REAL-TIME PRINCIPAL COMPONENT ANALYSIS USING PARALLEL KALMAN FILTER NETWORKS FOR PEAK PURITY ANALYSIS

A new approach for performing principal component analysis (PCA) during data acquisition is described. The method is based on a network of multilinear models which are fit to data with the discrete Kalman filter. Application to absorbance matrices such as those obtained In chromatography with multiwavelength detection is considered. Multivariate data projected into two- and three-dimensional subspaces are fit with linear and planar models, respectively. Model deviations, detected using principles of adaptive Kalman filtering, are used to elucidate the rank of the data set. Principal component eigenvectors are then constructed from the individual models. Results of this initial work using simulated and experimental data demonstrate that extraction of the first two principal components is readily accomplished and eigenvectors obtained are in good agreement with traditional batch PCA results. Extension to more principal components should be possible although it will increase the number and complexity of models. Advantages of the new algorithm include its recursive implementation, parallel structure, and ability to indicate model errors as a function of time. The procedure should prove particularly useful for self-modeling curve resolution applications in chromatography.