CNN-based acoustic identification of gas–liquid jet: Evaluation of noise resistance and visual explanation using Grad-CAM

Abstract

For the analysis of anomalies in a steam generator (SG) of a sodium-cooled fast reactor (SFR), we evaluate the noise resistance of CNN-based acoustic identification methods of gas–liquid two-phase jets and produce visual explanations for their decisions. First, we introduce the water flow sound and the three types of gas–liquid jet sounds, which simulate the background noise and the anomaly sounds, respectively. Second, we produce time–frequency representations for various signal-to-noise ratios (SNRs) and employ AlexNet, VGG16, and ResNet18 to the identification of the gas–liquid two-phase jets. As a result, the best CNN of ResNet18 achieves more than 92% for SNR=0,−4,−8,and−12 dB and 69% for SNR=−16and−20 dB. This result indicates that our proposed methods can identify the flow states of gas–liquid two-phase jets in low-level noise environments and detect the gas–liquid two-phase jets even in high-level noise environments. Also, Grad-CAM suggests that ResNet18 focuses on one of the spectrum peaks of the water flow sound and all or part of the signal intensity pattern of the gas–liquid jet sounds. Our proposed methods lead to the safe operation and fast, accurate, and accountable analysis of anomalies in SFR.