Research On Contamination Event Detection Method Based On Multiple Conventional Water Quality Sensors

Drinking water sources are important public infrastructures to ensure normal society operations and water supply security in urban and rural areas. Water quality directly affects the water quality of water distribution system and the drinking water security of residents. However, in recent years, more and more contaminant events occur frequently, and the severity of events is rising. Facing the practical problems in establishment and enhancement of water monitoring and early warning system, and considering the operability in practical construction, the present study is aimed to create a reliable and operable event detection method based on traditional online monitoring sensors.Most researchers focused on the optimization of different event detection methods. However, the event data were from computer simulations. These methods may not have good performance on real contaminant events. In this study, a pilot-scale experiment system was built to simulate fifteen kinds of contaminant events. The establishment of detection method was built based on the experiment data.In this study, a detection method based on Pearson-coefficient is proposed first time and optimized using non-dominated sorting genetic algorithm. Ten kinds of contaminant events(32 groups of events) were optimized and the Pareto-front was obtained, which provided reasonable selection of parameters to the decision makers. When considering the balance of true positive rate(TPR) and false positive rate(FPR), the TPR can reach 95.5% and FPR can reach 4.4%. When considering low FPR is more important, the TPR can reach 93.0% and FPR can reach 3.1%. Both solutions have good detection performances.Three datasets were utilized to test the reliability of this method. The first dataset A contained experiment data of same contaminant categories in the training datasets, and the TPR can reach 93.3%. The second dataset B contained experiment data of unknown contaminant categories, and the TPR can reach 93.6%. The third dataset C contained routine baseline data, and the FPR is only 0.35%. The results showed that this method has a high reliability.The robustness of the method was tested using dataset A and B based on latin hyper-cubic sampling. The results showed that the average robustness value of two datasets was 0.8520, which indicated a good performance.Finally, three event methods proposed by other researchers were tested and evaluated using the experiment data in this study. The results showed that the performance of multi-variable Euclidian distance method is better than the other two methods, but was far inferior compared with the proposed method.