Distributed Detection and Localization

Abstract

This thesis delves into the detection and localization aspects of distributed Wireless Sensor Networks (WSNs). Specifically, the research concentrates on WSNs in which sensors autonomously carry out detection tasks and transmit their decisions to a fusion center (FC). The FC’s role is to make a comprehensive decision about the presence of a specific event of interest and estimate its potential location. Given its broad significance, the thesis specializes in applying WSNs for industrial monitoring, particularly in the process and energy industry. Three distinct approaches are explored in this thesis: (i) per-sample/batch detection, (ii) quickest detection, and (iii) sequential detection. Each framework proposes a set of detection and associated localization rules. A primary objective of this work is to develop detection and localization strategies that leverage existing information about the monitored environment, bridging the gap between monitoring systems and the knowledge of the monitored system. In the proposed per-sample/batch detection approach, sensors make localized binary decisions about the presence of an adverse event. The FC aggregates these decisions to provide a more reliable global binary decision. A comparative analysis is conducted between the counting rule and the newly proposed modified Chair-Varshney rule. Threshold design is facilitated through the maximization of Youden’s Index. Upon detection, the FC offers an estimated position using four investigated localization algorithms: maximum a-posteriori localization, minimum mean square error localization, the centroid-based algorithm, and the newly proposed modified centroid-based algorithm. Regarding the quickest detection approach, two architectures are introduced that capitalize on diverse network structures for quickly detecting and pinpointing faults in industrial plants. Both incorporate a feedback mechanism transmitting parameters from higher to lower hierarchical levels. The first architecture, named three-layer architecture, involves multiple sensors overseeing a specific plant section, each independently conveying its local decisions to the FC. The FC gathers these local decisions spatially to generate a more comprehensive decision. Subsequently, a post-processing center (PPC) analyzes these global decisions over time, executing prompt detection and localization. In the second structure, named two-layer architecture, the FC engages in spatio-temporal aggregation to achieve swift detection, along with a potential estimation of the faulty component. The sequential detection approach proposes a WSN where sensors perform local sequential detection and transmit decisions to the FC. Two variations of this algorithm, the continuous sampling algorithm (CSA) and the decision-triggered sampling algorithm (DTSA), each with a unique transmission and detection rule, are introduced. In the DTSA, the FC processes only those transmissions corresponding to local decisions. The CSA instead functions as a time-aware detector by processing every sensor’s transmission, integrating the time of each transmission into the detection rule. The per-sample/batch and quickest detection approaches and their localization methods are tested in an oil and gas scenario involving an oil spill within subsea production systems. Conversely, an industrial facility’s carbon dioxide dispersion scenario tests the sequential detection approach. System performance evaluation encompasses the receiver operating characteristic curve, decision delay, localization error, computational complexity, and communication costs.