Artificial neural networks and statistical pattern recognition improve MOSFET gas sensor array calibration

It is noted that the poor selectivity of many gas sensors is disadvantageous when individual gases are studied in gas mixtures or when odors are identified. It has been shown that pattern recognition methods are very promising when gases or odors are identified by means of gas sensor arrays. The quality of predictive models, based on partial least square (PLS), nonlinear PLS, and artificial neural networks, has been studied. The authors have chosen a six-element gas sensor array containing Pd- and Pt-gate MOSFETs operating at elevated temperatures. The experiments show that hydrogen and ammonia concentrations can be well predicted in the presence of two other interfering gases, and in a three-component mixture without ammonia, hydrogen is well predicted. Ethylene and ethanol can be predicted as the sum of their concentrations.<>