Kernel-based nonlinear feature extractor and its application in electronic circuit fault diagnosis

Feature extraction of fault signals in analog circuits diagnosis aims to improve the separability of patterns belonging to different fault classes. Conventional linear feature extractor optimizes some separability criteria by linear transformation of pattern vectors. It can be generalized to its nonlinear versions by the introduction of kernel functions. A new separability measure defined by inter and intra-class scattering matrix and autocorrelation matrix of pattern samples is proposed in this paper, based on which a new class of nonlinear feature extractors are developed using kernel method. The proposed nonlinear feature extractor is used for Iris data transformation and analog circuit fault diagnosis. The experimental results shows that it outperforms nonlinear principle components analysis (PCA) feature extractor on the improvement of the separability of patterns in Iris data and the classification accuracy in electronic circuit fault diagnosis.