Acoustic monitoring of sodium boiling in a Liquid Metal Fast Breeder Reactor from autoregressive models

Issa Cherif Geraldo (1) , Tanmoy Bose (2) , Komi Midzodzi Pekpe (1) , Jean-Philippe Cassar (1) , Amiya Rajan Mohanty (2) , Kévin Paumel (3)

1 CRIStAL - Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189
2 IIT Kharagpur - Indian Institute of Technology Kharagpur
3 CEA Cadarache

This paper deals with acoustic monitoring of sodium boiling in a Liquid Metal Fast Breeder Reactor (LMFBR) based on Auto Regressive (AR) models which have low computational complexities. Some authors have used AR models for sodium boiling or sodium-water reaction detection. These works are based on the characterization of the difference between fault free condition and current functioning of the system. However, even in absence of faults, it is possible to observe a change in the AR models due to the change of operating mode of the LMFBR. This sets up the delicate problem of how to distinguish a change in operating mode in absence of faults and a change due to presence of faults. In this paper we propose a new approach for boiling detection based on the estimation of AR models on sliding windows. Afterwards, classification of the models into boiling or non-boiling models is made by comparing their coefficients by two statistical methods, multiple linear regression (LR) and Support Vectors Machines (SVM). The proposed approach takes into account operating mode informations in order to avoid false alarms. Experimental data include non-boiling background noise data collected from Phenix power plant (France) and provided by the CEA (Commissariat à l'Energie Atomique et aux énergies alternatives, France) and boiling condition data generated in laboratory. High boiling detection rates as well as low false alarms rates obtained on these experimental data show that the proposed method is efficient for boiling detection. Most importantly, it shows that the boiling phenomenon introduces a disturbance into the AR models that can be clearly detected.

Mots clés: Fault detection; Sound analysis; Nuclear plant; Autoregressive models; Classification methods

Domaines: Acoustique [physics.class-ph]; Traitement du signal et de l'image [eess.SP]; Machine Learning [stat.ML]

Issa Cherif Geraldo, Tanmoy Bose, Komi Midzodzi Pekpe, Jean-Philippe Cassar, Amiya Rajan Mohanty, et al.. Acoustic monitoring of sodium boiling in a Liquid Metal Fast Breeder Reactor from autoregressive models. Nuclear Engineering and Design, 2014, 278, pp.573-585. ⟨10.1016/j.nucengdes.2014.07.026⟩. ⟨hal-01298037⟩