Multivariate statistics play an important role in all forms of process analysis, from the design of experiments to the analysis of data. Research into multivariate statistics is becoming increasingly crucial to the success and validation of process systems. Multivariate statistics within the Marquardt Group are focused towards extracting as much relevant information from a system as possible. The systems we examine varying wildly from the tip of a sensor to crude oil refining. The varying projects lead to a diverse research direction including data fusion, calibration transfer, preprocessing and variable selection and image analysis.

Data Fusion
Data fusion is the process of molding together multiple sources of information such that the newly fused data set better describes the process system than the individual sources. This fits perfectly with our overall goal of extracting as much information from a process as possible. Research into data fusion is currently being driven by processes within petrochemical refining; we are currently examining and formulating the best methods to perform data fusion along with expanding the concept into areas such as pharmaceuticals and biofuel processing. Any process that records multiple measurements of a single sample is an ideal candidate for data fusion.

Preprocessing and Variable Selection
Typical process samples vary in physical and chemical properties; this can lead to a number of unwanted attributes being recorded in the spectra. We have been developing a series of tools that allows a user to remove this unwanted information leaving behind only the information that is relevant to the process. Our area of expertise involves the correction of Raman spectra for changes in the background fluorescence and cosmic rays.

Another way of overcoming unwanted variation in a set of data is to remove the problem variables. There are many algorithms and procedures for doing this, we have been advancing and improving these methods through application driven projects for building prediction models. In particular we focus on the application of genetic algorithms and forward selection backward elimination iterative routines. These methods select only the variables that pertain to the information of interest and remove the unwanted variation.

Image Analysis
Research in other parts of the group on luminescent sensors has created a need for a method to quantify the coverage of vapochromic materials on the tip of a sensor. Pictures of each of the sensor tips were taken and input into a series of in-house image analysis procedures. These procedures are able to find the edges of the sensors and separate the covered areas from exposed portions of the sensor tip. We also saw an increase in the image analysis method robustness by filtering out source illumination and construction of a platform. This allowed the analysis techniques to be applied in a fully automated fashion for the quantification of vapochromic material on a sensor tip.