Real　time　detection　of　production　quality　and　environmental　protection　indices　is　the　basis　for　realizing　operational　optimization　of　industrial　processes.　Insufficient　samples　are　one　of　the　main　problems　of　building　soft　sensor　model　for　such　difficulty-to-measure　key　indices.　Aiming　at　this　issue,　a　novel　virtual　sample　generation　(VSG)　strategy　based　on　feature　scaling　and　co-training　label　is　proposed.　First,　the　variational　autoencoder　(VAE)　is　used　to　make　feature　reduction　for　visualize　analysis.　Second,　the　typical　sample　cluster　and　its　boundary　are　selected　for　generating　the　virtual　sample　inputs.　Third,　the　feature　elevation　based　on　VAE　decoding　are　used　to　regain　the　original　input　dimension.　Fourth,　the　co-training　method　is　used　to　pseudo-label　and　select　the　optimal　virtual　sample　outputs.　Fifth,　the　hybrid　samples　are　used　to　construct　the　soft　measuring　model.　The　proposed　method　is　verified　by　the　dioxin　emission　data　of　an　actual　municipal　solid　waste　incineration　process　in　Beijing.　©　2023　IEEE.