基于贝叶斯优化FCN-DenseNet算法的供水管网爆管智能识别  

Research on burst intelligent identification in water distribution network based on Bayesian optimized FCN-DenseNet algorithm

作　　者：彭森[1] 程蕊 程景 吴卿[1] 田一梅[1] PENG Sen;CHENG Rui;CHENG Jing;WU Qing;TIAN Yi-mei(School of Environmental Science&Engineering,Tianjin University,Tianjin 300350,China)

机构地区：[1]天津大学环境科学与工程学院,天津300350

出　　处：《安全与环境学报》2022年第1期306-315,共10页Journal of Safety and Environment

基　　金：国家重点研发计划项目(2016YFC0802400)。

摘　　要：为了进一步拓展深度学习算法在供水管网爆管分析中的适用范围,提出了一种爆管区域识别方法。基于全连接稠密网络算法(Full Connect Network-DenseNet,FCN-DenseNet)构建了爆管区域识别模型,提取不同区域的爆管特征。同时,采用贝叶斯优化算法对识别模型的超参数组合进行优化和自动选取。以华东某园区的部分供水管网为研究对象,在建立水力模型的基础上,通过布置监测点和划分监测分区形成管网监测方案。综合考虑爆管位置、时间、流量、监测噪声及研究区域管网实际情况等因素,进行爆管工况模拟,建立爆管数据集,对识别模型进行训练和验证。结果表明,在验证数据集下,爆管区域平均识别率φ_(1)、φ_(2)分别可以达到93.5%、96.7%。贝叶斯优化的FCN-DenseNet算法增强了模型的性能和适用性,管网监测分区爆管可能性排序能够指导水司进行爆管分析和巡检。A method for burst district identification for water distribution network(WDN)was proposed to improve the application of deep learning algorithm to pipe burst analysis.A burst district identification model for WDN was established based on Full Connect Network-DenseNet(FCN-DenseNet)algorithm.The model takes the feature vector of pipe burst as the input and the monitoring district ID where the pipe bursts as the output.The FCN-DenseNet algorithm extracts the pressure features of different districts and identifies the district in which burst occurs.Hyperparameter is a group of parameters that define the attributes or training process of the FCN-DenseNet model.It has a great influence on the deep learning model performance.The identification accuracy of burst districts is the optimization objective and the Bayesian optimization algorithm optimizes and automatically selects the hyperparameter for FCN-DenseNet to achieve a better identification effect.K-means algorithm was used to analyze the node pressure sensitivity matrix based on a hydraulic WDN model of the research area located in an industrial district in East China.A monitoring scheme of WDN was developed,including pressure meter placement and network district arrangement.Different sets of burst scenario simulations were carried out and relevant factors such as the burst position,time,flow rate,monitoring noise,and actual maintenance records were taken into consideration.The identification model was trained and verified with the burst data sets generated according to the simulation.Results show that,compared with the models with stochastic hyperparameter,the performance of the Bayesian optimized model is significantly improved.With the validation data set,the average identification accuracy φ_(1) and φ_(2) can reach 93.5%and 96.7%,respectively.Bayesian optimized FCN-DenseNet algorithm enhances the performance and applicability of the model,and the possibility ranking of burst districts guides water companies to analyze pipe burst and formulate reasonable insp

关 键 词：安全管理工程 供水管网 爆管识别 管网分区 全连接稠密网络 贝叶斯优化 

分 类 号：X956[环境科学与工程—安全科学]