The　difficult-to-measure　parameters　such　as　product　quality　and　environmental　protection　indices　of　complex　industrial　processes　mostly　adopt　the　long-term　and　high-cost　off-line　detection　method,　which　makes　the　modeling　samples　for　building　the　risk　warning　model　of　these　difficult-to-measure　parameters　extremely　rare.　To　solve　the　above　problems,　a　virtual　sample　generation　method　combining　active　learning　(AL)　mechanism　and　generative　adversarial　network　(GAN)　for　process　data　is　proposed　and　used　for　risk　warning　modeling　of　difficult-to-measure　parameters.　Firstly,　the　difficult-to-measure　parameter　risk　level　is　added　as　condition　information　to　GAN,　so　that　the　generator　can　generate　virtual　samples　with　specified　risk　level.　Then,　the　virtual　samples　are　screened　by　using　the　maximum　mean　difference　(MMD)　approach　and　the　visual　results　are　obtained　by　using　principal　component　analysis　(PCA)　and　t-　distributed　stochastic　neighbor　embedding　(t-SNE).　Finally,　domain　experts　make　active　discrimination.　The　effectiveness　and　rationality　of　the　proposed　method　are　verified　by　using　benchmark　data　sets.