Detection and accommodation of outliers are crucial in a number of contexts, in which collected data from a given environment is subsequently used for assessing its running conditions or for data-based decision-making. Although a significant number of studies on this subject can be found in literature, a comprehensive empirical assessment in the context of local online detection in wireless sensor networks is still missing. The present work aims at filling this gap by offering an empirical evaluation of two state-of-the-art online detection methods. The first methodology is based on a Least Squares-Support Vector Machine technique, along with a sliding window-based learning algorithm, while the second approach relies on Principal Component Analysis and on the robust orthonormal projection approximation subspace tracking with rank-1 modification. The performance and implementability of these methods are evaluated using a generated non-stationary time-series and a test-bed consisting of a benchmark three-tank system and a wireless sensor network, where deployed algorithms are implemented under a multi-agent framework.
- Gaussian kernel
- Least-Squares Support Vector Machine
- Online implementation
- Outliers detection
- PCA and subspace tracking