Abstract In this work, a novel convolutional neural architecture (SE-CNN), which combines Spectrum Enhancement (SE) and Convolutional Neural Network (CNN), is proposed to detect the leak of gas pipeline. The SE has the effect of enhancing the leak signals and reducing background noise. CNN can automatically extract leak features and realize leak diagnosis. The experimental results show that the SE-CNN can achieve an average accuracy of 94.3 % for 6 categories and only requires 1.04 seconds of detection time. In this experiment, the diameters of the main pipeline and the branch pipeline are 125 mm and 25 mm. Due to its excellent accuracy and efficiency, the proposed enhanced spectrum convolutional neural architecture paves the way for real-time leak detection in industrial environments, which can ensure the process safety of gas pipeline transportation. Under strong background noise, the average accuracy of the SE-CNN can reach 94.3 % , which is 33 % , 3.7 % higher than that of SVM and CNN. In particular, the SE can be regarded as a data compression method, which can significantly reduce the original data size. The training time of the SE-CNN is 539 seconds, reducing 90.6 % compared with CNN.

中文翻译：

---

增强型频谱卷积神经架构：一种天然气管道的智能泄漏检测方法

摘要 在这项工作中，提出了一种结合频谱增强 (SE) 和卷积神经网络 (CNN) 的新型卷积神经架构 (SE-CNN) 来检测天然气管道的泄漏。SE 具有增强泄漏信号和降低背景噪声的作用。CNN可以自动提取泄漏特征，实现泄漏诊断。实验结果表明，SE-CNN 对 6 个类别的平均准确率可以达到 94.3%，并且只需要 1.04 秒的检测时间。本实验中，主管道和支管道的直径分别为125mm和25mm。由于其出色的准确性和效率，所提出的增强型频谱卷积神经架构为工业环境中的实时泄漏检测铺平了道路，从而保证燃气管道输送的过程安全。在强背景噪声下，SE-CNN 的平均准确率可以达到 94.3%，比 SVM 和 CNN 高 33%，3.7%。特别是SE可以看作是一种数据压缩方法，可以显着减少原始数据的大小。SE-CNN 的训练时间为 539 秒，与 CNN 相比减少了 90.6%。