Passive acoustic leak detection for sodium cooled fast reactors using hidden Markov models

Abstract

Acoustic leak detection for steam generators of sodium fast reactors have been an active research topic since the early 1970s and several methods have been tested over the years. Inspired by its success in the field of automatic speech recognition, we here apply hidden Markov models (HMM) in combination with Gaussian mixture models (GMM) to the problem. To achieve this, we propose a new feature calculation scheme, based on the temporal evolution of the power spectral density (PSD) of the signal. Using acoustic signals recorded during steam/water injection experiments done at the Indira Gandhi Centre for Atomic Research (IGCAR), the proposed method is tested. We perform parametric studies on the HMM+GMM model size and demonstrate that the proposed method a) performs well without a priori knowledge of injection noise, b) can incorporate several noise models and c) has an output distribution that simplifies false alarm rate control.

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

“…For our application, i.e. fault detection in industrial acoustic noise signals, a new feature calculation scheme based on the Welch PSD estimate [7] was presented in [8] along with some basic HMM theory. A brief review of this basic theory and the new feature calculation scheme explicitly stating the improvements that are new to this work will now be given.…”

“…Contrary to the normalization used in [8], this residual will not be indifferent to pure multiplicative changes in the signal (i.e. X(f j , m) → kX(f j , m)) since X 0 is constant.…”

“…Based on XR(f j , m) and J k (m) we now introduce a modified feature calculation scheme compared to the one of [8], given by…”

“…The relative importance factor of the first and second term of the feature, β, was in our previous work set by an ad-hoc trial and error procedure [8]. Since we are aiming for a method to be licensed for safety related systems in nuclear power plants, basing the method on well-defined and predictable procedures is of utmost importance.…”

“…In this work we develop such a method, based on a hidden Markov model (HMM) concept proposed in [8]. The performance is assessed by applying it to recordings from argon injection tests performed in the Phénix steam generator at full power.…”