Different strategies for the identification of gas sensing systems

We have shown that it is possible to extract dynamical models for non-linear gas sensors from experimental input-output data. Seven different methods (two linear and five non-linear) have been evaluated in terms of their prediction performance on the sensor response to white gaussian inputs. Two methods, artificial neural networks and modified Wiener kernels estimated by least squares, show very low prediction errors with models extracted from only 300 input-output data pairs. A detailed discussion on the advantages and disadvantages of every method is presented.