The design of an efficient online water quality monitoring
approach is complicated by several factors. First, changes
within the system including variation of source water or
the treatment process, the fluctuation of water demand,
and changes in the water quality sensors can create anomalous
water quality signatures, different from contamination
events. Distinguishing normal behavior from abnormal
behavior is difficult in these situations. Second, due to the
complex characteristics of water quality, a physical model
describing water quality behavior is difficult to derive and
implement in real time. This complexity has motivated the
extensive use of data-driven models for predicting water
quality behavior. Although data-driven approaches have
proven to be an efficient alternative, several challenges persist
including: accurate detection of contamination events
with low false positive rates (FPRs) and low detection