Testing and Evaluation of Water Quality Event Detection Algorithms

Event detection systems provide online analysis of water quality data for identification of significant water quality events. Two different online algorithms are tested here with multivariate data from two monitoring locations in an operating water distribution network. The data are split into training and testing sets and parameter identification is completed on the training data prior to application on the testing data. Water quality events are added to the testing data sets as perturbations from the measured water quality using 11 different event strengths. Resulting receiver operating characteristic curve areas quantify the relationship between probability of detection and false detections at the time step scale. Additionally, the proportion of events containing at least one detection is measured. Results show that both algorithms are capable of reliably detecting events that change the background water quality by 1.5 times the standard deviation of the water quality signal while limiting the false-positive results to 3–4% of the time steps. Trade-offs in the delay to detection versus the number of false-positive results are examined in the context of the event length.