Page 104 - 《应用声学》2022年第4期

第 41 卷 第 4 期 Vol. 41, No. 4
2022 年 7 月 Journal of Applied Acoustics July, 2022

⋄ 研究报告 ⋄

卷积神经网络在气体泄漏超声识别中的应用

韩鹏程 1† 燕 群 1 彭 涛 1 宁方立 2

(1 中国飞机强度研究所 西安 710065)
(2 西北工业大学机电学院 西安 710072)

摘要：为了克服现有气体泄漏检测方法的不足，提出一种基于卷积神经网络的气体泄漏超声信号识别方法。
在设计卷积神经网络网络结构时，通过多次预训练确定网络层数、卷积核数目和尺寸、全连接层神经元数目。
同时，选择 Inception 模块平衡网络宽度和深度，防止过拟合的同时提高网络对尺度的适应性。通过输气管道
泄漏实验平台模拟工况中常见的阀门泄漏和垫片泄漏，利用短时傅里叶变换进行时频图表征，在此基础上，建
立二分类模型和不同泄漏类型的三分类模型。结果表明，相比二分类模型，不同泄漏类型的三分类模型识别准
确率有所降低，添加 Inception 模块可以有效提高三分类模型的性能。
关键词：气体泄漏；卷积神经网络；时频图
中图法分类号: TP391 文献标识码: A 文章编号: 1000-310X(2022)04-0602-08
DOI: 10.11684/j.issn.1000-310X.2022.04.012

Application of convolutional neural network in ultrasonic identification of
gas leakage

HAN Pengcheng 1 YAN Qun 1 PENG Tao 1 NING Fangli 2

(1 Aircraft Strength Research Institute of China, Xi’an 710065, China)
(2 School of Mechanical Engineering, Northwestern Polytechnical University, Xi’an 710072, China)

Abstract: In order to overcome the shortcomings of existing gas leakage detection methods, an ultrasonic
signal recognition method of gas leakage based on convolutional neural network (CNN) was proposed. When
designing the CNN network structure, the number of network layers, the number and size of convolution
kernel and the number of fully connected layer neurons were determined by multiple pre-training. Meanwhile,
Inception module was selected to balance the width and depth of the network, prevent overfitting and improve
the adaptability of the network to scale. The valve leakage and gasket leakage in working conditions were
simulated by the gas pipeline leakage experimental platform, and the short-time Fourier transform was used to
characterize the time-frequency diagram. Based on this, two-class model and three-class model with different
leakage types were established. The results show that compared with two-class model, the recognition accuracy
of the three-class model with different leakage types is reduced, and the addition of Inception module can
effectively improve the performance of the three-class model.
Keywords: Gas leakage; Convolutional neural network; Time-frequency diagram

2021-08-21 收稿; 2021-11-30 定稿
作者简介: 韩鹏程 (1993– ), 男, 陕西西安人, 硕士, 研究方向: 声源识别。
† 通信作者 E-mail: hpc8163@126.com