Developing stacking ensemble models for multivariate contamination detection in water distribution systems

Generative adversarial networks for detecting contamination events in water distribution systems using multi-parameter, multi-site water quality monitoring

“…The contamination events were generated artificially by adding random disturbances to the normal data set, as performed in most other related studies [ 19 , 21 , 22 ]. The disturbances were generated randomly by considering amplitude, duration, direction (e.g., increase or decrease in value of water quality parameters), and the number of influenced water quality parameters.…”

“…To improve the performance of methods for detecting contamination events, recent work has focused on using multi-parameter fusion algorithms to detect anomalous water quality [ 17 , 18 ]. In more recent studies, time series water quality data represented by six water quality parameters (i.e., total chlorine, pH, electrical conductivity (EC), temperature, TOC, and turbidity) were analysed to provide fused anomaly alarms for contamination events [ [19] , [20] , [21] , [22] , [23] ]. Such multi-parameter fusion algorithms usually represent a simple fusion of anomalous results from individual parameters that fail to fully explore the correlations between multiple parameters.…”

“…In a statistical approach, the determination of contamination event detection is often based on the distribution of water quality parameter data [ 7 , 13 , 26 ]. However, owing to the nonlinear and nonstationary characteristics of water quality, statistical methods are usually unsuitable for detecting small abnormal changes in WDNs [ 21 ]. Hydraulic-model-based approaches detect contamination events by comparing observed real-time data with predicted values using a water quality and hydraulic network model [ [27] , [28] , [29] ].…”

“…The machine learning approach is considered an alternative for predicting real-time data of water quality parameters and identifying anomalous contamination events. Various machine learning algorithms have been applied for contamination event detection in WDNs, such as artificial neural networks [ 10 , 27 , 30 ], support vector machine [ 31 , 32 ], ensemble stacking models [ 21 ], and long short-term memory [ 33 ]. These models can capture the features of water quality time series data based on tests using a database compiled from the output of a single-site sensor.…”

Research on Modeling Method of Massive Launch Vehicle Time Series Data

“…The number of parameter fields in the slowly varying parameter set is uncertain, while the number of parameter fields in the instruction parameter set is determined. Due to the difference of two class, and the data models need to be designed respectively; 4) In aerospace engineering, there is more writing and less reading [11] with few operations of updating and deleting. Parameter sets' data sources are unique, meaning the data records of different parameter sets are independent; 5) Query applications pay more attention to the trend over a period of time [12] which is different from social applications , where every user pays more attention to each data record.…”