Data Anomaly Detection in Wireless Sensor Networks with Application to an Oil Refinery

This paper presents a study concerning the online detection of outliers in non-stationary time-series collected over Wireless Sensor Networks on scenario implemented at an oil refinery. Two different approaches are under assessment. One based on the Least Squares Support Vector Machine and sliding window learning and standard Gaussian kernel. The other consists in a modification to the standard Gaussian kernel in order to improve the detection performance for non-stationary data streams. The implementability and effectiveness of these methods are evaluate on an oil refinery test-bed over a Wireless Sensor Network.