| 谭震,郭新蕾,李甲振,郭永鑫,潘佳佳.基于多尺度卷积神经网络的管道泄漏检测模型研究[J].水利学报,2023,54(2):220-231 |
| 基于多尺度卷积神经网络的管道泄漏检测模型研究 |
| Multi-scale convolutional neural network model for pipeline leak detection |
| 投稿时间：2022-08-10 |
| DOI：10.13243/j.cnki.slxb.20220644 |
| 中文关键词: 管道泄漏检测 瞬变流模型 非恒定摩阻系数 多尺度泄漏特征 卷积神经网络 |
| 英文关键词: pipeline leak detection transient flow model unsteady friction coefficient multi-scale leak feature convolutional neural network |
| 基金项目:国家重点研发计划项目（2022YFC3200215，2022YFC3202500）；国家自然科学基金项目（U2243221，51979291，52179082） |
|
| 摘要点击次数: 3093 |
| 全文下载次数: 1429 |
| 中文摘要: |
| 如何有效检测管道泄漏是节水型社会建设迫切需要解决的关键和热点问题之一。近年来基于深度学习的管道泄漏检测方法发展迅速，本文针对传统单尺度卷积神经网络对泄漏特征提取不充分的问题，提出一种基于多尺度一维卷积神经网络（MS1DCNN）的管道系统泄漏检测模型。该方法利用多个不同卷积尺度的卷积通路并行提取管道泄漏的特征并进行泄漏信息的分类预测。基于经典的管道系统布置，利用瞬变流模型生成管道泄漏工况下的三个水压数据集对模型进行验证，三个数据集分别用于预测管道的泄漏位置、泄漏量和非恒定摩阻系数，对应样本数为39601、3980、4900，并将预测结果与其他深度学习方法和传统的机器学习方法进行对比分析。结果表明：MS1DCNN模型对数据集样本下泄漏位置、泄漏量、非恒定摩阻系数的分类准确率达到99.96%、98.48%、100%，三者平均预测精度比传统一维卷积神经网络（1DCNN）、BP神经网络、支持向量机（SVM）和k近邻算法（KNN）提高0.31%、2%、1.27%、22.8%；MS1DCNN在信噪比为-4～12 dB的噪声环境下各数据集的平均F1分数分别为99.2%、97.02%、100%，三者平均值分别比1DCNN、BP、SVM、KNN模型提高0.61%、2.3%、2.78%、28.59%，证明了MS1DCNN模型的预测性能。本文方法在管道泄漏参数、非恒定摩阻的同步预测方面有一定的潜力。 |
| 英文摘要: |
| How to detect pipeline leaks effectively is one of the key and hot research topics to be solved in the construction of water-saving society.Recently, pipeline leak detection methods based on deep learning have been developed rapidly.This paper proposes a pipeline leak detection model for pipeline systems based on multi-scale one-dimensional convolutional neural network (MS1DCNN) for the problem of inadequate extraction of leak features by traditional single-scale convolutional neural network.The model adopts multiple convolution paths with different convolution kernel sizes to extract the features of pipeline leaks and to classify the leak information.Based on the classical pipeline system layout, in order to validate the model, the transient flow model was used to generate three water pressure data sets which were used to identify the leak location, leakage and unsteady friction coefficient respectively under various pipeline leak cases.The sample numbers were 39601, 3980 and 4900.The model has been compared with other deep learning methods or traditional machine learning methods like traditional one-dimensional convolutional neural network (1DCNN), BP neural network, support vector machine (SVM) and k-Nearest Neighbor (KNN).The results show that the MS1DCNN model achieves 99.96%, 98.49% and 100% of the classification accuracy for leak location, leakage and unsteady friction coefficient under the data sets.The average prediction accuracy is 0.31%, 2.00%, 1.27% and 22.80% higher than other models, respectively.The average F1 scores of MS1DCNN model for each data sets are 99.2%, 97.02% and 100% respectively in the noise environment with SNR of -4-12 dB.They are 0.61%, 2.3%, 2.78% and 28.59% higher than those of 1DCNN, BP neural network, SVM and KNN, respectively, which proves the prediction performance of MS1DCNN model.The model presented in this paper is applicable for the synchronous prediction of pipeline leakage parameters and unsteady friction coefficient. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|