Iot and Neural Network Based Multi Region and Simultaneous Leakage Detection in Pipelines

Abstract

The increasing demand for water arising from global population growth and urbanization in recent years is stressing the water supply to its limits. By 2025, 1.8 billion people will experience absolute water scarcity, and 2/3 of the world will be living under water-stressed conditions [1-2]. Neural networks have proved to be an apt approach for water leakage detection as they have the best and most extensive reach on the problem owing to their computational nature. They do not any have basic flaws unlike alternate methods like leakage detection using acoustic sensors which cannot differentiate between spikes in flow and leakage. They are a flexible and efficient approach to detection of leakages in water distribution networks. According to an inquiry made by the International Water Supply Association (IWSA), the amount of lost or “unaccounted for water” (UFW) is typically in the range of 20–30% of production[3]. In this project, a neural network model is proposed for detection and location of leakages in the pipes based on pressure values from sensors deployed along the pipeline. The network is initially trained using these pressure values and can then be used to detect abnormalities in the readings which can be due to leakages. The open source tool used to develop this model is Neuroph Studio. Neuroph is a neural network framework written in Java. It can be used to create and train neural networks in Java programs. Neuroph provides Java class library as well as GUI tool to quickly create Java neural network components. The model consisting of a multilayer perceptron neural network identifies simultaneous leakages in multiple regions successfully. When the size of the input dataset increases from a set of 10 values to a set of 1500 values, the mean square error of outputs increases by 128 times. But when this change is from a set of 1500 values to a set of 12000 values, the mean square error increases by 1.6 times. Thus, the total mean square error decreases drastically with the increase in input size, leading to the conclusion that the model is stable and scalable.