Real-time probabilistic contaminant source identification and model-based event detection algorithms

The development of sensor-based contaminant warning systems (CWS) has extended beyond sensor placement algorithms into forensic algorithms such as event detection algorithms (EDA) and source identification algorithms. The objectives of the current study were focused on the development of a probabilistic contamination source identification algorithm (PCSI), as well as the development and evaluation of both localized and system-wide model-based EDAs. The PCSI algorithm was developed to overcome the limiting assumptions of other source identification algorithms (e.g., known hydraulics, perfect sensors, single injection location). Using the binary signals from a localized EDA, the resulting hydraulic path was traced backward in time and the probabilities of potential sources estimated with two different Bayesian updating procedures - a Beta-Binomial conjugate pair approach and a simpler Bayes' Rule approach. Results showed that the Beta-Binomial approach demonstrated better selectivity than the Bayes' Rule approach. Additionally, the PCSI algorithm was shown capable of accounting for false positive/negative responses as well as providing the flexibility for the successful identification of multiple source locations. Ultimately, CWS design and the performance of forensic tools are dependent on the performance of the EDAs. However, current EDA evaluation approaches do not typically account for transport characteristics within the network and/or the actual changes of common water quality parameters in response to a contaminant. Thus, water quality models were developed to represent the dynamics of chlorine, pH and conductivity in response to two contaminants (KCN and nicotine). The simulation studies demonstrated that current EDA evaluation approaches, as well as CWS design assumptions, may not adequately represent the EDA performance under conditions likely to be observed within a distribution system. These water quality models were also used to evaluate the model-based EDAs developed as part of this study. A model-based localized EDA was proposed to identify the "true" event from background noise by evaluating the likelihood of a series of multivariate error signals using multivariate kernel density estimation with a moving time-window. The evaluation was based on the use of both "synthetic" events as well as simulated water quality dynamics in response to two contaminants (discussed above). The results demonstrated the capabilities of the model-based EDA to detect anomalous events, as well as the significant impacts that network hydraulics and transport can have on EDA performance, which are typically not considered in current EDA evaluation approaches. A system-wide EDA was developed by integrating the binary signals from multiple localized EDAs through the PCSI algorithm with an alarm threshold based on the probability of network contamination. The proposed approach was compared against the performance of a localized system-wide EDA using two different simulated injection scenarios that produced different amounts of sensor information. In general, the results demonstrated that integrating the binary signals from the localized EDA provided better system-wide performance than relying solely on the individual localized EDAs. From a broader perspective, these results suggest that more realistic water quality dynamics should be considered when assessing EDA performance and be utilized to provide meaningful information for the design of sensor-based CWS.