Updating to the latest model

NIR spectroscopy is increasingly being used for on-the-spot analysis of industrial processes, allowing critical aspects of the production process to be monitored continuously and altered if necessary. For example, NIR spectroscopy is being used in diesel production to ensure that certain quality parameters, such as flash point and cloud point, are always at the correct values.

This involves building a model for predicting the parameter values from spectral data and then feeding this model with the data generated by monitoring diesel production with NIR spectroscopy. If the predicted values start to deviate from where they should be, then that is a strong indication there is a problem with the production process.

But only if the predictive model is still accurate and that may not be the case. As time goes by, there will be various unavoidable changes in the production process: for example, the composition of the diesel may change slightly or there may changes in temperature and flow rate. On top of that, the responses of the NIR spectrometer may change over time or it may even be replaced, with the new version producing slightly different responses.

These small unavoidable changes may not affect the values of the quality parameters, but they may well affect the accuracy of the predictive model, causing it to report that the values are deviating even if they're not. Ideally, what is needed is a predictive model that can update itself to take account of these unavoidable changes and that is what a team of Canadian chemical engineers from the University of Alberta and Suncor Energy have now succeeded in developing.

Using the latest spectral data, this model continuously updates itself by altering the wavelengths of infrared radiation that it bases its predictions on and by altering the coefficients utilised by the model. Of these, altering the wavelengths is most important, because any unavoidable changes in the production process will change which specific wavelengths provide most information about the quality parameters. By altering the wavelengths, the model can always focus on the most informative wavelengths for predicting the quality parameters and ignore those that no longer provide useful information.

As they report in Industrial & Engineering Chemistry Research, the chemical engineers tested this modelling approach on the diesel produced at a Canadian refinery and found that the resultant model could update itself to take account of unavoidable changes in the refining process. As a result, its predictions for the values of flash point and cloud point were much more accurate than those produced by models that didn't update.

Blog tags: