Water mains, especially old pipelines, are consistently threatened by the formation of leaks. Leaks inherit increased direct and indirect costs and impacts on various levels such as the economic field and the environmental level. Recently, financially capable municipalities are testing acoustic early detection systems that utilize wireless noise loggers. Noise loggers would be distributed throughout the water network to detect any anomalies in the network. Loggers provide early detection via recording and analyzing acoustic signals within the network. The city of Montreal adopted one of the leak detection projects in this domain and had reported that the main issue that hinders the installed system is false alarms. False alarms consume municipality resources and funds inefficiently. Therefore, this paper aims to present a novel approach to utilize more than one data analysis and classification technique to ameliorate the leak identification process. In this research, acoustic leak signals were analyzed using Fourier Transform, and the multiple frequency bandwidths were determined. Three models were developed to identify the state of the leak using Naïve Bayes (NB), Deep Learning (DL), and Decision Tree (DT) Algorithms. Each of the developed models has an accuracy ranging between 84% to 89%. An aggregator approach was developed to cultivate the collective approaches developed into one single answer. Through aggregation, the accuracy of leak detection improved from 89% at its best to 100%. The design, implementation approach and results are displayed in this paper. Using this method helps municipalities minimize and alleviate the costs of uncertain leak verifications and efficiently allocate their resources.

中文翻译：

---

改进泄漏检测系统的集体思维方法

自来水管，特别是旧管道，始终受到泄漏形成的威胁。泄漏会导致直接和间接成本的增加以及对各个层面（例如经济领域和环境层面）的影响。最近，有经济能力的市政当局正在测试利用无线噪声记录器的声学早期检测系统。噪声记录仪将分布在整个供水网络中，以检测网络中的任何异常情况。记录仪通过记录和分析网络中的声音信号来提供早期检测。蒙特利尔市采用了该领域的一个泄漏检测项目，并报告说，阻碍已安装系统的主要问题是虚假警报。虚假警报会低效地消耗市政资源和资金。因此，本文旨在提出一种新颖的方法，利用多种数据分析和分类技术来改善泄漏识别过程。在这项研究中，使用傅立叶变换分析了泄漏声信号，并确定了多个频率带宽。使用朴素贝叶斯（NB），深度学习（DL）和决策树（DT）算法，开发了三种模型来识别泄漏状态。每个开发的模型的精度在84％到89％之间。开发了一种聚合器方法，以培养发展为单个答案的集体方法。通过聚合，泄漏检测的准确性从最佳状态的89％提高到100％。本文展示了设计，实现方法和结果。