Gas detection and source localization: A Bayesian approach

This paper discusses modeling solutions that support detection of gaseous chemical substances and source localization in applications that are characterized by large numbers of noisy information sources, absence of calibrated concentration measurements and lack of detailed knowledge about the physical processes. In particular, we introduce a solution based on discrete Bayesian networks which allows tractable exploitation of large quantities of spatio-temporally distributed heterogeneous observations. The emphasis is on using coarse models avoiding assumptions about detailed aspects of the gas propagation processes. By considering properties of Bayesian networks we discuss the consequences of modeling simplifications and show with the help of simulations that the resulting inference processes are robust with respect to the modeling deviations.