PCA data preprocessing for neural network-based detection of parametric defects in analog IC

A new methodology for algorithmic selection of a proper training vector set for neural network learning in 2D PCA space is presented. In feed-forward neural networks with unsupervised learning, the training set selection plays a crucial role. In this paper, we propose a new approach to this selection using convex-hull graphics algorithms. Feed-forward neural network has been used for detecting parametric defects in a band pass filter circuit. As it is shown, well trained neural network is not only able to detect the faulty devices by classifying the analysed circuit´s parameter into a proper category but also identifies direction of an undesired deviation of the parameter