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

Abstract

Contamination events in water distribution networks (WDNs) can have a huge impact on water supply and public health; increasingly, online water quality sensors are deployed for real-time detection of contamination events. Machine learning has been used to integrate multivariate time series water quality data at multiple stations for contamination detection; however, accurate extraction of spatial features in water quality signals remains challenging. This study proposed a contamination detection method based on generative adversarial networks (GANs). The GAN model was constructed to simultaneously consider the spatial correlation between sensor locations and temporal information of water quality indicators. The model consists of two networks-a generator and a discriminator-the outputs of which are used to measure the degree of abnormality of water quality data at each time step, referred to as the anomaly score. Bayesian sequential analysis is used to update the likelihood of event occurrence based on the anomaly scores. Alarms are then generated from the fusion of single-site and multi-site models. The proposed method was tested on a WDN for various contamination events with different characteristics. Results showed high detection performance by the proposed GAN method compared with the minimum volume ellipsoid benchmark method for various contamination amplitudes. Additionally, the GAN method achieved high accuracy for various contamination events with different amplitudes and numbers of anomalous water quality parameters, and water quality data from different sensor stations, highlighting its robustness and potential for practical application to real-time contamination events.

Keywords: Contamination detection; Generative adversarial network; Multi-site time series data; Water distribution system; Water quality.

Graphical abstract

Fig. 1

Schematic of the proposed GAN-based method for spatial contamination event detection.

Fig. 2

The architecture of the proposed GAN model.

Fig. 3

Real-world WDN case study (the Yantian Network).

Fig. 4

Distribution and sequential increment of the GAN-based anomaly scores using single-site (a, c) and multi-site measurements (b, d).

Fig. 5

Event alarms of single-site, multi-site, and combined models for training (a) and testing (b) data sets with contamination events.

Fig. 6

The time series of the normalized water quality parameters monitored by Sensor Group 1 and the GAN-based anomaly scores of both single-site and multi-site models for the event are highlighted in Fig. 5.

Fig. 7

Receiver operating characteristic (ROC) curves of both the combined GAN-based and the MVE-based models using data from different groups of sensors during the testing experiments for contamination events with different amplitudes: a–d, ROC curves using Sensor Group 1 for events with amplitudes of 1.0–1.5 (a), 1.5–2.0 (b), 2.0–2.5 (c), and 2.5–3.0 (d); e–h, ROC curves using Sensor Group 2 for events with amplitudes of 1.0–1.5 (e), 1.5–2.0 (f), 2.0–2.5 (g), and 2.5–3.0 (h).

Fig. 8

Distribution of four evaluation indicators for different amplitudes and the number of influenced water quality parameters during the testing experiments: a, c, e, g, Group 1; b, d, f, h, Group 2.

Fig. 9

Event alarms of the proposed GAN-based model using Sensor Group 2 during one testing process (contamination characteristics: amplitude is 2.0–2.5 and the number of influenced water quality parameters is 5).

Fig. 10

Distribution of false alarms for events with different amplitude and different numbers of influenced water quality parameters using Sensor Group 2 during the testing experiments, with consideration of different event duration flags: a, alarms during the time of contamination injection (in this study, each contamination event lasted 10 h) are regarded as true alarms; b, alarms within 24 h of the start of contamination injection are regarded as true alarms.