Two-step feature extraction of acoustic emission signals for leakage detection of valves in gas pipelines

Abstract

Evaluation of valve conditions is necessary to maintain integrity of pipelines and can be achieved using Acoustic Emission (AE) technique. Effective feature extraction from AE signals is critical to improving the accuracy of evaluation on valve conditions. In this study, a new method is proposed for processing AE signals to extract a two-dimension feature, which can characterize valve conditions more accurately. Time-frequency features of AE signal are extracted through Mel-spectrum analysis and then a deeper feature is extracted by a Generative Adversarial Network (GAN) model. Experiments were implemented considering three working conditions under three pressure levels in the pipeline. Results show that based on the extracted feature, leakage condition and non-leakage condition can be entirely differentiated and investigated leakage conditions can be differentiated with a high accuracy. By extracting the new effective feature, the proposed method provides a new way to effectively evaluate valve conditions.