Outlier Detection from Non-Smooth Sensor Data

Timo Huuhtanen, Henrik Ambos, Alexander Jung

Abstract: Outlier detection is usually based on smooth assumption of the data. Most existing approaches for outlier detection from spatial sensor data assume the data to be a smooth function of the location. Spatial discontinuities in the data, such as arising from shadows in photovoltaic (PV) systems, may cause outlier detection methods based on the spatial smoothness assumption to fail. In this paper, we propose novel approaches for outlier detection of non-smooth spatial data. The methods are evaluated by numerical experiments involving PV panel measurements as well as synthetic data.