Comparison of two multivariate classification models for contamination event detection in water quality time series

This paper explores two applied classification models alerting for contamination events in water distribution systems. The models perform multivariate analysis of water quality online measurements for event detection. The developed models comprise an outlier detection algorithm and a following sequence analysis for the classification of events. The first model is an unsupervised minimum volume ellipsoid (MVE), which utilizes only normal operation measurements but requires calibration. The second is a supervised weighted support vector machine, which utilizes event examples and performs data-driven optimized calibration. The models were trained and tested on real water utility data with randomly simulated events that were superimposed on the original database. The models showed high accuracy and detection ability compared to previous studies. All in all, the MVE model achieved preferable results.