To　address　the　problem　that　the　model　unable　to　evaluate　accurately　when　the　distribution　difference　between　the　actual　fault　samples　and　the　training　samples,　an　evalution　model　based　on　directional　adversarial　transfer　is　proposed.　Firstly，a　traditional　adversarial　transfer　model　based　on　stacked　auto-encoder　is　built.　Through　adversarial　learning　between　training　samples　and　potential　samples，the　model　extracts　the　common　features　of　the　samples，thus　improving　the　ability　of　the　model　to　evaluate　potential　faults.　Then，a　directional　adversarial　method　is　added　to　the　traditional　adversarial　transfer　model　to　selectively　transfer　the　training　samples.　The　proposed　method　changes　the　weights　of　different　training　samples　in　the　adversarial　training　according　to　the　similarity　values　of　training　samples　and　potential　fault　samples，thus　reducing　the　negative　impact　of　large　difference　samples　on　transfer　training.　The　proposed　method　improves　the　accuracy　by　5.72%　compared　to　the　traditional　adversarial　transfer　model　in　the　real　system　simulation　examples.　The　test　results　show　that　the　proposed　method　can　effectively　improve　the　transferability　and　evaluation　accuracy　of　the　model.　©　2025　Science　Press.　All　rights　reserved.