"The goal for my on-line NIR is to be an automated copy of the lab technician"

On-line spectroscopy, over a wide range of wavelengths and technologies, is in wide use in many important industries for good reasons. But the hit rate of very successful implementations is far from 100%—why? Lack of success may have several reasons, ranging from hardware issues (including sampling-related issues) to the much overlooked area of
long-term validation and maintenance of models to reflect the real process evolution. We here warn against results of inferior accuracy and precision being allowed to flow unquestioned into Business Intelligence (BI) systems or Manufacturing Execution Systems (MES). This all comes down to continuous validation of the online sensor(s) employed. Only through a proper validation, which initially outlines
adequate procedures and thresholds for future performance qualification, can contemporary QC data be used as a support for the business case for which the equipment was installed. When asked “How
accurate?” or “How much in agreement is this on-line system?” almost everybody will go for a scheme in which extracted samples are readily compared to lab results. Any resulting systematic difference is then
usually totally assigned as a RMSE sensor inaccuracy, this “bias” is immediately adjusted and on we go…. But often this apparently rational validation does not lead
to improvements—why? Such a bias correction is but a Christmas fairytale, however, and does not exist in the real world. There are some very important lacunae
involved, as we shall explain… The main myth here concerns too small a number of QC samples—conventional practice does not lead to fit-for-purpose validation.