|M.Sc Student||Jonathan Arad|
|Subject||Water Distribution Systems Contamination Event Detection|
|Department||Department of Civil and Environmental Engineering||Supervisor||Full Professor Ostfeld Avi|
|Full Thesis text|
Urban water distribution systems consist of numerous elements and as such are inherently vulnerable. Physically securing each access point is not feasible. Research shows that it is reasonable to assert that regularly monitored water variables are affected in presence of a contaminant. For this reason the surrogate approach (i.e. identifying contaminants using regularly monitored water measurements), is appealing. This study focuses on interpreting data collected from sensors measuring routine variables from a single location to reveal outliers indicating possible contamination event intrusions.
A generic framework was constructed utilizing two estimation models, Decision trees and Artificial neural networks, in two different detection schemes: Fixed and Dynamic Thresholds approaches. Estimation models aim to model the behavior of historic data containing routine operation conditions water variables in an off-line manner. Imitating on-line conditions, estimation models predict normal behavior of variables and distinguish abnormal readings. Both Fixed and Dynamic schemes integrate sequential probability updating resulting in a contamination event probability.
Fixed threshold approach is based on the distinction between normal readings and abnormal behavior utilizing constant thresholds values. The method is based on a two staged approach - Training and execution procedures. In the execution procedure Bayes’ rule is recursively invoked increasing and decreasing the probability of a contamination event.
The Dynamic methodology is based on training and execution procedures as well. During the training stage, a genetic algorithm is utilized for tuning five decision variables: positive and negative filters, positive and negative dynamic thresholds, and window size. During the execution stage, Bayes’ rule is employed, invoking the five decision variables.
Outliers identified undergo further analysis to determine if they constitute a contamination event or a false alarm. Models are assessed through correlation coefficient, mean squared error, confusion matrices, receiver operating characteristic curves, and true and false positive rates.
Simulated contamination events test the ability of proposed approaches to successfully identify contamination events. The product of the methodologies consists of alarms indicating a possible contamination event. The objective of the constructed platform is to grant the decision maker tools for a comprehensive yet understandable view point of all available data gathered and analyzed.
Applications presented utilized measured data, available online, gathered from a utility in the United States. Applications results show the constructed platform function well under given conditions with relatively high success rate and robustness. Comparing approaches shows advantage to the Dynamic scheme in both true positive and false alarm rates.