CAMO Software
Home > Page


Octane number measurements

UOP Guided Wave Inc. made a feasibility study for one of Guided Wave'spectrochemical customers, where the monitoring of octane numbers in gasoline production is required for quality control. The reference method for octane number measurements is time consuming and expensive, involving extensive use of test engines, etc. If it is possible to replace such measurements with fast, inexpensive NIR-spectroscopy measurements, routine quality control could be effectively rationalized.

Twenty-six production gasoline samples were collected over a sufficient period of time, considered to span all the most important variations in the production variations. Two of the samples contained added alcohol, which increases the octane number. NIR absorbance spectra over 226 wavelengths, and the corresponding reference measurements of the octane number were recorded. A PLS model was made to predict the octane number from the spectra.

Score plot
Fig. 7 The score plot shows sample patterns. Here sample M52 and H59 are far away from all the others, indicating large dissimilarity.

The Unscrambler indicated two samples as extreme outliers. This can also be seen in the score plot, fig 7, where two samples contribute more to the model along the first PC (the horizontal axis) than all the other samples.

The plot of the prediction error , fig 8, displayed an increase, which is a clear sign that something is wrong - usually the presence of outliers which disturb the model.

Residual variance plot
Fig. 8 The prediction error should ideally decrease with increasing number of PLS-components. An increase indicates problems.

The loadings for PC1 were largest in band 1400-1420 nm indicating absorption due to alcohol, see fig 9. Obviously the model describes the variation in the data set due to alcohol. Since we are most interested in modeling octane number in general, we decide to remove the two outliers. Objects 25-26 were the samples with added alcohol; obviously they are so dissimilar to the others that we can't make a model for both types.

Loadings plot
Fig. 9 Large loadings indicate variables (here wavelengths) which are important in the model.

A new PLS model was made. Now the prediction error decreases as we add more PLS-components, see fig 10.

Residual Variance
Fig 10 The prediction error is low already for a model with only 2 PCs. 99% of the variance of Y is explained by 2 PCs.

The correlation between predicted and measured Y is 0.993, the offset close to 0, and the regression line close to 1, see fig 11. This is a very good model, with an average prediction error at validation of 0.24. Comparing this with octane numbers between 87 and 93 gives an average relative error of less than 0.5%.

Predicted vs. Measured
Fig.11 Good correlation between measured and predicted Y-values.

Since the feasibility study gave such promising results, the calibration model was extended with more samples, and validated further. It is now implemented for on-line prediction in a quality monitoring system for gasoline production.