Automatic pipeline leak detection based on AST deep learning using a free-swimming detector with acoustic resonance cavity

Abstract The free-swimming in-pipe spherical detector (SD) has found practical application in pipeline leak detection due to its proximity to the leak points, low risk of obstruction, and low cost. Acoustic resonance air cavity (ARAC) design enables the SD to capture small leak acoustic signals with high sensitivity, but it simultaneously renders the SD highly vulnerable to collision noise interference, increasing the difficulty of leak identification. For long-time-series acoustic data acquired by the SD, traditional manual analysis methods are inefficient and difficult to ensure accuracy. Additionally, valid leak samples of field pipelines are insufficient, and the actual leak events represent an extremely low proportion, which makes the collection result in a small sample imbalance dataset. To address these issues, this paper proposes an automatic identification method for pipeline leak sounds collected by the SD via incorporating Audio Spectrogram Transformer (AST) deep learning. First, a pipeline leak voice dataset is constructed using the field pipeline acoustic signals captured by the SD, which contained three types of acoustic events: normal rolling, collision, and leak. Second, the problems of category imbalance and sample insufficiency are addressed by combining data augmentation and transfer learning. Finally, the AST model is applied to pipeline leak sound identification and leak point localization. Experiments under different pressure conditions show that this method can accurately identify and locate 1mm aperture leaks, and the localization errors under 1MPa and 0.5MPa conditions are 2.46m and 3.07m, with relative errors of 0.37% and 0.47%, respectively. This research provides a new solution for the automation and intelligence of pipeline leak detection and localization, with good engineering application prospects.