Key　performance　indicators　of　complex　industrial　process　such　as　production　quality　and　pollutant　emissions　concentration　are　difficult　to　be　measured　online　due　to　limited　detection　technology　and　high　economical　cost.　Their　modeling　samples　have　high　dimension,　strong　uncertainty,　and　small　sample,　which　cannot　satisfy　the　needs　of　traditional　machine　learning　algorithms.　A　virtual　sample　generation　method　based　on　generative　adversarial　fuzzy　neural　network　(GAFNN)　is　proposed　to　address　the　abovementioned　problems.　First,　an　adaptive　feature　selection　algorithm　based　on　random　forest　is　used　to　reduce　input　feature　for　the　original　real　samples.　Second,　candidate　virtual　samples　are　generated　by　GAFNN　to　alleviate　the　problems　of　uncertainty　and　small　sample.　Third,　the　virtual　samples　are　screened　by　a　multi-constrained　selection　mechanism　to　improve　the　quality　of　virtual　samples.　Finally,　a　deep　forest　classification　model　is　constructed　on　the　basis　of　the　mixed　samples　in　terms　of　the　original　real　and　selected　virtual　samples.　The　effectiveness　of　the　proposed　method　is　verified　on　benchmark　and　real　industrial　data.