A　visualization　approach　for　process　fault　detection　using　bikernel　t-distributed　stochastic　neighbor　embedding　(bikernel　t-SNE)　is　described　in　this　paper.　Bikernel　t-SNE　preserves　the　dimension-reduction　ability　of　the　basic　t-SNE　and　enables　explicit　out-of-sample　extensions.　First,　the　explicit　projections　between　high-dimensional　data　and　low-dimensional　features　are　approximated　by　a　linear　combination　of　their　Gaussian　functions.　Second,　the　feature　kernel　matrix　is　transformed　into　two　latent　variables　through　entropy　component　analysis　(ECA).　With　the　bikernel　mapping　and　ECA,　outliers　are　depicted　away　from　the　inliers　in　two-dimensional　(2D)　scatter　plots.　Then,　the　squared　Mahalanobis　distance　is　used　as　the　fault　detection　index,　whose　confidence　boundary　is　represented　as　an　ellipse　in　the　2D　map.　The　performance　of　the　proposed　method　is　demonstrated　through　two　case　studies　on　the　Tennessee　Eastman　process　as　well　as　a　real　recombinant　human　granulocyte　colony-stimulating　factor　production　process.