Efficient Methods to Detect Atmospheric Concentration with Low Signal to Noise Ratio on a Sensor Network

Accidental or malicious releases of materials into the air can occur surreptitiously and lead to low concentrations involving very limited changes in the atmospheric background noise. This may result in signal variations on sensors placed in the environment characterized by a low signal-to-noise ratio (SNR). In this work, we aim to provide a selection of efficient online detection methods adapted to low signal to noise ratio (SNR). These methods take into consideration signals that are monitored by a subset of a network of sensors and for a limited period of time. We show that it is possible to detect with a very reasonable false alarm rate only knowing the statistics of the background noise. Derived from the Cumulative Sum (CUSUM), the methods presented and tested here are generalizations to multivariate time series considering space and time sparsity of the exposure of the network of sensors. We show that the methods previously presented in a Gaussian context can easily be extended to a Poisson model which is consistent with particle counting sensors like those used in the frame of radioactivity detection. We demonstrate the validity of the different methods on toy example and then prove its interest on data from the twin experiment of a fictitious radioactive release in an urban area.

Automated summary

This study aims to provide a selection of efficient online detection methods adapted to low signal-to-noise ratio (SNR). It shows that, by considering the space and time sparsity of multivariate time series, it is possible to detect the releases of low concentrations involving very limited changes in the atmospheric background noise with a reasonable false alarm rate.