Model-Based Event Detection for Contaminant Warning Systems

Abstract

Security issues have become increasingly important within distribution systems, which have led to the development of event detection algorithms (EDAs) to provide timely detection of intrusion events. The current study develops a localized model-based event detection algorithm that utilizes nonspecific water quality sensors to identify water quality anomalies. The proposed EDA focuses on evaluating a series of multivariate error signals between the observed signals and the model estimated signals based on a moving time-window of error statistics. The likelihood of the multivariate error signals is estimated using the product of univariate kernel density estimation (KDE), which is a type of nonparametric representation of the error distribution. A comprehensive analysis was performed using synthetic events to explore the combination of the moving window-pairs and bandwidth with respect to three injection strengths and two injection durations. In addition to the synthetic events, the EDA was also evaluated using a more realistic approach that simulates the water quality parameters in response to two real contaminants (KCN and nicotine) based on previously developed water quality dynamic models. Overall, the model-based EDA was capable of detecting anomalous water quality events through the statistical evaluation of multivariate error signals with performance related to the magnitude of the event. The results indicate that smaller events resulting from hydraulic/transport dynamics can have a significant impact on the EDA performance, which are not typically considered in most EDA evaluations. The results of the proposed EDA also illustrate that sensor performance should be considered within other water security activities such as the optimal design of sensor-based contaminant warning systems.