Airborne contaminant urban localization by the hybrid stochastic artificial neural network system.

One of the tasks of early response groups to threats, especially in urban areas, is to quickly trace the source of the danger and take appropriate steps to prevent its effects. For example, in the case of atmospheric contamination, the first thread indicators are the concentrations of the dangerous substance reported by a network of sensors in a given area. Thus, the dedicated emergency system should localize the source of contamination based on the data from the sensors. Moreover, such a system should work in realtime. The paper presents the reconstruction system pointing to the most probable source of the airborne toxin in the urban terrain in real-time. The method is based on the stochastic event reconstruction approach using the Approximate Bayesian Computation (ABC) methodology. The sequential ABC scanning algorithm dynamically updates the posterior distributions of the searched parameters by the most up-to-date reported concentrations. The operation of such a system in real-time is possible thanks to the use of the surrogate model in place of the computationally expensive urban dispersion model. The surrogate model exploits the feedforward neural network (FFNN) characterizing with the fast reply. The process of training dedicated FFNN and methods of FFNN model verification for the system dedicated to the given urbanized area is presented. Training of the artificial neural network requires an extensive, reliable set of contaminant dispersion data, impossible to obtain from actual field experiments. Thus, the training dataset was generated using the Quick Urban Industrial Complex Dispersion Modeling system (QUIC), developed by Los Alamos National Laboratory. Finally, the proposed reconstruction system's efficiency is tested using synthetic releases and actual data from a full-scale field experiment DAPPLE conducted in central London in 2007.