Deep learning-based health monitoring and prognostics for nuclear power plants: a systematic review

To ensure the safe and economical operation of nuclear power plants, reliable and accurate anomaly detection and operational trend prediction for key performance parameters are critical. Deep learning has been increasingly adopted for health monitoring and prognostics, yet focused systematic reviews of recent deep learning-based studies remain limited. This review synthesizes 99 studies on deep learning-enabled health monitoring and prognostics for nuclear power plants from a performance-centric, multi-level perspective. It summarizes the major technical challenges, classifies research subjects across system, component/equipment, and parameter levels, and analyzes the main data sources. It further examines state-of-the-art model architectures and end-to-end health monitoring and prognostics pipeline design choices, including data preprocessing, multi-source data fusion, health indicator construction, threshold setting, and evaluation metrics. In addition, the review highlights key technical trends, persistent deployment barriers, and practical design trade-offs relevant to stringent nuclear applications. Based on these findings, it further articulates a performance-centric, multi-level HM&P framework as a forward-looking synthesis to support future research on deep learning-enabled nuclear power plant health management.