9.5.   Quantification of the Refrigerants R22 and R134a in Mixtures: Part II

This section is an example how the calibration of old data sets can be improved by a time-resolved data analysis. The data set, which was recorded by the RIfS array with different polymers, was already introduced and investigated in chapter 3. During the development of the time-resolved measurements, this data set was investigated once more, since the sensor response of the Makrolon layer had also been recorded in a time-resolved mode (see section 4.5.1.2).

The parallel growing neural network framework (250 parallel runs for each analyte) was applied to the data set with all 23 variables. The framework selected 6 variables for an optimal prediction. Among these variables the 3 static sensor responses of HBP, PDMS and UE 2010 15% and additionally 3 time-resolved variables of Makrolon were selected. Using these variables instead of the 6 static sensor responses the test data were significantly better predicted with an improvement of 33% for the RMSE of R22 and 9% for the RMSE of R134a (see table 9). The parallel framework was also applied to the calibration data using only the time-resolved sensor signals of Makrolon. The predictions of the optimized neural nets with 18 variables are shown in table 9. Compared with the static sensor signals of 6 sensors, the predictions using the time-resolved sensor signals of the Makrolon sensor were by 12% better for R22 and by 68% worse for R134a. When comparing the mean error of the test data of this single sensor approach with the 15 combinations of the 2-sensor approach (see table 1 in section 3.4) it is astonishing that the time-resolved single sensor approach is better than most of the 2-sensor approaches (6th position).

Variables

Test R22

Test R134a

Test Mean

6 Static

0.00183

0.00630

0.00406

3 Static + 3 Time-resolved

0.00141

0.00575

0.00358

18 Time-resolved

0.00161

0.01061

0.00611

table 9:      RMSE of the time-resolved data analysis compared with the static data analysis performed in chapter 3, see also table 1.

As a conclusion of this section it might be said that even for data sets, which were recorded without the intention of a time-resolved data analysis and without any optimization of the measurement parameters, the time-resolved data analysis can improve the calibration.