Due　to　the　difficulty　of　detection　technology,　and　high　time　and　economic　cost,　the　modeling　samples　of　soft-sensing　model　with　difficult　parameters　have　some　problems,　such　as　small　numbers,　sparse　distribution,　and　imbalance,　which　seriously　restrict　the　generalization　performance　of　data-driven　models.　To　solve　the　above　problems,　a　virtual　sample　generation　(VSG)　method　based　on　multi-objective　particle　swarm　optimization　(MOPSO)　hybrid　optimization　is　proposed.　First,　the　population　representation　mechanism　of　the　integrated　learning　particle　swarm　optimization　algorithm　is　designed,　so　that　it　can　simultaneously　encode　the　continuous　and　the　discrete　variables.　Then,　the　fitness　function　of　the　integrated　learning　particle　swarm　optimization　algorithm　with　multistage　and　multi-objective　characteristics　is　defined　to　minimize　the　number　of　virtual　samples　while　ensuring　the　generalization　performance　of　the　model.　Finally,　a　multi-objective　hybrid　optimization　task　is　generated　for　virtual　samples　to　improve　the　integrated　learning　particle　swarm　optimization　algorithm,　so　that　it　can　adapt　to　the　variable　dimension　characteristics　of　the　virtual　sample　optimization　process　and　improve　the　convergence　speed.　At　the　same　time,　the　comprehensive　evaluation　index　and　distribution　similarity　index　are　proposed　for　evaluating　the　quality　of　virtual　samples　by　referring　to　metric　learning　for　the　first　time.　In　this　paper,　two　benchmark　datasets　and　an　actual　industrial　dataset　are　used　to　verify　the　effectiveness　and　superiority　of　the　proposed　method.　©　2024　Science　Press.　All　rights　reserved.