In　complex　industrial　processes,　accurate　prediction　of　difficult-to-detect　parameters　such　as　product　quality　and　environmental　indicators　based　on　data-driven　modeling　is　helpful　for　the　realization　of　intelligent　detection,　parameter　optimization　and　intelligent　control.　However,　the　acquisition　of　the　truth　value　of　the　difficult-to-detect　parameters　is　limited　by　detection　technology,　labor　cost　and　economic　cost.　Therefore,　the　modeling　samples　have　problems　such　as　insufficient　feature　information,　large　information　gap　and　unbalanced　distribution.　That　leads　to　the　poor　generalization　performance　of　the　difficult-to-detect　parameter　prediction　model.　The　method　of　generating　virtual　samples　by　filling　the　information　gap　between　small　samples　can　expand　the　number　of　samples,　thereby　improving　the　accuracy　of　model　prediction.　In　view　of　this,　virtual　sample　generation　(VSG)　for　industrial　process　data　regression　modeling　is　reviewed　in　this　paper.　Firstly,　the　definition　of　VSG,　characteristics　of　small-sample　data　of　industrial　processes　and　their　causes　are　described.　Then,　VSG　methods　are　reviewed　from　three　perspectives,　i.e.,　small-sample　distribution,　interpolation　and　optimization.　Next,　through　the　comparative　analysis　of　the　existing　VSG　methods,　the　corresponding　research　difficulties　are　put　forward.　Finally,　we　summarize　the　existing　problems　and　look　forward　to　the　future　research　direction.　©　2021　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.