Multi-level information fusion for spatiotemporal monitoring in water distribution networks

Highlights

• Multi-level information fusion offers improved water quality monitoring under uncertainty.

• The NF-BRB and high-level BRB systems provide flexible inference frameworks based on WQP features.

• Hybrid learning and dynamic fusion algorithms significantly enhance the detection performance.

• Validation strategy with projecting anomaly signal patterns allows performance evaluation.

Abstract

This paper deals with online water quality monitoring in distribution networks based on surrogate water quality parameters (WQPs). The present strategy is based on multi-level information fusion using hierarchical belief rule-based (BRB) systems. Networked fuzzy belief rule-based (NF-BRB) and high-level BRB systems are introduced for information fusion at the feature level. Primary and secondary features are extracted from online WQP signals. Primary features are analyzed using the NF-BRB system that is built through knowledge elicitation from experts. Secondary features are interpreted through the high-level BRB system that employs a fuzzy partitioning on the feature sets and a hybrid learning strategy for its rule base construction. Finally, the dynamic fuzzy evidential fusion is introduced to aggregate the local and spatial assessments in each analysis window. As an important contribution of this paper, we propose a new validation method for event detection in the water distribution network (WDN) based on adaptive projection of the signal patterns attributed to anomaly events, obtained through contamination experiments in a pilot facility, to the real WQP signals measured across the WDN. Single and composite contamination events based on several biological and chemical contaminants are simulated to evaluate the performance of the proposed framework in event detection. The proposed multi-level information fusion framework obtains a high detection rate and a reduced number of false negative and positive results.

Introduction

Online monitoring of distributed infrastructure systems such as water distribution networks is necessary to assure the safety of water supply systems and management of health risks. Since water quality can deteriorate after leaving the treatment plant, monitoring policies are required at the distribution network level. Often various patterns of uncertainty and information deficiency about spatial relationships and interdependency of distributed observations impede effective monitoring and data interpretation in distributed systems. As a result, monitoring of distributed complex systems requires specific signal processing, data interpretation, and event analysis methods. One major shortcoming of the current water quality monitoring systems is the lack of inference mechanisms that provide interpretable operational information to the operators and facility management team.

The emerging sensor technology allows online measurements of surrogate WQPs (Ernest et al., 2008, WERF, 2008). Successful event detection and alarming in water treatment plants and distribution networks is largely dependent on the employed computational data analysis and their discriminatory ability. The current methods to detect anomalies in the WDN can be categorized in three groups including statistical (Allgeier et al., 2005, McKenna et al., 2007, Murray et al., 2010, Shang et al., 2007), empirical AI-based (Allgeier et al., 2005, Raciti et al., 2012), and data mining (Koch and McKenna, 2011, McKenna et al., 2007, Murray et al., 2010, Yang et al., 2008) methods. Some of these event detection systems are only suited for data collected in a single monitoring station to indicate the occurrence of a contamination event (Byer and Carlson, 2005, Cook et al., 2006, Murray et al., 2010, Yang et al., 2009) while other methods are based on two monitoring locations to improve the aggregation results using one of the nodes as the reference to compensate for the calibration error of the other node, varying time delays, and background noise (Yang et al., 2008, McKenna et al., 2007). Kumar et al. (2007) studied some of these event detection algorithms and demonstrated that tested algorithms do not provide unique responses in similar situations. Many of these problems are attributed to the inconsistency in the indications of WQPs and nonlinearity in the inference process. In order to improve the event detection performance, specific data fusion methods were introduced to incorporate other types of information including operational data (Hart, Mckenna, Murray, & Haxton, 2010), sampling location specific features (Raciti et al., 2012), and multiple sampling locations data (Hou et al., 2013, Koch and McKenna, 2011, Murray et al., 2010).

Due to the complex nature of water quality degradation process, various hypotheses can be associated with the deviation of WQPs from their regulatory ranges (Adams, 2009, Dawsey et al., 2006, Francisque et al., 2009, Lu and Huang, 2009, Mounce et al., 2009). While above methods use empirical AI-based or data mining algorithms, they do not employ a framework that allow incorporating various types of uncertainty in causal spatiotemporal relationships of WQPs to water quality events. Thus, more flexible knowledge representation frameworks that allow a set of hypotheses as the plausible indications of significant variations in WQPs can provide a promising solution.

The belief rule-based (BRB) system with an embedded belief structure in its consequent can model nonspecificity when the expert cannot make strong judgments. Alternatively, imprecise and incomplete data may be used to build the rule base that governs the BRB system (Chen et al., 2013, Liu et al., 2004, Yang et al., 2006). A well-known BRB system called “rule base inference methodology using evidential reasoning (RIMER)” has been employed in various applications that involve uncertainty and incompleteness (Wang et al., 2006, Xu et al., 2007, Yang and Xu, 2002a, Yang and Xu, 2002b). BRB systems were successfully employed for improved water quality assessment in one sampling location (Aghaarabi et al., 2014, Aminravan et al., 2011, Aminravan et al., 2012). However, BRB systems have never been adopted for spatiotemporal monitoring in the WDN.

This work involves the application of BRB systems and evidential reasoning algorithms to a real municipal problem: the determination of water quality degradation events through the fusion of surrogate online monitored WQPs. This problem is formulated as nonlinear function approximations of features extracted from surrogate WQPs, spatial variations of the monitored WQPs, and imprecise temporal relationships between monitored WQPs at different nodes. A multi-level information fusion framework is proposed to alleviate some deficiencies in data interpretation in previous research works for distributed information fusion in the WDN. The multi-level information fusion framework efficiently performs data aggregation in the WDN at measurement, feature, and decision levels. The multi-level information fusion framework encompasses three sub-modules: (i) spatiotemporal sensor data preprocessing, (ii) primary and secondary (parametric) feature extraction, and (iii) a flexible dynamic hierarchical BRB system. The schematic of the multi-level information fusion is presented in Fig. 1.

The first two modules prepare the initial WQP signals and extract features indicative of anomalies to be used as inputs to the hierarchical BRB system. Sensor data preprocessing module adaptively transforms the WQP sensor measurements and effectively reduces the effects of baseline variations and other data discrepancies. Then, primary and secondary feature extraction methods quantify significant deviation of preprocessed spatiotemporal data from regulatory patterns. The vectors of quantified features are mapped into water quality levels using the hierarchical BRB system. To aggregate the assessed quality levels in each analysis window a dynamic fuzzy evidential fusion algorithm is introduced. Extensive simulated water quality degradation events, originated from real dataset of the Quebec City’s main WDN and contamination tests in a pilot distribution facility, were exploited to validate the efficacy of the proposed framework when employed for event detection in the WDN.