Multi-class gas pipeline leak detection using Hilbert transform and a lightweight convolutional neural network

Abstract

Ensuring the safety and sustainability of gas distribution networks depends heavily on the timely detection of pipeline leaks, since leaks that go unnoticed can cause substantial losses and create serious safety risks. When applied to multi-class leak types, however, conventional detection approaches often prove insufficiently accurate. To address this limitation, the present study proposes a novel multi-class gas pipeline leak detection approach based on acoustic signal analysis, aiming to enhance both detection accuracy and processing efficiency. In this approach, features are first extracted from leak-generated acoustic signals through application of the Hilbert transform, after which classification is performed using a lightweight convolutional neural network (CNN) architecture referred to as LightCNN. The method was tested using the gas pipeline leakage assessment-12 (GPLA-12) dataset, comprising 12 distinct leak categories. According to the evaluation outcomes, the proposed approach peaked at 96.18% accuracy while employing a more streamlined model architecture than earlier techniques. Overall, this work contributes meaningfully to advancing leak detection systems that are more reliable, faster, and more efficient, with potential for broad application across the gas pipeline industry and in signal processing technologies more generally.