Unsupervised learning-based acoustic detection of gas leakage in liquid: Evaluation of noise resistance based on parametric ROC analysis

Early detection of water/steam leakage is important in the prevention of failure propagation of heat transfer tubes in a steam generator of a sodium-cooled fast reactor. This study proposes an unsupervised learning-based acoustic method to detect gas leakage in liquid and evaluates its noise resistance based on parametric receiver operating characteristic (ROC) analysis. An autoencoder is trained, validated, and tested on time–frequency representations of simulated noise and leak signals for various signal-to-noise ratios (SNRs). To calculate a false positive rate and a true positive rate, the probability density function is assumed to be either as a normal distribution, a power transformed normal distribution, or a power normal distribution. As a result, the power normal distribution that shows the best goodness-of-fit was used as the probability density function to draw an ROC curve. The predictive ability of the autoencoder is evaluated as excellent for SNR = 0 , − 4 , − 8 , and − 12 dB, good for SNR = − 16 dB, and poor for SNR = − 20 dB. The autoencoder can detect leakage at relatively low-noise levels and has the potential to detect leakage at relatively high-noise levels equivalent to actual noise levels. Segmentation of the noise and leak signals can also be achieved from input, reconstructed, and residual images. These results suggest that the proposed method contributes to laying the foundation for detection and accident analysis of water/steam leakage in a steam generator of a sodium-cooled fast reactor.