Sensor-based contamination warning systems are being developed to detect intentional intrusion events and provide information on day-to-day performance of water distribution systems. Once a contamination warning system has detected contamination at one or more sensor locations, remedial actions should apply methods to identify the possible source locations. Current research on contamination source identification assumes that the information on contamination detection comes from perfect sensors. This assumption ignores the possibility of false positives (treating them as "true" events) and false negatives (treating every negative as though no event has occurred), which can presumably degrade effectiveness of remedial actions, and result in severe effects on public health. This research studies the impacts of imperfect sensor measurements (false positive or negative readings) on the contamination source identification. A probabilistic approach based on Bayes' theorem is applied to estimate the probability of each node in the water distribution system being a contamination source. Specifically, the probability of a paired location and time being a potential source is updated based on: 1) single or multiple sensor signals, which are dependent on the hydraulic connectivity determined through a flow path analysis model, and 2) the true and false reading information, which is dependent on background water quality and sensor behavior. Hence, the probabilistic state of a possible source can be iteratively modified in real-time using an inference system that combines the previous knowledge and information from every new sensor reading in the network. A simulation study using a network with a hypothetical sensor system is presented to illustrate the time varying probabilities associated with the candidate contamination sources.