Evaluation of Machine Learning-based Methods for Continuous Water Quality Data Analysis

Abstract

Wastewater treatment facilities are increasingly installing sensors to monitor water quality. As these datasets have increased in size and complexity, it has become difficult to identify abnormal readings in a timely manner either manually or using simple rules that might have been sufficient previously. Two ammonia sensors were installed at the Dundas Wastewater Treatment Plant in November 2017. The collected ammonia concentration data shows a daily pattern. A learning-based method is implemented in this thesis to identify any readings which violate this daily pattern. The data points which were predicted to be anomalous were qualitatively ranked based on the severity and the likelihood of being faulty. The result of the learning-based method was evaluated and compared to other traditional detection methods.