基于CNN和MFCC的供水管网漏损声信号识别方法  

A Method for Identifying Leakage Acoustic Signals in Water Supply Network Based on CNN with MFCC

作　　者：陈炯禧 王琦[1,2] 詹凡[1] 陈彦冰 黄颀 张宏洋 王志红 陈贡发[1,2] 赵志伟 辛萍 CHEN Jiong-xi;WANG Qi;ZHAN Fan;CHEN Yan-bing;HUANG Xin;ZHANG Hong-yang;WANG Zhi-hong;CHEN Gong-fa;ZHAO Zhi-wei;XIN Ping(School of Civil and Transportation Engineering,Guangdong University of Technology,Guangzhou 510006,China;Cross Disciplinary Research Institute of Marine Engineering Safety and Sustainable Development Innovation,Guangdong University of Technology,Guangzhou 510006,China;School of Water Conservancy and Hydroelectric Power,Hebei University of Engineering,Handan 056038,China;Shenzhen ANSO IOT Co.Ltd.,Shenzhen 518052,China)

机构地区：[1]广东工业大学土木与交通工程学院,广东广州510006 [2]广东工业大学海洋工程安全与可持续发展创新交叉研究院,广东广州510006 [3]河北工程大学水利水电学院,河北邯郸056038 [4]深圳拓安信物联股份有限公司,广东深圳518052

出　　处：《中国给水排水》2024年第23期13-19,共7页China Water & Wastewater

基　　金：广州市重点研发计划农业和社会发展科技专题项目(2023B03J1333)。

摘　　要：针对供水管网漏损识别效率低和对人工经验依赖性强等问题,基于卷积神经网络(CNN)和梅尔频率倒谱系数(MFCC)提出了一种供水管网漏损声信号识别方法。对噪声记录仪和水音传感器采集的漏损声信号提取MFCC及其一、二阶差分作为漏损声信号特征,得到了包含漏损特征的特征图像,将其输入到CNN模型,通过超参数优化后最终得到了漏损识别模型。结果表明,使用MFCC与MFCC的一阶差分特征参数组合作为输入特征训练模型时的识别效果最好,其测试集准确率达到95.26%,F1分数达到89.22%,具备优良的漏损识别能力。Aiming at the problems of low efficiency and strong dependence on artificial experience in leakage identification of water supply network,a leakage acoustic signal identification method of water supply network was proposed based on convolutional neural network(CNN)and Mel frequency cepstral coefficient(MFCC).MFCC and its first-and second-order differences were extracted from the leakage sound signals collected by noise recorders and water sound sensors as leakage sound signal features.Feature images containing leakage features were obtained and input into a CNN model.After hyperparameter optimization,the leakage identification model was finally obtained.The results showed that using the combination of MFCC and its first-order differential feature parameters as input features to train the model yielded the best identification performance,with a test set accuracy of 95.26%and an F1 score of 89.22%,demonstrating excellent leakage identification ability.

关 键 词：供水管网 漏损识别 声学信号 卷积神经网络 梅尔频率倒谱系数 

分 类 号：TU991[建筑科学—市政工程]