Dealing With Outliers in Wireless Sensor Networks: An Oil Refinery Application

Wireless sensor networks (WSNs) have become an important area of research because of their inherent characteristics, such as flexibility, low operational and maintenance costs, and scalability. When dealing with system monitoring in industrial environments, WSNs can be used for detecting and classifying transitory events or be integrated into networked control systems. As such, it is essential that the collected data is reliable, ensuring the quality of received information. A particular case of loss of reliability stems from outliers in raw data collected from the environment through built-in transducers or external transmitters attached to analog-to-digital converter ports. To avoid sending inaccurate data to the base station, it is required to implement a real-time data analysis to be launched at sensor nodes, which takes into account the nodes' natural computing and storage limitations. This brief proposes an outlier detection and accommodation methodology relying on univariate statistics in the form of Shewhart control charts, and formalized through a distributed hierarchical computational entities topology. The proposed scheme is evaluated on a real monitoring scenario implemented in a major oil refinery plant. Results from in situ experiments demonstrate the feasibility and relevance of the proposed approach.