Leak detection in pipelines based on acoustic emission and growing neural gas network utilizing unlabeled healthy condition data

Abstract

Leak detection in pipelines is crucial to prevent economic losses, guarantee safety and reduce undesirable environmental impacts. Hence, it becomes necessary to detect leaks as early as possible. Recent studies reported several leak detection techniques that primarily analyze Acoustic emission. (AE) data through artificial intelligence (AI) algorithms for investigating pipeline leaks. As prior labeled leakage data is typically required for training these AI models, their usage in industrial settings is impracticable. Moreover, existing works suggest various AE-based features that respond differently to premature leaks, which calls for development of comprehensive leak indicators to arrive at clear-cut decisions. This paper proposes a novel approach based on AE data and growing neural gas (GNG) network to build health indicators for real-time tracking of pipeline condition and sensing incipient leaks. GNG models the data topology in an unsupervised manner without making any prior assumptions about their structure and distribution. First, the AE signals are acquired from the test pipeline and the time- and frequency-domain features are extracted from AE signals. Then, the GNG is trained with feature data belonging to pipeline healthy condition only. In the end, the desired health indicator is obtained by supplying incoming feature samples to the already trained GNG model. The proposed approach is validated on the pipeline data collected from both laboratory experimental setup and actual field. The results indicate that the developed pipeline health indicator is capable of recognizing very small-size leaks and predicting their severity as leak-size cultivates. Moreover, higher average leak detection accuracy and leak sensitivity for different pressure conditions, i.e. 100 % and 96.5 %, are obtained. The proposed GNG-based-leak diagnosis framework outperforms the existing techniques such as Gaussian mixture models (GMM) and Support vector data description (SVDD) by improving the leak detection accuracy and leak sensitivity via 24.1 % and 5.4 %, respectively.