Robustness of Episodic Acoustic Leak Detection Across Noise Levels

Abstract

Acoustic leak detection is critical for pipeline integrity, yet it faces challenges from environmental noise and limited labeled data. To our knowledge, no prior work has systematically quantified noise robustness in episodic acoustic leak detection. We tested a Conformer-based prototypical network on the GPLA-12 dataset over an SNR 0-20dB range with five different random seeds. The results for clean, 10dB, and 0dB conditions were 65.88% ±1.36%, 68.04% ±0.66%, and 63.07% ±1.30%, respectively. The addition of 10dB SNR enhanced performance by 2.16% (p < 0.05), which was linked to a regularization effect observed in the training-validation curves, while 0dB SNR reduced accuracy by 2.81% (p < 0.05). While fully-supervised methods achieve higher absolute accuracy, our 5-shot episodic result of 65.88% provides the first baseline for few-shot leak detection on this dataset, demonstrating feasibility under extreme data scarcity.