Presentation Information
[3D05]Acoustic anomaly detection of gas leakage in liquid based on unsupervised learning
*Nao Mikami1, Kosuke Aizawa1, Akikazu Kurihara1, Yoshitaka Ueki2 (1. JAEA, 2. TUS)
Keywords:
Gas-liquid two-phase flow,Acoustic technique,Machine learning,Unsupervised learning
In a sodium-cooled fast reactor (SFR), the early-stage detection of water leakage from a steam generator tube is important to prevent the tube failure propagation caused by the sodium-water reaction. Separating noise and leak sounds is a key issue to an acoustic technique with a short-time response. Recent studies have applied machine learning to address this issue. In particular, unsupervised learning is expected to be useful for anomaly detection in SFR. The present study focused on this potential and validated the basic feasibility of an unsupervised learning-based acoustic technique using an autoencoder. The autoencoder learned features of time-frequency representations of the simulated noise and leak sounds. Receiver operating characteristic (ROC) curves were estimated and areas under ROC curves were evaluated.