Abstract Galvanised Steel Pipe (GSP) is the most common gas pipeline in populated areas. Existing leak detection research aimed at welded steel pipe is not suitable for GSP system due to their differences in line pressure, connection method, and leak path. This paper presents a gas leak detection method for galvanised steel pipe based on acoustic emission. An experimental setup composed of eight segments is designed to quantitatively simulate gas leak in GSP network considering flow-induced noise. The experiments verify that internal flow noise demonstrates similarity to leak-induced signals and thus interferes with leak detection based on shallow machine learning approaches. Convolutional Neural Network (CNN) is therefore introduced to solve the problem. Different network architectures are investigated and evaluated. Two types of inputs are discussed, namely time-domain signal and time-frequency distribution. Leak detection result show that the proposed method is robust to internal flow noise. When leak rate is greater than 0.03 L/s, the best model achieves overall accuracy more than 93% in both the test set and the cross-validation set. The model performances indicate that traditional frequency analysis is ineffective to improve the flow-noise robustness of the CNN-based leak detector.

中文翻译：

---

基于卷积神经网络的带内部流动噪声的镀锌钢管气体泄漏检测

摘要 镀锌钢管（GSP）是人口稠密地区最常见的输气管道。现有针对焊接钢管的检漏研究，由于管线压力、连接方式和泄漏路径的差异，不适用于GSP系统。提出一种基于声发射的镀锌钢管气体泄漏检测方法。一个由八段组成的实验装置旨在定量模拟 GSP 网络中考虑到流动引起的噪声的气体泄漏。实验验证了内部流动噪声与泄漏引起的信号相似，因此会干扰基于浅层机器学习方法的泄漏检测。因此引入了卷积神经网络（CNN）来解决这个问题。研究和评估了不同的网络架构。讨论了两种类型的输入，即时域信号和时频分布。泄漏检测结果表明所提出的方法对内部流动噪声具有鲁棒性。当泄漏率大于 0.03 L/s 时，最佳模型在测试集和交叉验证集上的总体准确率均超过 93%。模型性能表明，传统的频率分析对于提高基于 CNN 的检漏仪的流动噪声鲁棒性是无效的。