288
B. Igne et al.
• Preprocessing: used to reduce irrelevant variance, enhance relevant variance
and/or linearize the relationship between the spectral data and the parameter of
interest. Preprocessing options should be selected to address the particular needs
of the model based on the sample matrix. For instance, if a difference in particle
size exists across the samples, varying scattering intensities could be mitigated
by standard normal variate [17] or multiplicative scatter correction [18].
• Model complexity: over- or under-determined models can significantly impact
robustness, and the model complexity should be set so that the model is able to
handle future variability while still delivering adequate accuracy.
12.3.5.4 Sources of Method Errors
The final method error is the combination of several sources of error that not only
have the potential to impact the model accuracy but also the measurement suitability.
Sources of error to consider include the sampling error, the error of the laboratory,
the error due to instrument variability, etc.
Ensuring the measurement is representative of the process being analyzed is critical to the quality of the method outputs. There are several items to consider: the scale
of scrutiny (the volume of sample analyzed by the instrument), the spectral collection frequency, the sampling method, and the process variability. If a spectrometer
analyzes the entire sample volume but the sample is collected in a way that does
not represent the variability of the process, the model error may be low (meeting
the predefined criteria for figures of merit) but method error would be high. If the
sampling is unbiased, the measurement volume is appropriate, but the frequency of
measurement is low, the manufacturing process may not be appropriately monitored.
Ensuring that the sample measured is representative of the process is critical to a
method success.
As a secondary analytical method, NIRS relies on the determination of the parameter(s) of interest by a primary method (i.e., scale, chromatography and spectroscopy,
etc.) for model building. The larger the error in the reference methodology (error of
the laboratory), the larger the error of the NIRS model.
Another element related to the error of the laboratory is the quantity being
measured as reference values (or the unit in which it is expressed) and its relationship
with what NIRS measures. In a publication, the impact of selecting the unit of the
parameter of interest (volume fraction vs. weight fraction) was shown to result in
nonlinearities with the authors commenting on the fact that the sensitivity of NIRS to
volume fraction over weight is distorting the established concept of artificial design of
experiments based on weight [19]. The work was done on liquid samples but should
apply to other forms of samples. Authors argue that the nonlinearity generated by
using the “wrong” unit cannot be fully accommodated for by spectral preprocessing
methods.
Instrumental error should also not be under-estimated. While modern instruments
are highly reproducible with nearly noise-free detectors, variability in lamp intensity
can be observed as a function of time and care should be taken to ensure a model
B. Igne et al.
• Preprocessing: used to reduce irrelevant variance, enhance relevant variance
and/or linearize the relationship between the spectral data and the parameter of
interest. Preprocessing options should be selected to address the particular needs
of the model based on the sample matrix. For instance, if a difference in particle
size exists across the samples, varying scattering intensities could be mitigated
by standard normal variate [17] or multiplicative scatter correction [18].
• Model complexity: over- or under-determined models can significantly impact
robustness, and the model complexity should be set so that the model is able to
handle future variability while still delivering adequate accuracy.
12.3.5.4 Sources of Method Errors
The final method error is the combination of several sources of error that not only
have the potential to impact the model accuracy but also the measurement suitability.
Sources of error to consider include the sampling error, the error of the laboratory,
the error due to instrument variability, etc.
Ensuring the measurement is representative of the process being analyzed is critical to the quality of the method outputs. There are several items to consider: the scale
of scrutiny (the volume of sample analyzed by the instrument), the spectral collection frequency, the sampling method, and the process variability. If a spectrometer
analyzes the entire sample volume but the sample is collected in a way that does
not represent the variability of the process, the model error may be low (meeting
the predefined criteria for figures of merit) but method error would be high. If the
sampling is unbiased, the measurement volume is appropriate, but the frequency of
measurement is low, the manufacturing process may not be appropriately monitored.
Ensuring that the sample measured is representative of the process is critical to a
method success.
As a secondary analytical method, NIRS relies on the determination of the parameter(s) of interest by a primary method (i.e., scale, chromatography and spectroscopy,
etc.) for model building. The larger the error in the reference methodology (error of
the laboratory), the larger the error of the NIRS model.
Another element related to the error of the laboratory is the quantity being
measured as reference values (or the unit in which it is expressed) and its relationship
with what NIRS measures. In a publication, the impact of selecting the unit of the
parameter of interest (volume fraction vs. weight fraction) was shown to result in
nonlinearities with the authors commenting on the fact that the sensitivity of NIRS to
volume fraction over weight is distorting the established concept of artificial design of
experiments based on weight [19]. The work was done on liquid samples but should
apply to other forms of samples. Authors argue that the nonlinearity generated by
using the “wrong” unit cannot be fully accommodated for by spectral preprocessing
methods.
Instrumental error should also not be under-estimated. While modern instruments
are highly reproducible with nearly noise-free detectors, variability in lamp intensity
can be observed as a function of time and care should be taken to ensure a model