Sodium boiling Detection in a LMFBR Using Autoregressive Models and SVM

International audience This paper deals with acoustic detection of sodium boiling in a Liquid Metal Fast Breeder Reactor (LMFBR) cooled by liquid sodium. As sodium boiling induces acoustic emission, the method consists in real time analysis of acoustic signals measured through wave guides. AutoRegressive (AR) models are estimated on sliding windows and are classified in boiling or non-boiling models using Support Vector Machines (SVM). One of the difficulties to cope with is disturbances due to the influence of some environment noises like the liquid coolant cavitation, vortex flow, shaft vibration and mechanical pump noise. These disturbances can generate false alarms or mask the boiling. The proposed method is designed to be robust toward these disturbances. Furthermore, the SVM are designed to be robust toward the operating mode changing. The application for online monitoring is made on data obtained from French nuclear power plant Phenix and boiling sound signals generated from Laboratory experiments. Different acoustic boiling sound levels are used and the effectiveness of the method is shown by the good detection rate and its low false alarm rate even for low acoustic boiling sound level.