基于生成式对抗网络的供水管道漏水信号检测方法

Water Supply Pipeline Leakage Signal Detection Method Based on Conditional Generative Adversarial Network

针对供水管道漏水声音信号数据集匮乏、漏水情况多样需反复采集确定漏水、检测准确率低等问题,提出一种基于条件生成对抗网络的增强漏水信号数据集的方法。将深度对抗网络与条件生成对抗网络相结合对漏水信号数据集进行数据增强,用扩充后的数据集对一维卷积神经网络进行训练并对不同实地采集的样本进行漏水信号识别。验证表明:一种管质的某种程度漏水信息经对抗网络进行数据增强后,具有该管质未采集的漏水信号特征,能用于更加细微的漏水信号检测。该方法也适用于其它管质各种情况的漏水检测,具有良好的实用性。

In light of lack of sound signal data set of water supply pipeline leakage, difference of the water leakage conditions, demand of the repeated collection and determination of water leakage, and the low accuracy of detection, a method for enhancing water leakage signal data set based on conditional generative adversarial network was proposed in the paper. The data set of water leakage signal was enhanced through combining the depth adversarial network with the conditional generative adversarial network, and the different samples collected in the field were recognized following the one-dimensional convolutional neural network were trained with the expanded data set. The empirical study demonstrated that signals of a certain degree of water leakage from a kind of pipe had the characteristics of signals of uncollected pipe of same kind after the data was enhanced by the adversarial network, which was able to be used for more subtle water leakage signal detection. The method was also applicable to water leakage detection of other kinds of pipes under various conditions and had good applicability.