The　deep　learning　models　have　swept　the　field　of　intelligent　fault　diagnosis,　especially　the　deep　convolutional　neural　network　(DCNN)　with　ability　to　automatically　extract　features　from　large-scale　data.　On　one　hand,　as　most　of　the　existing　DCNN　models　are　specifically　designed　for　image　task,　direct　application　of　these　models　to　machine　fault　diagnosis　will　cause　domain　bias.　On　the　other　hand,　it　is　time-consuming　and　often　dependent　on　expert　knowledge　to　design　a　DCNN　model　for　a　given　fault　diagnosis　task　from　a　specific　dataset.　In　this　paper,　a　differentiable　architecture　search　method　through　gradient　optimization　is　proposed　to　design　a　new　network　model　for　aeroengine　bevel　gear　fault　diagnosis.　This　method　can　design　efficient　network　from　a　complex　search　space　for　a　specific　dataset　via　constructing　the　hyper-network　and　gradient-based　search　strategy.　The　proxy　model　and　optimization　approximation　are　used　to　speed　up　the　searching　process.　Compared　with　the　traditional　discrete　neural　architecture　search　methods,　this　method　has　great　superiority　in　computational　cost.　The　aeroengine　bevel　gear　dataset　is　used　to　verify　the　performance　of　this　method,　and　only　3　GPU　hours　are　required　for　this　dataset　to　search　the　optimized　network.　The　proposed　method　achieves　the　state-of-the-art　performance　under　different　signal-to-noise　ratios　via　comparisons　with　other　mainstream　manually　designed　networks.　©　2020　IEEE.