The　data　of　dissolved　gas　in　oil　is　an　important　state　parameter　of　the　power　reactor,　and　its　content　directly　reflects　the　working　state　of　the　reactor.　Fully　considering　the　irregularity　of　the　data　distribution　of　dissolved　gas　in　oil　and　the　problem　that　the　traditional　algorithm　model　cannot　effectively　deal　with　high-dimensional　data,　this　paper　proposes　an　abnormal　recognition　method　for　UHV　reactors　based　on　the　Deep　Auto-encoding　Gaussian　Mixture　Model　(DAGMM).　Based　on　historical　detection　data,　the　abnormal　state　of　high-dimensional　dissolved　gas　data　in　oil　is　realized　using　end-to-end　training,　combined　with　data　dimensionality　reduction　capabilities　of　autoencoders　and　Gaussian　mixture　model　clustering.　It　is　verified　with　an　example　analysis　that　the　proposed　method　can　accurately　identify　the　abnormal　state　of　the　UHV　reactor.　©　Published　under　licence　by　IOP　Publishing　Ltd.