Metrohm Vision Manual - Manual (Page 109)

▪▪▪▪▪▪▪ 107
If the library has a clustering method developed, Vision allows the creation of separate identification
methods for each leaf (lowest level) cluster. This is a completely local method, independent of
methods created for other clusters. This approach allows the reliable identification of similar
substances in a diverse library that includes a variety of products. It also allows for the inclusion of
groups of products with different physical and chemical properties within a single library.
Since the PC models are local, it is strongly recommended that each library product contain at least
30 sample spectra when identification by Mahalanobis distance or Residual Variance is attempted.
4.5.2 Cross-checking of Library Samples
The identification method development program checks for misidentified samples and conflicting
products when the method parameters are calculated. This in fact is a full validation of the method.
The results are displayed in the form of a 3-D histogram and in numerical format.
4.5.3 Identification Methods and Types of Thresholds
There are four identification methods available: Mahalanobis Distance in Principal Component space,
Residual Variance in Principal Component space, Maximum Distance in Wavelength space, and
Correlation in Wavelength space.
For Mahalanobis distance and Residual Variance methods it is possible to calculate the probability
that a given spectrum belongs to the distribution represented by the product spectra. Therefore, for
those methods, Vision offers two kinds of thresholds: match value (the actual number representing
distance or residual variance), and the probability level. For the two remaining methods, only the
match value type of threshold is possible.
Any combination of available math pre-treatments can be applied to the spectra with each of the
methods.
Mahalanobis Distance in Principal Component space
In this method, the local Principal Component model is calculated for each product in the library.
During the analysis, the unknown spectrum PC scores are calculated for each product model and
Mahalanobis distance is calculated. The unknown is identified as a product when the Mahalanobis
distance for this product is within the threshold value. The default threshold is 0.6 for match value
and 0.84 for probability level.
Residual Variance in Principal Component space
In this method, the local Principal Component model is calculated for each product in the library.
Each product Principal Component model is used to reconstruct the unknown spectrum. The
difference between original and reconstructed spectrum is used to calculate the residual variance.
The unknown is identified as a product when the residual variance for this product’s Principal
Component model is within the threshold value. The default threshold is 3 for match value and 0.84
for probability level.
Maximum Distance in Wavelength Space
Maximum distance belongs to the group of wavelength methods. For each product, the training set
of spectra is used to calculate the mean product spectrum and the inflated standard deviation
spectrum. During the analysis, the unknown spectrum is subtracted from the mean product spectrum
and divided by the standard deviation at each wavelength. The unknown is identified as a product