The　complexity　and　uncertainty　of　vibration　signals　from　distribution　transformers　pose　significant　challenges　for　diagnosing　mechanical　faults.　To　address　this,　this　paper　proposes　a　novel　fault　diagnosis　model　for　distribution　transformers,　which　combines　a　cross-domain　fusion　multi-scale　convolutional　autoencoder　(CFMS-CAE)　with　an　open-set　domain　adaptation　classifier　(OSDA-C).　Specifically,　in　order　to　extract　more　comprehensive　features,　a　convolutional　autoencoder　(CAE)　model　based　on　multi-output　objectives　is　constructed　to　extract　the　time-frequency　domain　characteristics　of　transformer　vibration　signals.　Multiple-scale　convolutional　layers　are　incorporated　into　the　convolutional　autoencoder　to　enable　multi-range　feature　extraction.　Additionally,　parameter　optimization　is　achieved　using　the　crayfish　optimization　algorithm　(COA).　Subsequently,　an　open-set　domain　adaptation　module　is　integrated　into　the　convolutional　neural　network　classifier　to　establish　boundaries　for　each　category　and　facilitate　the　identification　of　transformer　fault　categories,　including　unknown-type　faults.　The　experimental　results　demonstrate　that　the　proposed　method　is　effective　for　fault　identification　in　both　dry-type　and　oil-immersed　transformers,　with　average　accuracy　reaching　99.35%　and　99.62%,　respectively.　For　unknown-type　faults,　the　accuracy　also　achieved　100%　and　97.5%,　respectively.　©　2024　Elsevier　Ltd