FEASIBILITY OF SURFACE-ACOUSTIC-WAVE (SAW) SENSOR ARRAY-PROCESSING WITH FORMAL NEURAL NETWORKS

The poor selectivity of sensors, used to detect different species mixe d in a real gas mixture, requires efficient signal processing devices to improve it. We have tried to apply neural techniques to this proble m in order to obtain some 'electronic' selectivity. This work is the f irst step of a feasibility study of an electronic device using SAW gas sensors. Using linear approximation of experimental and static result s of SAW sensors given in previous data, we also attempted to apply ne ural networks for multisensor array signal processing. In order to ext end the operating range, the saturation effects yield by chemical laye rs was considered. To optimize the neural network, several activation functions were tested. A design close to radial basis functions networ ks was successfully applied. A new network connectivity led to increas ed interpolation capability. For real time processing, it is important to escape from slow kinetic adsorption of chemical layers. In this pa per, the theoretical possibility of response time compensation has bee n shown by using a deconvolution process.