Artificial neural networks for nonlinear extensions of principal component analysis

One of the challenges in data intensive computing is the ability to deal with complexities intrinsic to vast data sets in high-dimensional spaces. Dealing with these properties becomes increasingly more difficult as both the dimensionality and the amount of the data increases. This type of problem is a challenging task for human visual perception. Therefore, multivariate statistical methods are generally used to identify and eliminate unnecessary dimensions so as to encourage a more parsimonious representation of the data set while retaining the maximum information content possible. The traditional multivariate statistical methods, though useful, are limited as the applications becoming more complex to unveil nonlinear relationships among the variables. These methods typically lack the capability to efficiently handle a large amount of on-line data. This research proposes a unified state of the art methodology for discovering both linear and nonlinear information in the data using linear and nonlinear projections without biasing the results by imposing preconceived parametric (model) structures. The new methods advocated here are nonlinear extensions of principal component analysis (PCA) by means of artificial neural networks (ANN) to address the important problem of on-line adaptive parameter estimation for effectively dealing with data intensive computing. Several industrial applications such as data compression, automotive air-fuel ratio modeling and intake valve carbon deposit analysis will be solved using the proposed techniques.