Sensor network data denoising via recursive graph median filters

Abstract

In wireless sensor networks (WSN) for environmental monitoring, the sensor nodes typically have limited computational and communication resources. Furthermore, the sensors often operate in a harsh environment, and the measurements can be subjected to significant levels of noise. In this work, with these considerations in mind, we propose an efficient method to denoise the sensor data. Using concepts from graph signal processing, the WSN is first modelled using an extended graph. The time-vertex graph jointly models the correlations between neighbouring sensor nodes and across the time dimension. A recursive graph median filter is developed that can be highly localized, and can be implemented with distributed processing. The filter is applied to the denoising of data that is subjected, simultaneously, to Gaussian noise and impulsive noise. Extensive experimental results, using three real-world measurement datasets, will demonstrate that the recursive filter significantly outperforms the equivalent linear filter and nonrecursive median filter, at high noise levels.