Auto news on Youtube Dec 16 2017

BWIQ is B&W Tek's chemometric software

designed for both on and off -line quantitative and qualitative spectroscopy measurements

Today we are going to demonstrate how to build a classification model with Raman spectral data

In this video, we will import the spectra into the BWIQ software

define the class for each sample

classify the spectra using a Principal Component Analysis-Mahalanobis Distance algorithm

and finally use the model to classify new samples

In classification models, spectra of samples corresponding to specific groups are collected

and then modeled according to their similarity to form a class

For this example, we will create a model using the Raman spectra of amino acids l-alanine

l-aspartic acid and l-cysteine hydrochloride

that were collected on B&W Tek's NanoRam handheld Raman system

20 spectra were collected for each amino acid

To import the spectra into the BWIQ software, click File - Import

and select the file format you wish to use

Locate the data set and import the files into BWIQ

The data files now appear in the Spectra window along with the corresponding spectra

When developing classification models, we need to define the class for each spectrum in the software

Click the blue addition button in the upper left corner of the Spectra window

a third column will appear

You can rename the column as "Class"

by right-clicking the header and selecting "Rename" from the dropdown menu

Here, we will enter an integer value for each distinct class

so that all aspartic acid spectra are designated class 1

all alanine spectra are class 2

and all l-cysteine hydrochloride spectra are class 3

There are several ways to designate spectra files as calibration or validation samples in BWIQ

Spectra files can be manually designated as calibration, validation, or ignored files

by clicking the drop-down button from the "Usage" column

Files designated "Ignored" will not be included in the final analysis

There are also several sampling algorithms available in the BWIQ software

which can be selected under "Sampling"

The parameters for the algorithms selected will appear in the algorithm properties panel

In the algorithm properties panel, you can change the ratio of calibration files to validation files

for instance, designating 60% of the files as calibration files, and 40% as validation files

To apply the sampling algorithm to the data

click the blue "execute" triangle in the algorithm properties window

the usage column in the Spectra window

will show which spectra are designated as calibration samples and which are validation samples

For this type of qualitative analysis, we designate all files as calibration files

Next, we can add pre-processing steps to remove variation in the data set

that is not related to chemical differences

but instead may result from scattering, instrumental variation, spectral noise, or background differences

The Raman data presented were collected using an automatic integration time feature

so the integration times for each spectrum are noticeably different

To normalize the data intensities, we will use a standard normal variate normalization algorithm

Click Pre-Process, then Standard Normal Variate

The normalization now appears in the algorithms window

When building classification models, an important step is to mean center the data

as we are interested in the difference of the data from a centered point

not how far the samples are from that center

Click "Pre-processing" then "Center"

The step is now added to the algorithms window

Manual variable selection enables us to use the entire spectrum for analysis

or alternatively to restrict the analysis to selected regions of the data

Using the whole spectrum for analysis will allow the model to be more sensitive to contaminants

or changes in the samples that introduce signal in other spectral regions

However, it can be helpful to remove non-informative or noisy regions of the data from the analysis

For example, in this data we can see that above 1800 wave numbers

there is little Raman signal

Under Variable, choose Manual Selection

A Manual Variables selection is automatically added to the selected algorithms window

Now that we have finished adding pre-processing steps

we can add a classification algorithm

For this data, we will select a Principal Component Analysis Mahalanobis Distance classification method

Principal Component Analysis, or PCA

is an excellent exploratory chemometric tool

that uses a reduced variable space to define the greatest variance within a data set

With PCA, model scores and loadings are computed

and the 2D scores plot provides a visualization of natural groupings of samples

based on the similarity of the data

With each sample represented by a single point in the new Principal Component space

To classify new samples with the model, we'll use Principal Component Analysis

in combination with the parameter known as the Mahalanobis Distance

which measures the distance of a new sample to the center of each class

The Mahalanobis Distance for each new sample is calculated for every specified class

The sample is classified by the software as the group that corresponds to the shortest calculated Mahalanobis Distance

Choose Classification, and then PCA-MD

Click the blue addition button in the Algorithm Properties window

to add the PCA-MD algorithm to the model

The algorithms that are listed in the selected algorithms window

now make up the steps of our classification model

To save this training file with the extension .train

Go to File, then Save As, and save the file

Training files can be edited and saved at any time in Method Development

To execute the model, click the blue triangle at the top of the window

This will perform the steps in sequence

Because our model includes a Manual Variable Selection

The manual variables window appears upon execution of the model

In this window, we can type in the the spectral regions of interest

or use the cursor to define the analysis regions

Because there is no significant Raman signal above 1800 wave numbers

for any of these amino acids in this data set

we will limit our analysis from 200-1800 wave numbers in these spectra

The data shown have been pre-processed and are now mean centered

with peaks in both directions around the zero line

Upon execution of the method, a new model file is created

The model file is a .cmml file that can be saved and used to classify new samples on-line

while connected to a B&W Tek i-Raman series instrument

or offline using previously collected data

To view the scores plots, click the Scores icon in the toolbar

The scores plot shows the clusters of the samples in principal component space

In this plot, we observe three distinct, well separated clusters

These three clusters correspond to the three different classes of amino acids

Based on how close they fall to these three clusters

we can classify new spectra as belonging to one of these groups

There are several other plots available to view data

including loadings and variance

To classify unknown samples from previously acquired data

click Predict, and load the prediction files in the correct format

The prediction results list the individual spectral file names

along with the calculated Mahalanobis distances from each class in the model

The column to the right labeled Class shows the final classification for the sample

In this case, either one, two, or three

The class showing the lowest calculated Mahalanobis distance

is the class chosen by the BWIQ software to represent that sample

Under Generate PDF Report, you can also save a PDF report

which details the method parameters and the classification predictions

For more information on using the BWIQ software, please refer to the user manual

as well as our other support videos

You can also reach us at www.bwtek.com/support

Thank you for watching!