Agile Process Analytics
Application knowledge and chemometrics play a vital role in the processing of all types of multivariate data into application-specific information and has been doing so for at least 50 years. There has been a not-so-subtle shift in thinking as we integrate basic concepts and the occasional hallucination in the data mining, artificial intelligence, machine learning worlds. The target is to identify combinations of our technical tools to augment or replace tasks that consume brainpower where timely response is valued, and profits are at risk. The biggest focus of chemometrics has been in the calibration of optical spectrometers. It is worth considering the subtasks:
- Optimizing the instrument settings for a given application;
- Optimizing the method parameters – preprocessing, transformations, wavelength ranges;
- Handling of calibration transfer; and
- Optimizing models for inliers and rank in pursuit of routine processing and adjusting to changes in ingredients and unit operation.
The first two tasks are a set-once method development and the third may be generic across all applications. This paper tackles subtask 4 with a project that combined traditional approaches in statistics, database organization, pattern recognition, and chemometrics with some newer concepts tied to better understanding of data mining, neurocomputing, and machine learning. The future goal is to automate spectroscopy calibrations such that it is possible to have instrument systems tune themselves.