Utilizing　fake　data　(simulated　based　on　mechanism　models　or　generated　through　data-driven　models)　for　data　enhancement　is　a　popular　approach　to　solve　the　problem　of　fault　diagnosis　with　small　samples.　Consequently,　the　quality　of　such　fake　data　impacts　fault　diagnosis　accuracy.　This　article　proposes　a　data　model　fusion　(DMF)-driven　framework　for　small　sample　fault　diagnosis.　This　framework　integrates　the　digital　twin　model　(DTM)　and　the　conditional　deep　convolutional　generative　adversarial　network　(C-DCGAN).　Digital　twin　data　(DTD)　under　various　fault　conditions　is　first　obtained　in　the　data　generation　stage　based　on　DTM　simulation.　Then,　a　data　generation　method　based　on　DTM-C-DCGAN　is　proposed.　The　method　adopts　DTD　as　the　soft-physics　constraint　input　to　the　generator　of　C-DCGAN.　Hence,　the　generator　is　induced　to　generate　data　that　is　more　consistent　with　the　failure　mechanism　and　closer　to　the　real　data.　During　the　fault　diagnosis　stage,　the　generated　data　(GD)　are　used　to　enhance　the　training　process　of　the　fault　diagnosis　model,　improving　its　generalization　ability.　Finally,　the　effectiveness　of　the　proposed　method　is　comprehensively　verified　via　two　publicly　rolling　bearing　datasets.　Compared　with　the　existing　single　data-driven　and　physics-based　methods,　the　experimental　results　demonstrate　that　the　proposed　DMF　method　can　significantly　enhance　the　quality　of　the　GD　and　improve　the　accuracy　of　fault　identification,　achieving　an　average　accuracy　of　97.31%.