Intelligent　fault　diagnosis,　which　is　mainly　based　on　neural　network,　has　been　widely　used　in　machinery　monitoring.　Although　such　deep　learning　methods　are　effective,　the　new　architectures　are　mainly　handcrafted　by　series　of　experiments　that　require　ample　time　and　substantial　efforts.　To　automate　process　of　building　neural　networks　and　save　designing　time,　a　novel　differentiable　neural　architecture　search　method　is　proposed.　By　gradually　reducing　candidate　operations　while　retaining　trained　parameters　during　pruning,　computation　consumed　by　each　stage　of　neural　architecture　search　is　decreased,　which　accelerates　search　process.　To　improve　inferential　efficiency　of　subnetworks,　specially　designed　penalty　terms　are　introduced　into　the　objective　function　for　searching　optimal　numbers　of　layers　and　nodes,　which　can　reduce　complexity　of　subnetworks　and　save　calculation　time　of　signal　analysis.　In　addition,　exclusive　competition　between　candidate　operations　is　broken　by　changing　discretization　and　selection　methods　of　operations,　which　provides　a　basis　for　channel　fusion.　Effectiveness　of　the　proposed　method　is　verified　by　two　datasets.　Experiments　show　that　this　method　can　generate　subnetworks　of　lower　complexity　and　less　computational　cost　than　other　state-of-art　neural　architecture　search　techniques,　while　achieving　competitive　result.　©　2021　Elsevier　Ltd