Abstract: Most of manifold learning algorithms such as locality and preserving projection(LPP) only concern the neighborhood samples and ignore those farthest ones, which may lead to missing out the global information. Therefore, a novel algorithm, named nearest and farthest distance preserving projection(NFDPP), is proposed to explore the relationship between the sample and its farthest samples as well as that between it and its nearest neighbors simultaneously. Dimension reduction can be performed by NFDPP with the “nearest neighbor matrix” and “farthest distance matrix”. Simulation and experiments on the engine misfire are conducted. Experiments results demonstrate that, comparing with principal component analysis(PCA), LPP, neighborhood preserving embedding(NPE) and linear discriminant analysis(LDA), the NFDPP can recognize the engine misfire fault effectively.

Key words: engine misfire, fault recognition, feature extraction, manifold learning