A neural network approach for the reduction of the dimensionality of slowly time-varying electromagnetic inverse problems

In real electromagnetic problems, it is often required to rapidly interpret a lot of experimental raw data provided by a set of sensors with the aim of controlling the time evolution of a system under observation. This is the case, for example, of the real time control of a plasma discharge in a tokamak device for nuclear fusion experiments. Some procedures apt to carry out principal component analysis (PCA) by using suitable artificial neural network (ANN) models are presented. The related models allow one to adaptively extract the PCs directly from the input data without estimating in advance the covariance matrix over the sample database. The proposed architectures may also cope with non-stationary problems. Two examples of application in electromagnetics are presented which concern respectively the reduction of dimensionality in a typical identification problem and the adaptive recovery of the PCs during a slow change of the statistics of the simulated experiment.