Efficient root cause localization in IoT-enabled water distribution networks by hierarchical anomaly analysis

Distribution networks such as water, power and gas networks are usually subject to failures in the network elements, bad design or even abuse. Broken pipes or junctions, faulty pumps or valves and malfunctioning reservoirs in Water Distribution Networks (WDN) are examples of failures that cause a lot of resource losses and adversely affect the quality of service. In order to detect and fix these kinds of incidents in a timely manner, distribution network elements are equipped with sensors and IoT devices. Due to the massive amounts of measured data by sensors, efficient de-centralized analysis methods over Fog Computing Architecture (FCA) have been proposed recently to detect and localize anomalous behaviors in the network elements. Although these methods are useful in reporting global anomalies efficiently, they are unable to detect the real causes of anomalies in many scenarios accurately. To overcome this deficiency, in this paper a novel Hierarchical Anomaly Analysis (HAA) model with a majority-based spatial anomaly detection technique based on an unsupervised clustering algorithm is presented to accurately localize the actual network element that caused the abnormal behavior in a de-centralized manner. The presented HAA, which is based on the concept of Cause Detection Hierarchy (CDH) over FCA, was able to outperform the previous FCA-based root cause detection approach, achieving an average increase of 32% in the detection F-score across three datasets. To minimize the number of sub-model transmissions in CDH and also to tolerate the failures of CDH agents, a novel entropy-aware publishing mechanism based on K-Nearest Neighbor (KNN) method is presented that prevents the CDH agents from transmitting predictable sub-models to upper layers. It was observed that by choosing a proper value for the Model Quantizing Factor (QF), up to 52% reduction in the number of published messages is possible.