Deep　learning　methods　have　been　widely　investigated　in　machinery　fault　diagnosis　owing　to　their　powerful　feature　learning　capability.　However,　high　accuracy　is　hard　to　achieve　due　to　the　limited　fault　information　in　a　single　domain　when　the　data　volume　is　small.　In　this　paper,　an　optimized　Hilbert　curve　(OHC)　method　is　developed,　which　can　generate　a　novel　domain　to　highlight　the　fault　impulses　of　vibration　signals.　To　fully　mine　the　fault　information,　a　bidirectional-channel　convolutional　neural　network　with　an　attention　mechanism　is　further　proposed,　in　which　two　channels　are　constructed　and　a　transmission　channel　selection　is　conducted　by　a　novel　improved　convolutional　block　attention　module.　First,　the　OHC　images　and　the　time-frequency　representations　are　obtained　by　OHC　and　wavelet　transform　respectively.　Second,　the　two　types　of　representations　are　fed　into　the　channels　respectively　for　feature　learning.　Finally,　the　learned　features　are　allocated　to　different　attention　mechanism　for　feature　fusion　and　classification.　The　proposed　method　is　evaluated　via　the　datasets　of　rolling　bearings　and　planetary　gearboxes,　and　results　show　that　it　outperforms　the　comparison　methods.