Event Detection from Water Quality Time Series

Detection of anomalous events in water distribution systems is of interest for both daily operations focused on delivery of high quality water as well as for identification of accidental or intentional contamination events. In lieu of network-wide deployment of in-situ contaminant-specific sensors, data streams resulting from in-situ monitoring of ambient water quality are employed as input to event detection algorithms to identify periods of anomalous water quality. The basis of these approaches is prediction of the future water quality values (state estimation) and then comparison of the prediction errors, the differences between predicted and measured water quality signals, to identify outliers in an on-line framework. These algorithms generally rely on a stationary time series and large, sudden changes within the time series make outlier detection difficult. Here we propose an approach to improving the identification of events, defined as a cluster of outliers, that will also identify changes in the baseline water quality. This approach is called the binomial event discriminator (BED) and it uses a failure model based on the binomial distribution to determine the probability of an event existing based on r outliers occurring within n time steps. If the consecutive number of outliers exceeds an upper limit, a change in the baseline water quality is declared. The BED is applied to observed water quality collected at a location within a utility distribution system. The BED is able to reduce the number of false positive event identifications by several orders of magnitude compared to not using the BED. The BED is also identifies two locations as baseline water quality changes.