Deep Autoencoder Based Leakage Detection in Water Distribution SCADA Systems

Abstract

Recent advances in industrial automation linked to the development and pervasive implementation of Industrial Internet of Things (IIoT) bring greater networking and communication possibilities, as well as higher volumes of available data about the systems and processes. Modern Supervisory Control and Data Acquisition (SCADA) systems follow these advances by encompassing the convergence between IT (information technology) and OT (operational technology), leading to increased possibilities for data analysis and predictions, providing faster, smarter and more flexible systems with early error detection. Such systems can bring great improvements in water distribution systems, which are becoming increasingly complex following the population growth, and should be reliable and safe since they directly influence public health. In these big and complex systems early error detection is crucial, as it can reduce maintenance costs and detect errors before they reach catastrophic stage. One of the common problems in such systems is water leakage, which usually happens due to pipe bursts and poor infrastructure. With increased system complexity, error detection is getting more complicated and often can be discovered too late. In this paper we propose a model based on a deep autoencoder neural network for leakage detection in water distribution systems. We train and test the model using the LeakDB benchmark algorithm. Such a model can be applied to real world data in a SCADA system to predict and discover water leakage errors, thus lowering the repair costs and improving overall system reliability.