7 NIR Data Exploration and Regression by Chemometrics—A Primer
185
7.8.2 Outro
The idea of using ASCA to partition the variances of the experimental design parameters has great potential and has not been fully exploited in NIRS literature. ASCA
brings a valuable link between the multivariate data analysis and statistics, as it is able
to provide significance testing of the different design factor effects and interactions.
7.9 Process Analytical Technology, Machine Learning
and Other NIRS Trends
Collecting large quantities of extremely low-quality data will not be the
recipe for success!
—Tom Fearn, British chemometrician
Due to its unique capabilities and complex, holistic spectra, NIR spectroscopy has
served as the perfect playground for the development of multivariate data analysis
and chemometrics, and there is no sign for this to stop in the near future. The majority
of spectroscopic sensors used in process analytical technology (PAT) are based on
NIR technology [57, 58]. This may be called “the second green analytical revolution
of NIRS analysis.” The first was introduced by Williams and Norris in 1975 when
replacing sulfuric acid demanding Kjeldahl analysis with clean NIRS analysis [59].
NIRS analysis in the PAT context has perhaps a much larger potential to change the
way that we produce sustainably and the way that we optimize processes for a new
circular and green economy (see Fig. 7.35). Moreover, portable NIRS sensors are
omnipresent in agriculture, we are beginning to see NIRS sensors on drones, NIRS
hyperspectral imaging and NIRS sensors on mobile phones are emerging. This will
drastically increase the amount of NIR data collected from practically all aspects of
life.
This technology revolution has created a strong quest for new and more efficient
data analytical tools. Artificial intelligence, machine learning and deep learning are
increasingly exploited to facilitate efficient information extraction from the enormous data collections, and applications are starting to emerge for the spectroscopic
disciplines [60]. These methods can deal with nonlinear effects (abundant in NIR
spectroscopy), but are generally less interpretable (black box) for the scientist.
The developments of machine learning methods are primarily made in the NIR
imaging field, where neural networks have proven quite efficient in decoding the
complexity of hyperspectral images in their original multidimensional form [61], a
natural extension to the more traditional approach of applying PCA on deconvoluted
images.
The true application of “deep” neural processing, where the neural network is fed
raw sensor data and trained to form self-organizing feature detectors, is an obvious
185
7.8.2 Outro
The idea of using ASCA to partition the variances of the experimental design parameters has great potential and has not been fully exploited in NIRS literature. ASCA
brings a valuable link between the multivariate data analysis and statistics, as it is able
to provide significance testing of the different design factor effects and interactions.
7.9 Process Analytical Technology, Machine Learning
and Other NIRS Trends
Collecting large quantities of extremely low-quality data will not be the
recipe for success!
—Tom Fearn, British chemometrician
Due to its unique capabilities and complex, holistic spectra, NIR spectroscopy has
served as the perfect playground for the development of multivariate data analysis
and chemometrics, and there is no sign for this to stop in the near future. The majority
of spectroscopic sensors used in process analytical technology (PAT) are based on
NIR technology [57, 58]. This may be called “the second green analytical revolution
of NIRS analysis.” The first was introduced by Williams and Norris in 1975 when
replacing sulfuric acid demanding Kjeldahl analysis with clean NIRS analysis [59].
NIRS analysis in the PAT context has perhaps a much larger potential to change the
way that we produce sustainably and the way that we optimize processes for a new
circular and green economy (see Fig. 7.35). Moreover, portable NIRS sensors are
omnipresent in agriculture, we are beginning to see NIRS sensors on drones, NIRS
hyperspectral imaging and NIRS sensors on mobile phones are emerging. This will
drastically increase the amount of NIR data collected from practically all aspects of
life.
This technology revolution has created a strong quest for new and more efficient
data analytical tools. Artificial intelligence, machine learning and deep learning are
increasingly exploited to facilitate efficient information extraction from the enormous data collections, and applications are starting to emerge for the spectroscopic
disciplines [60]. These methods can deal with nonlinear effects (abundant in NIR
spectroscopy), but are generally less interpretable (black box) for the scientist.
The developments of machine learning methods are primarily made in the NIR
imaging field, where neural networks have proven quite efficient in decoding the
complexity of hyperspectral images in their original multidimensional form [61], a
natural extension to the more traditional approach of applying PCA on deconvoluted
images.
The true application of “deep” neural processing, where the neural network is fed
raw sensor data and trained to form self-organizing feature detectors, is an obvious