Leak detection in a gas pipeline using spectral portrait of acoustic emission signals

This paper proposes a reliable method to detect leaks and recognize their various sizes in a gas pipeline based on the spectral portrait of acoustic emission (AE) signals. Through analyzing the relation of leak syndromes and AE signals in the frequency domain, this paper demonstrates that leak identification only by signatures obtained from measured individual signals might be unreliable due to strong attenuation in wave propagation along pipelines. The proposed method constructs a vector quantity in the frequency domain from two AE signals simultaneously acquired by two sensor channels, combined with a wave propagation model. This quantity characterizes the spectral portrait of AE signals emitted from a leak, its signatures therefore reflect leakage syndromes the same as leak signals, whereas those of individual sensor-measurements involve considerable distortions. Hence, the accuracy and reliability of leak detection and recognition are improved by the approach using the spectral portrait-related quantity. This paper also describes bank filters designed to encode the spectral envelope of AE signals instead of observing their total continuous frequency span. Finally, a trained multi-class support vector machine (SVM) classifier is responsible for the leak categorization, exploiting a set of the most discriminative encoded spectral envelope (ESE) features. The experimental results show that the proposed method outperforms conventional algorithms in pattern separability quantified by Kullback-Leibler distance and classification accuracy. (C) 2019 Elsevier Ltd. All rights reserved.