The　modeling　data　of　difficult-to-measure　parameters　such　as　industrial　process　quality　indicators　and　environmental　indicators　have　characteristics　of　small　samples　and　sparse　distribution.　A　new　virtual　sample　generation　(VSG)　method　based　on　probability　density　distribution　(PDF)　of　reduced　features　is　proposed　for　modeling　data　augmentation.　Firstly,　the　principal　component　analysis　(PCA)　is　used　to　reduce　the　feature　dimension　and　the　kernel　density　estimation　(KDE)　is　performed　on　the　obtained　independent　principal　components　to　generate　candidate　virtual　principal　components.　By　using　orthogonally　sampling　approach,　the　obtained　virtual　principal　components　are　used　to　re-construct　the　inputs　of　virtual　sample.　Then,　in　order　to　balance　the　accuracy　and　randomness　of　the　mapping　model,　an　ensemble　mapping　model　is　constructed　by　using　random　forest　(RF)　and　random　weight　neural　network　(RWNN)　to　obtain　the　outputs　of　virtual　samples.　Finally,　principal　component　contribution　rate,　KDE　smoothing　index,　number　of　candidate　virtual　principal　components　and　virtual　samples,　mapping　model　parameters　and　ensemble　weights　that　affect　the　quality　of　virtual　samples　are　selected　by　comprehensive　learning　particle　swarm　optimization　(CLPSO)　algorithm　for　obtaining　the　optimized　virtual　samples.　The　experimental　results　on　benchmark　dataset　and　dioxin　(DXN)　datasets　of　municipal　solid　waste　incineration　process　show　the　rationality　and　effectiveness　of　the　proposed　method.　©　2024　South　China　University　of　Technology.　All　rights　reserved.