Acoustic data condensation to enhance pipeline leak detection

• Published in: Nuclear engineering and design Vol. 327; pp. 198 - 211
• Main Authors: Oh, Se Won, Yoon, Doo-Byung, Kim, Gwan Joong, Bae, Ji-Hoon, Kim, Hyeon Soo
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.02.2018; Elsevier BV
• Subjects: Acoustic noise; Acoustic signal; Artificial intelligence; Autocovariance; Damage prevention; Data condensation; Data processing; Leak detection; Learning algorithms; Machine learning; Monitoring; Nuclear energy; Nuclear power plants; Remote sensors; Sensors; Steam electric power generation; Structural damage; Acoustic signal; Data condensation; Leak detection; Autocovariance; Machine learning
• ISSN: 0029-5493; 1872-759X
• DOI: 10.1016/j.nucengdes.2017.12.006

•For pipeline leak detection air-borne acoustic signals are remotely captured.•A method to find the key predictors that affect leak distinction is presented.•Data conversion of the acoustic signals facilitates distinguishing pipeline leaks.•Performance metrics are derived by several learning algorithms for classification.•Experiments show leaks can be better distinguished using less amount of predictors. Acoustic monitoring techniques are widely adopted for identifying various leaks from plant facilities to prevent loss of resources and any further structural damages. As the conventional sensing devices have measured acoustic signals at predesignated positions inside or very close to the object being observed, the need for more sophisticated and automated monitoring of more complex infrastructure has increased both the number of sensors to be installed and the amount of data to be analyzed. Thus, in order to diagnose the high-pressure steam leakage efficiently, this research proposes a novel method to find and condense the distinguishable features from the acoustic signals, which are captured by remotely dispersed microphone sensor nodes around a laboratory scale nuclear power plant coolant system. The performance of the proposed method is evaluated by several quantitative metrics resulting from the five state-of-the-art machine learning algorithms, together with the condensed data ratio. Experimental results show that the proposed method can transform the original acoustic signals into a smaller number of featured predictors, even less than ten-thousandths of the original data amount, while improving classification accuracy despite loud machine-driven noises nearby.