A dependability framework for WSN-based aquatic monitoring systems

Abstract

Wireless Sensor Networks (WSN) are being progressively used in several application areas, particularly to collect data and monitor physical processes. Moreover, sensor nodes used in environmental monitoring applications, such as the aquatic sensor networks, are often subject to harsh environmental conditions while monitoring complex phenomena. Non-functional requirements, like reliability, security or availability, are increasingly important and must be accounted for in the application development. For that purpose, there is a large body of knowledge on dependability techniques for distributed systems, which provides a good basis to understand how to satisfy these non-functional requirements of WSN-based monitoring applications. Given the data-centric nature of monitoring applications, it is of particular importance to ensure that data is reliable or, more generically, that it has the necessary quality. The problem of ensuring the desired quality of data for dependable monitoring using WSNs is studied herein. With a dependability-oriented perspective, it is reviewed the possible impairments to dependability and the prominent existing solutions to solve or mitigate these impairments. Despite the variety of components that may form a WSN-based monitoring system, it is given particular attention to understanding which faults can affect sensors, how they can affect the quality of the information, and how this quality can be improved and quantified. Open research issues for the specific case of aquatic monitoring applications are also discussed. One of the challenges in achieving a dependable system behavior is to overcome the external disturbances affecting sensor measurements and detect the failure patterns in sensor data. This is a particular problem in environmental monitoring, due to the difficulty in distinguishing a faulty behavior from the representation of a natural phenomenon. Existing solutions for failure detection assume that physical processes can be accurately modeled, or that there are large deviations that may be detected using coarse techniques, or more commonly that it is a high-density sensor network with value redundant sensors. This thesis aims at defining a new methodology for dependable data quality in environmental monitoring systems, aiming to detect faulty measurements and increase the sensors data quality. The framework of the methodology is overviewed through a generically applicable design, which can be employed to any environment sensor network dataset. The methodology is evaluated in various datasets of different WSNs, where it is used machine learning to model each sensor behavior, exploiting the existence of correlated data provided by neighbor sensors. It is intended to explore the data fusion strategies in order to effectively detect potential failures for each sensor and, simultaneously, distinguish truly abnormal measurements from deviations due to natural phenomena. This is accomplished with the successful application of the methodology to detect and correct outliers, offset and drifting failures in real monitoring networks datasets. In the future, the methodology can be applied to optimize the data quality control processes of new and already operating monitoring networks, and assist in the networks maintenance operations.