Metal semiconductor sensors (MOS) would be ideal for ambient gas monitoring if not for their nonlinearity and unselectivity. Fortunately, the response of a multisensor array can be computationally deconvolved enabling selective gas measurement. The proposed research extends this technique to real time applications by using networks of connected processors, or neural networks. Phase I research will emulate multi-layer feed-forward neural networks, train them using back-propagation and use them to selectively measure ethyl alcohol, hydrogen, and methane with an array of MOS sensors. A goodness of fit output will also be incorporated and its efficacy for the detection of input channel faults and phantom analytes will be evaluated. Research results should be broadly applicable to other problems of nonlinear multicomponent spectral deconvolution. Markets for lower cost gas detectors are dominated by foreign manufacturers. By attacking the most significant obstacle to the development of a new generation of cost-effective analytical instruments, this research should result in greater domestic competitiveness in this important field.
