Data Analytics for Automated Near Real Time Detection of Blockages in Smart Wastewater Systems

Abstract

Blockage events account for a substantial portion of the reported failures in the wastewater network, causing flooding, loss of service, environmental pollution and significant clean-up costs. Increasing telemetry in Combined Sewer Overflows (CSOs) provides the opportunity for near real-time data-driven modelling of the sewer network. The research work presented in this thesis describes the development and testing of a novel system, designed for the automatic detection of blockages and other unusual events in near real-time. The methodology utilises an Evolutionary Artificial Neural Network (EANN) model for short term CSO level predictions and Statistical Process Control (SPC) techniques to analyse unusual CSO level behaviour. The system is designed to mimic the work of a trained, experience human technician in determining if a blockage event has occurred. The detection system has been applied to real blockage events from a UK wastewater network. The results obtained illustrate that the methodology can identify different types of blockage events in a reliable and timely manner, and with a low number of false alarms. In addition, a model has been developed for the prediction of water levels in a CSO chamber and the generation of alerts for upcoming spill events. The model consists of a bi-model committee evolutionary artificial neural network (CEANN), composed of two EANN models optimised for wet and dry weather, respectively. The models are combined using a non-linear weighted averaging approach to overcome bias arising from imbalanced data. Both methodologies are designed to be generic and self-learning, thus they can be applied to any CSO location, without requiring input from a human operator. It is envisioned that the technology will allow utilities to respond proactively to developing blockages events, thus reducing potential harm to the sewer network and the surrounding environment.