Abstract:
The ability to detect deliberate or accidental contamination of a water distribution system is of real concern to the safety and security of our nation’s drinking water. To address these concerns, increased attention has been placed on sophisticated monitoring of water distribution systems and the use of robust statistical analysis. Using existing data from in-situ water quality sensors, this paper explores the ability to detect anomalies in water quality using multivariate techniques. The algorithm developed in this study uses a multivariate distance measure between the current water quality measurement and the closest observation in multivariate space within a moving window of previous observations. To discriminate between normal and anomalous water quality, the distance measure is compared to a constant threshold. To test the algorithm, we utilize both simulated anomalous events and laboratory based events that correspond to real contaminants. These events are superimposed onto in-situ water quality recorded at four different locations within a single utility network. Measured water quality parameters include free chlorine, pH, temperature and electrical conductivity. Robust discrimination methods have a high probability of detecting anomalies with a low false alarm rate. Here, receiver operating characteristic (ROC) curves are used to test the ability of the multivariate classification algorithm to detect anomalous water quality while keeping false alarms low. This analysis explores the false alarm rate associated with detecting a range of anomalous water quality observations.
