9.3.4.   Conclusions

Ternary mixtures of methanol, ethanol and 1-propanol measured by the RIfS array setup and by the 4l setup could be successfully quantified. The application of the growing neural network framework instead of the non-optimized neural networks resulted in significantly improved calibrations again. The variable selection of the framework for the array setup is quite astonishing, since only the sensor signals of 2 sensors out of 4 sensors are used. As the framework selects the most predictive variables, it can be concluded that the time domain of 2 sensors contains more information than the parallel (static) information of all 4 sensors together. This is an impressive demonstration, how the time-resolved measurements of few sensors can render the application of many parallel sensors with different sensitivities redundant. The static evaluation of the 4 sensors, which corresponds to the sensor signals at the end of exposure to analytes, shows that the time-resolved evaluation of the sensor signals is highly superior even though the static sensor evaluation is not a mathematically underdetermined system. It was demonstrated that smoothing improves the calibration of measurements performed by thin sensitive layers whereas the calibration deteriorates for measurements performed by thick sensitive layers when smoothing the sensor signals. Furthermore, the 4l setup can be used as single sensor device for a multicomponent quantification. Compared with the array setup, the price of miniaturization and cost reduction has to be paid in terms of extended measurement times respectively higher prediction errors.