Detection and accommodation of outliers in Wireless Sensor Networks within a multi-agent framework

Graphical abstractDisplay Omitted HighlightsInvestigated three techniques for outliers detection in Wireless Sensor Networks.A Least Squares-Support Vector Machine-based technique with a sliding window-based learning.Principal Component Analysis method along with subspace tracking with rank-1 modification.Univariate statistics-based scheme within an oversampling environment.All methods are implemented within a hierarchical multi-agent framework. This paper studies three techniques for outliers detection in the context of Wireless Sensor Networks, including a machine learning technique, a Principal Component Analysis-based methodology and an univariate statistics-based approach. The first methodology is based on a Least Squares-Support Vector Machine technique, together with a sliding window learning. A modification to this approach is also considered in order to improve its performance in non-stationary time-series. The second methodology relies on Principal Component Analysis, along with the robust orthonormal projection approximation subspace tracking with rank-1 modification, while the last approach is based on univariate statistics within an oversampling mechanism. All methods are implemented under a hierarchical multi-agent framework and compared through experiments carried out on a test-bed.