Early warning in Molten Salt Fast Reactors based on a data-driven method for the online incident detection and diagnosis

• The study presents a data-driven method for online incident detection in MSFR operation. • The methodology monitors key plant parameters to identify abnormal operating conditions. • The algorithm is trained on simulated deviations of main MSFR operating variables. • The model achieves 89–99 % accuracy in detecting and classifying plant anomalies. • The framework performs probabilistic inference to identify likely root-cause scenarios. This paper presents an innovative online incident detection and classification method, which aims at improving the safety, reliability and availability of Molten Salt Fast Reactor (MSFR) power plant, focusing on scenarios characterized by deviations from normal operational conditions. The first part of the paper is devoted to describing and discussing the proposed online data-driven incident detection and classification methodology (based on adaptive Singular Value Decomposition-SVD and kNN algorithm), which aims at identifying abnormal plant conditions thanks to a continuous monitoring of some measurable parameters and variables (e.g., the molten salt temperatures in the secondary circuit). The developed incident detection algorithm is trained on a set of simulated scenarios featured by deviations of the main MSFR plant parameters from their nominal values. The data-driven model is then assessed considering increasingly complex incident classification rules and tasks, showing satisfactory performances in detecting and classifying plant anomalies (with an accuracy ranging between 89 % and 99 %). Finally, a fault diagnosis framework is proposed to carry out probabilistic inference on the most likely root causes (or precursors) – e.g., combinations of physical parameter values and component failures – that lead the system to the detected abnormal states.