Demonstration of an improved passive acoustic fault detection method on recordings from the Phénix steam generator operating at full power

Abstract : A hidden Markov model method proposed earlier for passive acoustic leak detection in sodium fast reactor systems has been improved to eliminate ad-hoc setting of model parameters and allow smaller amounts of training data. The method is based on training the hidden Markov model on known background noise only and optimizing its free model parameters by a parametric study of detection performance for synthetic noises superposed onto the same background noise. Using recordings of background noise as well as from argon injection tests performed at full power in the Phenix sodium fast reactor plant, it is estimated that the resulting method will detect leak-like deviations from the background noise with a detection delay of a few seconds, a false alarm rate close to 10$^{−8}$ and at signal-to-noise ratio conditions at least corresponding to an additive signal at $\sim$10 dB. The method is one-channel, i.e. using input from one single acoustic sensor only.