To　detect　and　intelligently　identify　the　defects　of　the　vermicular　cast　iron　cylinder　head,　defective　casting　samples　were　made　corresponding　to　each　type　of　the　actual　defects.　We　setup　the　ultrasonic　testing　system　to　examine　the　defective　samples.　The　detected　defect　signals　were　processed　to　obtain　the　characteristic　spectrograms　of　the　defects,　which　were　further　sorted　and　classified　into　a　sample　database.　An　algorithm　based　on　a　convolutional　neural　network　was　proposed　to　identify　the　defects　intelligently.　A　convolutional　neural　network　model　was　established.　The　network　structure　and　parameters　were　optimized.　It　shows　that　a　neural　network　with　3　3　convolution　kernel　dimension,　3　convolution　layers,　20　convolution　kernels　in　each　layer　and　a　learning　rate　of　0.0005　can　effectively　identify　the　spectrograms　of　the　defects.　The　results　show　that　the　identification　accuracy　of　the　proposed　algorithm　is　97.14%.　The　model　meets　the　practical　requirements　of　cylinder　head　defect　detection.　The　detection　efficiency　has　improved　significantly.　©　Published　under　licence　by　IOP　Publishing　Ltd.