A neural network approach to fuel cell instrumentation for the detection of gas contaminants

The authors demonstrate that when artificial neural networks are used to process the output from a fuel cell, high levels of discrimination are possible. It is shown that contamination of ethanol with methanol can be detected (with 100% accuracy in the test set) down to levels of 5%. In addition the pure gases, ethanol, methanol and iso-propanol can be identified (again with 100% accuracy in the test set). When trained to recognise pure gases but tested on gas mixtures the outputs were generally not reliable in detecting the component gases in the mixture but were successful in identifying the fact that the gas was not pure. In the pure gas experiments reported it was found that frequency transformation prior to neural processing was advantageous. The use of curve fit polynomial coefficients as input data provides a compact and robust representation with computational complexity suitable for implementation in portable instrumentation