Summary:
A nonlinear model-based approach is taken to quantitatively analyze time series data generated by analytical instruments.  An automated system is presented which takes as input an Analytical Instrument Association (AIA) network common data format (NetCDF) data file and generates an estimate of the concentrations of specific analytes of interest.  The system consists of three primary modules which, when combined, provide accurate and precise knowledge about unknown sample matrices, especially difficult-to-analyze mixture samples.  A preprocessing module extracts peak parameter estimates for the exponentially-modified Gaussian (EMG) model from the raw signal and utilizes nonlinear optimization techniques to fit the model to the observed data.  A novel sliding window approach ensures that the influence of neighboring peaks is included in the model fitting without the requirement for arbitrary established peak endpoints.  Modeled peak parameters are available for both instrument performance assessment and use in the analysis module.  Several traditional analysis algorithms are implemented in parallel on the raw and extracted data.  A complete analyte-based model-analysis algorithm is also developed for the analysis of complex mixture samples.  This algorithm utilizes concentration dependent, complete analyte models derived from calibration standards to model the observed signal in a unified manner.  Each analysis algorithm generates analyte concentration and confidence estimates and an additional performance measure.  The third module utilizes a fuzzy logic inference system to fuse the results of the multiple analysis algorithms into a single comprehensive sample characterization.  Software modules implement the described algorithms and interface to the supervisory controller of an automated chemical analysis system. Experimental results from gas chromatography data generated from simulated, standard, and actual environmental samples are presented along with conclusions drawn regarding the increase in accuracy and performance of the system over traditional methods.

This work was conducted at IRIS lab by Martin Hunt under the supervision of M. A. Abidi (Thesis Chair).