Relative to plant-wide sensing surveys and generator-centric AI reviews, the contribution is an HTGU-wide assessment of which method families are validated on operational plants versus rotating-machinery benchmarks. Across the reviewed studies, several general patterns emerge: hydropower-specific work is dominated by unsupervised anomaly detection, classical machine learning approaches remain a strong, plant-validated baseline, and high accuracies of recent Transformer/SSM architectures should be interpreted as architectural potential on related assets, not as demonstrated HTGU deployability. We translate the identified gaps into a short-/medium-/long-term roadmap toward transferable, explainable, and industrially validated PdM.